Currently, in the efforts of consolidating models based on theories that simulate cognitive and learning human processes, the Meta-Learning paradigm (MeL) has emerged in the Machine Intelligence area. MeL has been applied in combination with other approaches for Machine Learning and Deep-Learning, for instance, to solve new tasks with few samples in the few-shot learning approach. However, to consider MeL as a well-defined learning paradigm aimed at successfully simulating learning human processes, it is essential to find a from to evaluate its learning-to-learn process (LTLP). Metrics reported in the literature to evaluate the LTLP are insufficient and shallow to understand the results reported by MeL-based models. The most common metric used to evaluate and compare MeL-based models is the metric accuracy. Therefore, this work proposes the use of the Centered Kernel Alignment (CKA) metric to evaluate the LTLP of MeL-based models for few-shot learning. The CKA metric helps to quantify how much of the meta-knowledge and meta-experience learned in the meta-training of the meta-models are used to solve new tasks, contributing to the understanding of the LTLP. A MeL methodology is implemented using 9 prior-models in the meta-training of meta-models. In the experiments of the meta-models, the CKA metric is used to analyze similarities between the meta-features obtained by each prior-model. Besides, a proposed meta-supervised algorithm uses the CKA metric to select the meta-parameters to solve new tasks in the base-domain. The results showed CKA scores of 0.674 and 0.799 (these scores mean the level in which the new base-tasks to solve are different from the meta-learned tasks), and accuracies of 92.27% and 99.73%, for solving the new base-tasks using the CUB and Mini-ImageNet datasets, respectively.

As the studies on Machine Intelligence continue, new methods emerge with the purpose of improving computational learning, especially in deep-learning to overcome the need of high amount of data used to solve tasks. Therefore, paradigms like Meta Learning for few-shot learning intents to work out this situation. Nevertheless, current meta-learning methods address this situation using a single model trained in different levels of learning: meta-level and base-level. Instead, a new methodology to implement meta-learning for few-shot learning is proposed in this paper. This methodology is based on the concepts of modularity and reuse of optimized and previously implemented models and it is divided into elements of the process of learning. The elements are prior-models, meta-model, few-shot-base-model and a meta-algorithm. The proposed meta-algorithm generates a distribution over trained meta-models. Then, the best meta-model whose meta-tasks are more similar to the new tasks that the few-shot-base-model will solve is chosen. Also, a new concept of meta-few-shot sampling is proposed in order to evaluate how many meta-tasks are used in the meta-training stage. The experiments with 5-way 5-shot configuration show a competitive result compared with state-of-the-art methods. The proposed method achieves 97.56±1.57% and 83.33±8.33%in the CUB-200-2011 dataset, using the EfficientNetB7 and the DenseNet121 as prior-models, respectively; and, 99.56±0.63% (EfficientNetB7) and 94.89±4.28% (DenseNet121) in the Mini-ImageNet dataset.

Birds of prey especially eagles and hawks have a visual acuity two to five times better than humans. Among the peculiar characteristics of their biological vision are that they have two types of foveae; one shallow fovea used in their binocular vision, and a deep fovea for monocular vision. The deep fovea allows these birds to see objects at long distances and to identify them as possible prey. Inspired by the biological functioning of the deep fovea a model called DeepFoveaNet is proposed in this paper. DeepFoveaNet is a convolutional neural network model to detect moving objects in video sequences. DeepFoveaNet emulates the monocular vision of birds of prey through two Encoder-Decoder convolutional neural network modules. This model combines the capacity of magnification of the deep fovea and the context information of the peripheral vision. Unlike algorithms to detect moving objects, ranked in the first places of the Change Detection database (CDnet14), DeepFoveaNet does not depend on previously trained neural networks, neither on a huge number of training images for its training. Besides, its architecture allows it to learn spatiotemporal information of the video. DeepFoveaNet was evaluated in the CDnet14 database achieving high performance and was ranked as one of the ten best algorithms. The characteristics and results of DeepFoveaNet demonstrated that the model is comparable to the state-of-the-art algorithms to detect moving objects, and it can detect very small moving objects through its deep fovea model that other algorithms cannot detect.

In recent years, motor imagery has been used as a communication alternative in brain computer interface systems. In this paper an evaluation of five different types of feature extraction methods to recognize two motor imagery signals based on common spatial patterns (CSP), and time-frequency transformations are evaluated; CSP, continuous Wavelet transform (CWT), Stockwell transform (ST), and the following combinations CSP+CWT and CSP+ST. The classifier employed to recognize between right-hand, and left-hand was a support vector machine. The proposed methods were evaluated in two know datasets and compared with other state of the art methods. The best performance achieved with the proposed methods, considering a correct recognition rate, was 79.87% ± 10.73% and 73.25% ± 08.04% with the datasets BCI Competition IV dataset 2a and EEGdataset, showing better performances than the reported works. Besides, the CSP+SVM method requires a processing time of only 1.73 seconds which make it suitable for real time applications.

This study proposes a new method to evaluate the performance of algorithms for moving object detection (MODA) in video sequences. The proposed method is based on human performance metric intervals, instead of ideal metric values (0 or 1) which are commonly used in the literature. These intervals are proposed to establish a more reliable evaluation and comparison, and to identify areas of improvement in the evaluation of MODA. The contribution of the study includes the determination of human segmentation performance metric intervals and their comparison with state‐of‐the‐art MODA, and the evaluation of their segmentation results in a tracking task to establish the impact between performance and practical utility. Results show that human participants had difficulty with achieving a perfect segmentation score. Deep learning algorithms achieved performance above the human average, while other techniques achieved a performance between 88 and 92%. Furthermore, the authors demonstrate that algorithms not ranked at the top of the quantitative metrics worked satisfactorily in a tracking experiment; and therefore, should not be discarded for real applications.

This paper presents a new motion detection algorithm called Fuzzy Gray Difference Histogram (FGDH). The method consists of three main stages. The first stage obtains the gray level differences of a pixel centered in a LxL window with respect to its neighbors by the local binary similarity patterns (LBSP) operