Convolutional Neural Networks (CNNs) are widely used in image classification tasks, but their performance can degrade significantly under poor illumination conditions. Although numerous image enhancement methods have been developed to mitigate this issue, identifying the most appropriate technique for a specific image remains challenging. This work aims to determine the most effective image enhancement algorithm for a potentially dimmed input image to enhance the classification prediction performance. To do this, a trained regressor, which evaluates different features of an image, ascertains the increase or decrease in classification prediction performance between the input image and the enhanced image. These predictions are then used to select the most suitable image enhancement algorithm to be applied to the input image for optimal CNN classification. Experimental results using various CNN architectures and enhancement techniques demonstrate that the proposed strategy consistently improves classification performance under challenging lighting conditions.

Estimating the speed of nearby vehicles is essential for driver assistance. A real-time, camera-only pipeline is presented that uses an onboard monocular camera: vehicles are detected and tracked with an off-the-shelf one-stage CNN (YOLOv8); distance is approximated from bounding-box angular width using class-dependent priors (2.0[Formula: see text]m cars; 2.5[Formula: see text]m larger vehicles) and camera intrinsics; a Nadaraya-Watson kernel smooths the distance sequence, and its analytic derivative yields relative speed. The approach supports multiple targets without dedicated ranging hardware. Evaluation on CARLA synthetic video with ground truth analyzes estimated versus ground-truth distance, estimate/ground-truth ratio versus image-center displacement, and kernel-based speed versus a polynomial trend. Results show a positional bias away from the image center and a stability-lag trade-off due to smoothing. The contribution is a detector-agnostic distance-speed head that couples angular geometry with analytic Nadaraya-Watson smoothing and differentiation for real-time operation, positioned as a low-cost alternative or complement to active sensors, with limitations and paths to real-world validation outlined.

• In-depth study of the impact of intensity value regularization on 3D MRI segmentation methods: nnU-Net, WNet, and Primus. • A broad range of five intensity regularization approaches used in deep learning literature has been included. • The analysis focuses on the influence of the number of channels in deep learning architectures, specifically structural elements like Convolutional Neural Networks, Transformers, and hybrid models. • Comprehensive dataset compendium: three relevant open-access neurological disorders datasets were considered: glioblastoma, multiple sclerosis, and epilepsy, using T1 and FLAIR MRI sequences. • An exhaustive statistical analysis was conducted to evaluate the reported results and support the findings thoroughly. In this work, a comprehensive analysis of the impact of intensity value regularization methods on 3D MRI segmentation for three neurological disorders: glioblastoma, multiple sclerosis, and epilepsy, is presented. The experiments were conducted through three architectures: nnU-Net (convolutional neural network), WNet (hybrid combining convolutional and transformer elements), and Primus (transformer-based), considering both FLAIR and T1-weighted images, as well as FLAIR-only scenarios. The statistical analysis conducted underscores the crucial role of intensity regularization in the performance. The results indicate that among the intensity regularization methods tested in this study, KDE, White-stripe, and Z-score standardizations proved to be particularly effective. Furthermore, nnU-Net is the most robust architecture against intensity variability, with small improvements of around 3%. Meanwhile, methods incorporating TF elements are more sensitive to these variations. WNet demonstrates slightly greater gains, around 6%. While Primus can be less stable and underperform compared to nnU-Net and WNet in most cases; nonetheless, it remains a promising and competitive option. Additionally, it has been demonstrated that adding an extra channel does not necessarily guarantee improved performance, while also increasing computational cost.

• Denoising Diffusion Implicit Model (DDIM) sampling generates latent representations from noise conditioned on low-resolution 3D MRI slices. • Low-resolution 3D MRI volume is sliced along one axis to create latent space representations. • Latent space interpolation enhances 3D MRI resolution by generating additional slices. • Decoder reconstructs high-resolution 3D MRI slices from interpolated latent representations. • Enhancing MRI 3D volume super-resolution across all axes. In various image processing tasks, enhancing resolution is a fundamental challenge, particularly along specific axes where resolution tends to be lower. This limitation can hinder the performance of models in tasks such as medical image analysis. Traditional approaches often involve interpolation techniques, but they may lead to loss of information or introduce artifacts. Recently, deep learning-based methods, especially those utilizing latent spaces, have shown promise in addressing this issue. Because typical super-resolution methods are designed for 2D images, they can easily be applied to increase resolution in two of the axes in a volumetric MRI, but not the other axis. While volumetric (3D) deep learning models for super-resolution have been proposed, they have very high computational requirements, even if the region of interest to super-resolve does not span the whole volume. In our work, we propose a novel approach that uses a diffusion latent model to increase resolution along an arbitrary axis. Our method involves transforming input images into a latent space, where a U-Net model is employed to capture high-level features. Crucially, just before decoding, we introduce a linear interpolation in the latent space to enhance resolution along the specified axis. This interpolated latent representation is then decoded by the decoder, yielding images with increased resolution, thus achieving a resolution across all axes and, therefore, an increase in resolution of the entire volume, using a 2D deep learning model rather than a fully-fledged 3D model. The proposal has been extensively tested with a wide range of brain lesions and brain tumor images of T1, T2, and FLAIR modes. The experimental comparison with several state-of-the-art methods has consistently shown the advantages of our approach.

The real-time synthesis of 3D views, facilitated by convolutional neural networks like NeX, is increasingly pivotal in various computer vision applications. These networks are trained using photographs taken from different perspectives during the training phase. However, these images may be susceptible to contamination from noise originating from the vision sensor or the surrounding environment. This research meticulously examines the impact of noise on the resulting image quality of 3D views synthesized by the NeX network. Various noise levels and scenes have been incorporated to substantiate the claim that the presence of noise significantly degrades image quality. Additionally, a new strategy is introduced to improve image quality by calculating consensus among NeX networks trained on images pre-processed with a denoising algorithm. Experimental results confirm the effectiveness of this technique, demonstrating improvements of up to 1.300 dB and 0.032 for Peak Signal Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM), respectively, under certain scenes and noise levels. Notably, the performance gains are especially significant when using synthesized images generated by NeX from noisy inputs in the consensus process.

Abstract Creating an annotated medical image dataset is challenging and traditionally reliant on labor-intensive manual annotations. Additionally, these datasets often present substantial imbalances regarding sensing devices, class of medical disorders, and patient ethnicity and phenotype. Recently, there has been a research interest in mitigating these issues by employing data augmentation with generative models. However, the quality of images and semantics in medical image datasets are critical for computer vision tasks such as image segmentation. This paper presents DatasetGAN2-ADA, which aims to mitigate these difficulties by presenting an innovative deep-style interpreter robust against anomalous synthesis and designed to automate annotated image generation entirely. By leveraging the capabilities of StyleGAN2-ADA with an improved architecture of DatasetGAN and an enhanced execution framework integrated with an anomaly detector based on custom features, we propose a combined strategy for eliminating flawed synthetic images and masks. Furthermore, we propose exploiting image projections and preexisting semantics, eliminating the need for manual annotations to train our deep-style interpreter. The experimental results obtained with a magnetic resonance image (MRI) dataset demonstrate that DatasetGAN2-ADA is strongly effective in improving the efficiency and quality of synthetic generation, rejecting the synthesis of a substantial amount of low-quality images and masks. Then, an extension of this method is evaluated for detecting anomalous latent vectors a priori of the image synthesis, achieving up to 95.24% precision and illustrating its compelling potential for practical applications in medical imaging.

Large image datasets with annotated pixel-level semantics are necessary to train and evaluate supervised deep-learning models. These datasets are very expensive in terms of the human effort required to build them. Still, recent developments such as DatasetGAN open the possibility of leveraging generative systems to automatically synthesise massive amounts of images along with pixel-level information. This work analyses DatasetGAN and proposes a novel architecture that utilises the semantic information of neighbouring pixels to achieve significantly better performance. Additionally, the overfitting observed in the original architecture is thoroughly investigated, and modifications are proposed to mitigate it. Furthermore, the implementation has been redesigned to greatly reduce the memory requirements of DatasetGAN, and a comprehensive study of the impact of the number of classes in the segmentation task is presented. • An exhaustive study was presented, outperforming DatasetGAN architecture. • An extensive analysis of the overfitting in the original architecture was presented. • A novel architecture using adjacent pixels was proposed, enhancing performance. • A new computation strategy was offered, reducing RAM requirements drastically. • The impact of class count on DatasetGAN’s performance was thoroughly examined.

Typically, deep learning models for image segmentation tasks are trained using large datasets of images annotated at the pixel level, which can be expensive and highly time-consuming. A way to reduce the amount of annotated images required for training is to adopt a semi-supervised approach. In this regard, generative deep learning models, concretely Generative Adversarial Networks (GANs), have been adapted to semi-supervised training of segmentation tasks. This work proposes MaskGDM, a deep learning architecture combining some ideas from EditGAN, a GAN that jointly models images and their segmentations, together with a generative diffusion model. With careful integration, we find that using a generative diffusion model can improve EditGAN performance results in multiple segmentation datasets, both multi-class and with binary labels. According to the quantitative results obtained, the proposed model improves multi-class image segmentation when compared to the EditGAN and DatasetGAN models, respectively, by [Formula: see text] and [Formula: see text]. Moreover, using the ISIC dataset, our proposal improves the results from other models by up to [Formula: see text] for the binary image segmentation approach.

Autonomous driving requires information such as the velocity of other vehicles to prevent potential hazards. This work proposes a real-time deep learning-based framework to estimate vehicle speeds from image captures through an onboard camera. Vehicles are detected and tracked by the proposed deep neural networks and a tracking algorithm, which analyzes the trajectories. Finally, a linear regression model estimates the speed of a vehicle based on its position and size in the camera frame. This proposal has been tested with two sequences of the Prevention dataset with satisfactory results. The system can estimate the speed of multiple vehicles simultaneously. It can be integrated easily with onboard computer systems, thus allowing to development of a low-cost solution for speed estimation in an everyday vehicle. The potential applications include vehicle safety systems, driver assistance, and autonomous driving technologies.

Abstract Detection of small objects is one of the main challenges to be improved in deep learning, mainly due to the small number of pixels and scene’s context, leading to a loss in performance. In this paper, we present an optimized approach based on deep object detection models that allow the detection of a higher number of elements and improve the score obtained for their class inference. The main advantage of the presented methodology is that it is not necessary to modify the internal structure of the selected convolutional neural network model or re-training for a specific scene. Our proposal is based on detecting initial regions to generate several sub-images using super-resolution (SR) techniques, increasing the number of pixels of the elements, and re-infer over these areas using the same pre-trained model. A reduced set of windows is calculated in the super-resolved image by analyzing a computed graph that describes the distances among the preliminary object detections. This analysis is done by finding maximal cliques on it. This way, the number of windows to be examined is diminished, significantly speeding up the detection process. This framework has been successfully tested on real traffic sequences obtained from the U.S. Department of Transportation. An increase of up to 44.6% is achieved, going from an average detection rate for the EfficientDet D4 model of 14.5% compared to 59.1% using the methodology presented for the first sequence. Qualitative experiments have also been performed over the Cityscapes and VisDrone datasets.

Abstract In the area of medical imaging, one of the factors that can negatively influence the performance of prediction algorithms is the limited number of observations for each class within a labeled dataset. Usually, in order to increase the samples, a second set of unlabeled images is used. However, this set adds two new problems (i) finding patient observations with different pathologies than those observed in the labeled data set and (ii) finding images belonging to a different distribution from the dataset used in the model training process. This way, merging datasets from different sources can have an adverse effect on the distribution of features. Encountering this type of data (better known as out-of-distribution data) within the deployment environments may also lead to varying degrees of performance degradation as can be seen in the different experimental results obtained. In this research, a study of the behavior of Feature Density is made, as a mathematical model for the estimation of predictive uncertainty in supervised classification algorithms, in order to improve the behavior when out-of-distribution data are presented in the dataset. The Feature Density method is based on the estimation of feature density by means of histogram calculation (or Probability Density Function). The advantage of this method over the baseline approach (Mahalanobis distance) is that it does not assume a Gaussian-type distribution of sample characteristics and serves to estimate the uncertainty. This work focuses on the binary classification of mammography X-ray images from three different datasets simulating the condition of a different degree of contamination with out-of-distribution sample. According to the obtained results, the performance of the proposed method depends directly on the architecture of the implemented neural network.

The exponential increase in the use of technology in road management systems has led to real-time visual information in thousands of locations on road networks. A previous step in preventing or detecting accidents involves identifying vehicles on the road. The application of convolutional neural networks in object detection has significantly improved this field, enhancing classical computer vision techniques. Although, there are deficiencies due to the low detection rate provided by the available pre-trained models, especially for small objects. The main drawback is that they require manual labeling of the vehicles that appear in the images from each IP camera located on the road network to retrain the model. This task is not feasible if we have thousands of cameras distributed across the extensive road network of each nation or state. Our proposal presented a new automatic procedure for detecting small-scale objects in traffic sequences. In the first stage, vehicle patterns detected from a set of frames are generated automatically through an offline process, using super-resolution techniques and pre-trained object detection networks. Subsequently, the object detection model is retrained with the previously obtained data, adapting it to the analyzed scene. Finally, already online and in real-time, the retrained model is used in the rest of the traffic sequence or the video stream generated by the camera. This framework has been successfully tested on the NGSIM and the GRAM datasets.

Noise robustness is crucial when approaching a moving detection problem since image noise is easily mistaken for movement. In order to deal with the noise, deep denoising autoencoders are commonly proposed to be applied on image patches with an inherent disadvantage with respect to the segmentation resolution. In this work, a fully convolutional autoencoder-based moving detection model is proposed in order to deal with noise with no patch extraction required. Different autoencoder structures and training strategies are also tested to get insights into the best network design approach.

Neural rendering is a relatively new field of research that aims to produce high quality perspectives of a 3D scene from a reduced set of sample images. This is done with the help of deep artificial neural networks that model the geometry and color characteristics of the scene. The NeX model relies on neural basis expansion to yield accurate results with a lower computational load than the previous NeRF model. In this work, a procedure is proposed to further enhance the quality of the perspectives generated by NeX. Our proposal is based on the combination of the outputs of several NeX models by a consensus mechanism. The approach is compared to the original NeX for a wide range of scenes. It is found that our method significantly outperforms the original procedure, both in quantitative and qualitative terms.

Abstract The detection of small objects is one of the problems present in deep learning due to the context of the scene or the low number of pixels of the objects to be detected. According to these problems, current pre‐trained models based on convolutional neural networks usually give a poor average precision, highlighting some as CenterNet HourGlass104 with a mean average precision of 25.6%, or SSD‐512 with 9%. This work focuses on the detection of small objects. In particular, our proposal aims to vehicle detection from images captured by video surveillance cameras with pre‐trained models without modifying their structures, so it does not require retraining the network to improve the detection rate of the elements. For better performance, a technique has been developed which, starting from certain initial regions, detects a higher number of objects and improves their class inference without modifying or retraining the network. The neural network is integrated with processes that are in charge of increasing the resolution of the images to improve the object detection performance. This solution has been tested for a set of traffic images containing elements of different scales to check the efficiency depending on the detections obtained by the model. Our proposal achieves good results in a wide range of situations, obtaining, for example, an average score of 45.1% with the EfficientDet‐D4 model for the first video sequence, compared to the 24.3% accuracy initially provided by the pre‐trained model.

Social conflicts appearing in the media are increasing public awareness about security issues, resulting in a higher demand of more exhaustive environment monitoring methods. Automatic video surveillance systems are a powerful assistance to public and private security agents. Since the arrival of deep learning, object detection and classification systems have experienced a large improvement in both accuracy and versatility. However, deep learning-based object detection and classification systems often require expensive GPU-based hardware to work properly. This paper presents a novel deep learning-based foreground anomalous object detection system for video streams supplied by panoramic cameras, specially designed to build power efficient video surveillance systems. The system optimises the process of searching for anomalous objects through a new potential detection generator managed by three different multivariant homoscedastic distributions. Experimental results obtained after its deployment in a Jetson TX2 board attest the good performance of the system, postulating it as a solvent approach to power saving video surveillance systems.

In this paper, we present a new framework for Content Based Image Retrieval (CBIR), based on Dynamic Ensemble Selection (DES) of classifiers. Herein, the classifiers consist of Convolutional Neural Networks (CNNs) that output the class probability vector of each input image. First, a diverse ensemble is built by training several weak classifiers on different training subsets, from the retrieval database. Then, each training image is passed throughout all the candidate classifiers to extract its class probability vector. These extracted vectors are then combined to make its final representation. When a new query image is input, the level of competence of all the candidate classifiers is measured on the region of competence of this image, in the training set. Then, only the most competent classifiers are selected to extract the class probability vectors from each query image. They are then combined to make a more discriminant image representation. To the best of our knowledge DES has never been applied to the CBIR area, which is the novelty of our research. The obtained results demonstrate the effectiveness of the dynamic selection, compared to the use of all the ensemble members, in terms of precision, recall, and mean average precision.

One of the most invasive cancer types which affect women is breast cancer. Unfortunately, it exhibits a high mortality rate. Automated histopathological image analysis can help to diagnose the disease. Therefore, computer aided diagnosis by intelligent image analysis can help in the diagnosis tasks associated with this disease. Here we propose an automated system for histopathological image analysis that is based on deep learning neural networks with convolutional layers. Rather than a single network, an ensemble of them is built so as to attain higher recognition rates, which are obtained by computing a consensus decision from the individual networks of the ensemble. A final step involves the optimization of the set of networks that are included in the ensemble by a genetic algorithm. Experimental results are provided with a set of benchmark images, with favorable outcomes.

Computer aided diagnosis for mammogram images have seen positive results through the usage of deep learning architectures. However, limited sample sizes for the target datasets might prevent the usage of a deep learning model under real world scenarios. The usage of unlabeled data to improve the accuracy of the model can be an approach to tackle the lack of target data. Moreover, important model attributes for the medical domain as model uncertainty might be improved through the usage of unlabeled data. Therefore, in this work we explore the impact of using unlabeled data through the implementation of a recent approach known as MixMatch, for mammogram images. We evaluate the improvement on accuracy and uncertainty of the model using popular and simple approaches to estimate uncertainty. For this aim, we propose the usage of the uncertainty balanced accuracy metric.

Image restoration, which is the process of denoising noisy images in order to recover their latent clean images, has been frequently addressed. The importance of this field resides in the impact of noisy images on the performance of computer vision systems. In this work, a deep autoencoder neural network architecture is proposed to denoise images affected by Gaussian noise. The performance of the system is enhanced by using a test time augmentation scheme. Experiments have been carried out by considering different levels of Gaussian noise. Results demonstrate the suitability of the proposed methodology in order to enhance the quality of the image restoration process in images affected by Gaussian noise.

Skin cancer is one of the most prevalent diseases among people. Physicians have a challenge every time they have to determine whether a diseased skin is benign or malign. There exist clinical diagnosis methods (such as the ABCDE rule), but they depend mainly on the physician's experience and might be imprecise. Deep learning models are very extended in medical image analysis, and several deep models have been proposed for moles classification. In this work, a convolutional neural network is proposed to support the diagnosis procedure. The proposed MobileNetV2-based model is improved by a shifting technique, providing better performance than raw transfer learning models for moles classification. Experiments show that this technique could be applied to the state-of-the-art deep models to improve their results and outperform the training phase.

Ellipses are among the most frequently used geometric models in visual pattern recognition and digital image analysis. This work aims to combine the outputs of an ensemble of ellipse fitting methods, so that the deleterious effect of suboptimal fits is alleviated. Therefore, the accuracy of the combined ellipse fit is higher than the accuracy of the individual methods. Three characterizations of the ellipse have been considered by different researchers: algebraic, geometric, and natural. In this paper, the natural characterization has been employed in our method due to its superior performance. Furthermore, five ellipse fitting methods have been chosen to be combined by the proposed consensus method. The experiments include comparisons of our proposal with the original methods and additional ones. Several tests with synthetic and bitmap image datasets demonstrate its great potential with noisy data and the presence of occlusion. The proposed consensus algorithm is the only one that ranks among the first positions for all the tests that were carried out. This demonstrates the suitability of our proposal for practical applications with high occlusion or noise.

Foreground object detection algorithms should be insensitive to noise present in the analyzed video sequences. In this work, a study of a type of non-supervised deep learning network, called autoencoder, is performed. They are suited to reduce input dimensionality and capture the most relevant information present in a region or image. Therefore, different types of autoencoders, deterministic and variational, with different architectures, activation functions and number of layers, are analyzed. This neural network is combined with a probabilistic mixture model which attempts to classify each video frame region as background and foreground.

The acquisition of 3D MRIs is adversely affected by many degrading factors including low spatial resolution and noise. Image enhancement techniques are commonplace, but there are few proposals that address the increase of the spatial resolution and noise removal at the same time. An algorithm to address this vital need is proposed in this presented work. The proposal tiles the 3D image space into parallelepipeds, so that a median filter is applied in each parallelepiped. The results obtained from several such tilings are then combined by a subsequent median computation. The convergence properties of the proposed method are formally proved. Experimental results with both synthetic and real images demonstrate our approach outperforms its competitors for images with high noise levels. Moreover, it is demonstrated that our algorithm does not generate any hallucinations.

Homography estimation from a pair of natural images is a problem of paramount importance for computer vision. Specialized deep convolutional neural networks have been proposed to accomplish this task. In this work, a method to enhance the result of this kind of homography estimators is proposed. Our approach generates a set of tentative color transformations for the image pair. Then the color transformed image pairs are evaluated by a regressor that estimates the quality of the homography that would be obtained by supplying the transformed image pairs to the homography estimator. Then the image pair that is predicted to yield the best result is provided to the homography estimator. Experimental results are shown, which demonstrate that our approach performs better than the direct application of the homography estimator to the original image pair, both in qualitative and quantitative terms.

Convolutional Neural Networks (CNNs) are widely used due to their high performance in many tasks related to computer vision. In particular, image classification is one of the fields where CNNs are employed with success. However, images can be heavily affected by several inconveniences such as noise or illumination. Therefore, image enhancement algorithms have been developed to improve the quality of the images. In this work, the impact that brightness and image contrast enhancement techniques have on the performance achieved by CNNs in classification tasks is analyzed. More specifically, several well known CNNs architectures such as Alexnet or Googlenet, and image contrast enhancement techniques such as Gamma Correction or Logarithm Transformation are studied. Different experiments have been carried out, and the obtained qualitative and quantitative results are reported.

Self-organizing maps (SOMs) are aimed to learn a representation of the input distribution which faithfully describes the topological relations among the clusters of the distribution. For some data sets and applications, it is known beforehand that some regions of the input space cannot contain any samples. Those are known as forbidden regions. In these cases, any prototype which lies in a forbidden region is meaningless. However, previous self-organizing models do not address this problem. In this paper, we propose a new SOM model which is guaranteed to keep all prototypes out of a set of prespecified forbidden regions. Experimental results are reported, which show that our proposal outperforms the SOM both in terms of vector quantization error and quality of the learned topological maps.

Automatic video surveillance systems are usually designed to detect anomalous objects being present in a scene or behaving dangerously. In order to perform adequately, they must incorporate models able to achieve accurate pattern recognition in an image, and deep learning neural networks excel at this task. However, exhaustive scan of the full image results in multiple image blocks or windows to analyze, which could make the time performance of the system very poor when implemented on low cost devices. This paper presents a system which attempts to detect abnormal moving objects within an area covered by a PTZ camera while it is planning. The decision about the block of the image to analyze is based on a mixture distribution composed of two components: a uniform probability distribution, which represents a blind random selection, and a mixture of Gaussian probability distributions. Gaussian distributions represent windows in the image where anomalous objects were detected previously and contribute to generate the next window to analyze close to those windows of interest. The system is implemented on a Raspberry Pi microcontroller-based board, which enables the design and implementation of a low-cost monitoring system that is able to perform image processing.

This paper presents a methodology for estimating pollution on roads by analyzing traffic video sequences. The objective is to take advantage of the huge network of static cameras which is possible to lind in the road system of any state or country to estimate the pollution on each area. This proposal uses deep learning neural networks for the object detection, and a pollution estimation model based on the frequency of vehicles and their speed. The experiments show promising results which suggest that the system can be used alone or combined with existing systems for measuring pollution on roads.

In this paper, a new self-organizing neural gas model that we call Growing Hierarchical Bregman Neural Gas (GHBNG) has been proposed. Our proposal is based on the Growing Hierarchical Neural Gas (GHNG) in which Bregman divergences are incorporated in order to compute the winning neuron. This model has been applied to anomaly detection in video sequences together with a Faster R-CNN as an object detector module. Experimental results not only confirm the effectiveness of the GHBNG for the detection of anomalous object in video sequences but also its self-organization capabilities.

Enhancing resolution is a permanent goal in magnetic resonance (MR) imaging, in order to keep improving diagnostic capability and registration methods. Super-resolution (SR) techniques are applied at the postprocessing stage, and their use and development have progressively increased during the last years. In particular, example-based methods have been mostly proposed in recent state-of-the-art works. In this paper, a combination of a deep-learning SR system and a random shifting technique to improve the quality of MR images is proposed, implemented and tested. The model was compared to four competitors: cubic spline interpolation, non-local means upsampling, low-rank total variation and a three-dimensional convolutional neural network trained with patches of HR brain images (SRCNN3D). The newly proposed method showed better results in Peak Signal-to-Noise Ratio, Structural Similarity index, and Bhattacharyya coefficient. Computation times were at the same level as those of these up-to-date methods. When applied to downsampled MR structural Tl images, the new method also yielded better qualitative results, both in the restored images and in the images of residuals.

Abstract The construction of a model of the background of a scene still remains as a challenging task in video surveillance systems, in particular for moving cameras. This work presents a novel approach for constructing a panoramic background model based on the neural gas network and a subsequent piecewise linear interpolation by Delaunay triangulation. Furthermore, an ensemble model of neural gas networks is also proposed. The approach can handle arbitrary camera directions and zooms for a pan‐tilt‐zoom camera‐based surveillance system. After testing the proposed approach on several indoor sequences, the results demonstrate that the proposed methods are effective and suitable to use for real‐time video surveillance applications.

One of the most important challenges in computer vision applications is the background modeling, especially when the background is dynamic and the input distribution might not be stationary, i.e. the distribution of the input data could change with time (e.g. changing illuminations, waving trees, water, etc.). In this work, an unsupervised learning neural network is proposed which is able to cope with progressive changes in the input distribution. It is based on a dual learning mechanism which manages the changes of the input distribution separately from the cluster detection. The proposal is adequate for scenes where the background varies slowly. The performance of the method is tested against several state-of-the-art foreground detectors both quantitatively and qualitatively, with favorable results.

In this paper a controller for PTZ cameras based on an unsupervised neural network model is presented. It takes advantage of the foreground mask generated by a non-parametric foreground detection subsystem. Thus, our aim is to optimize the movements of the PTZ camera to attain the maximum coverage of the observed scene in presence of moving objects. A growing neural gas (GNG) is applied to enhance the representation of the foreground objects. Both qualitative and quantitative results are reported using several widely used datasets, which demonstrate the suitability of our approach.

The construction of a model of the background of a scene still remains as a challenging task in video surveillance systems, in particular for moving cameras. This work presents a novel approach for constructing a panoramic background model based on competitive learning neural networks and a subsequent piecewise linear interpolation by Delaunay triangulation. The approach can handle arbitrary camera directions and zooms for a Pan-Tilt-Zoom (PTZ) camera-based surveillance system. After testing the proposed approach on several indoor sequences, the results demonstrate that the proposed method is effective and suitable to use for real-time video surveillance applications.

The growing neural gas (GNG) self-organizing neural network stands as one of the most successful examples of unsupervised learning of a graph of processing units. Despite its success, little attention has been devoted to its extension to a hierarchical model, unlike other models such as the self-organizing map, which has many hierarchical versions. Here, a hierarchical GNG is presented, which is designed to learn a tree of graphs. Moreover, the original GNG algorithm is improved by a distinction between a growth phase where more units are added until no significant improvement in the quantization error is obtained, and a convergence phase where no unit creation is allowed. This means that a principled mechanism is established to control the growth of the structure. Experiments are reported, which demonstrate the self-organization and hierarchy learning abilities of our approach and its performance for vector quantization applications.

In this work, a novel self-organizing model called growing neural forest (GNF) is presented. It is based on the growing neural gas (GNG), which learns a general graph with no special provisions for datasets with separated clusters. On the contrary, the proposed GNF learns a set of trees so that each tree represents a connected cluster of data. High dimensional datasets often contain large empty regions among clusters, so this proposal is better suited to them than other self-organizing models because it represents these separated clusters as connected components made of neurons. Experimental results are reported which show the self-organization capabilities of the model. Moreover, its suitability for unsupervised clustering and foreground detection applications is demonstrated. In particular, the GNF is shown to correctly discover the connected component structure of some datasets. Moreover, it outperforms some well-known foreground detectors both in quantitative and qualitative terms.

The task of single-image superresolution is to obtain a high resolution image from a low resolution input. Many strategies have achieved considerable success by learning from image databases so as to find suitable replacements for the missing information. However, little effort has been devoted to coping with significant amounts of noise in the input. Noise makes the problem even harder, since its distribution is unknown in many practical cases. Here an approach is proposed to solve this problem irrespective of the noise type, which supports both integer and fractional zoom factors. It is based on the application of the median filter on parallelogram shaped windows chosen according to a suitable probability distribution. The resulting outputs are then median filtered again to obtain the final output. Experimental results are reported for artificial and natural images under Gaussian and impulsive uniform noise. It is found that the proposed approach outperforms several state-of-the-art single-image superresolution methods both quantitatively and qualitatively.

The selection of the best color space is a fundamental task in detecting foreground objects on scenes. In many situations, especially on dynamic backgrounds, neither grayscale nor RGB color spaces represent the best solution to detect foreground objects. Other standard color spaces, such as YCbCr or HSV, have been proposed for background modeling in the literature; although the best results have been achieved using diverse color spaces according to the application, scene, algorithm, etc. In this work, a color space and color component weighting selection process is proposed to detect foreground objects in video sequences using self-organizing maps. Experimental results are also provided using well known benchmark videos.

Growing hierarchical self-organizing models are characterized by the flexibility of their structure, which can easily accommodate for complex input datasets. However, most proposals use the Euclidean distance as the only error measure. Here we propose a way to introduce Bregman divergences in these models, which is based on stochastic approximation principles, so that more general distortion measures can be employed. A procedure is derived to compare the performance of networks using different divergences. Moreover, a probabilistic interpretation of the model is provided, which enables its use as a Bayesian classifier. Experimental results are presented for classification and data visualization applications, which show the advantages of these divergences with respect to the classical Euclidean distance.

The estimation of multivariate probability density functions has traditionally been carried out by mixtures of parametric densities or by kernel density estimators. Here we present a new nonparametric approach to this problem which is based on the integration of several multivariate histograms, computed over affine transformations of the training data. Our proposal belongs to the class of averaged histogram density estimators. The inherent discontinuities of the histograms are smoothed, while their low computational complexity is retained. We provide a formal proof of the convergence to the real probability density function as the number of training samples grows, and we demonstrate the performance of our approach when compared with a set of standard probability density estimators.

The quality of self-organizing maps is always a key issue to practitioners. Smooth maps convey information about input data sets in a clear manner. Here a method is presented to modify the learning algorithm of self-organizing maps to reduce the number of topology errors, hence the obtained map has better quality at the expense of increased quantization error. It is based on avoiding maps that self-intersect or nearly so, as these states are related to low quality. Our approach is tested with synthetic data and real data from visualization, pattern recognition and computer vision applications, with satisfactory results.

In this paper, a novel methodology for using feature selection in hand biometric systems, based on genetic algorithms and mutual information is presented. A hand segmentation algorithm based on adaptive threshold and active contours is also applied, in order to deal with complex back grounds and non-homogeneous illumination. The aim of this methodology is two-fold. On the one hand, getting robust features in biometric systems with no restriction in the hand-pose and in its orientation with regard to the camera. On the other hand, providing a subset of features which reduce the complexity of the identification process and maximize the generalization rate of the classifiers. By using the IITD Palmprint Database, which is an example of such free hand-pose biometric systems, the experimental results show that it is not always necessary to apply sophisticated classification methods to obtain good accuracy results. Simple classifiers such as kNN and LDA together with this feature selection approach, get even better generalisation rates than other more elaborate and complex methods.

Background modeling and foreground detection are key parts of any computer vision system. These problems have been addressed in literature with several probabilistic approaches based on mixture models. Here we propose a new kind of probabilistic background models which is based on probabilistic self-organising maps. This way, the background pixels are modeled with more flexibility. On the other hand, a statistical correlation measure is used to test the similarity among nearby pixels, so as to enhance the detection performance by providing a feedback to the process. Several well known benchmark videos have been used to assess the relative performance of our proposal with respect to traditional neural and non neural based methods, with favourable results, both qualitatively and quantitatively. A statistical analysis of the differences among methods demonstrates that our method is significantly better than its competitors. This way, a strong alternative to classical methods is presented.

Since the introduction of the growing hierarchical self-organizing map, much work has been done on self-organizing neural models with a dynamic structure. These models allow adjusting the layers of the model to the features of the input dataset. Here we propose a new self-organizing model which is based on a probabilistic mixture of multivariate Gaussian components. The learning rule is derived from the stochastic approximation framework, and a probabilistic criterion is used to control the growth of the model. Moreover, the model is able to adapt to the topology of each layer, so that a hierarchy of dynamic graphs is built. This overcomes the limitations of the self-organizing maps with a fixed topology, and gives rise to a faithful visualization method for high-dimensional data.

The original self-organizing feature map did not define any probability distribution on the input space. However, the advantages of introducing probabilistic methodologies into self-organizing map models were soon evident. This has led to a wide range of proposals which reflect the current emergence of probabilistic approaches to computational intelligence. The underlying estimation theories behind them derive from two main lines of thought: the expectation maximization methodology and stochastic approximation methods. Here, we present a comprehensive view of the state of the art, with a unifying perspective of the involved theoretical frameworks. In particular, we examine the most commonly used continuous probability distributions, self-organization mechanisms, and learning schemes. Special emphasis is given to the connections among them and their relative advantages depending on the characteristics of the problem at hand. Furthermore, we evaluate their performance in two typical applications of self-organizing maps: classification and visualization.

Robustness against outliers is a desirable property of any unsupervised learning scheme. In particular, probability density estimators benefit from incorporating this feature. A possible strategy to achieve this goal is to substitute the sample mean and the sample covariance matrix by more robust location and spread estimators. Here we use the L1-median to develop a nonparametric probability density function (PDF) estimator. We prove its most relevant properties, and we show its performance in density estimation and classification applications.

We present a new neural model which extends the classical competitive learning (CL) by performing a Probabilistic Principal Components Analysis (PPCA) at each neuron. The model also has the ability to learn the number of basis vectors required to represent the principal directions of each cluster, so it overcomes a drawback of most local PCA models, where the dimensionality of a cluster must be fixed a priori. Experimental results are presented to show the performance of the network with multispectral image data.

In this paper, we present a probabilistic neural model, which extends Kohonen's self-organizing map (SOM) by performing a probabilistic principal component analysis (PPCA) at each neuron. Several SOMs have been proposed in the literature to capture the local principal subspaces, but our approach offers a probabilistic model while it has a low complexity on the dimensionality of the input space. This allows to process very high-dimensional data to obtain reliable estimations of the probability densities which are based on the PPCA framework. Experimental results are presented, which show the map formation capabilities of the proposal with high-dimensional data, and its potential in image and video compression applications.

We present a new neural model that extends the classical competitive learning by performing a principal components analysis (PCA) at each neuron. This model represents an improvement with respect to known local PCA methods, because it is not needed to present the entire data set to the network on each computing step. This allows a fast execution while retaining the dimensionality-reduction properties of the PCA. Furthermore, every neuron is able to modify its behavior to adapt to the local dimensionality of the input distribution. Hence, our model has a dimensionality estimation capability. The experimental results we present show the dimensionality-reduction capabilities of the model with multisensor images.

This paper proposes a new method to split images into regions. It consists of two subsystems: cluster detection and cluster fusion. The cluster detection is performed by a competitive neural network or the k-means algorithm, followed by an algorithm which obtains connected clusters. The cluster fusion involves a procedure that is based on the theory of equivalence relations. Proofs are given for the significant properties that we have found. It is not necessary to specify the number of regions in advance, which is a significant improvement over the standard competitive-style strategies. Finally, simulation results are given to demonstrate the performance of this method for some images.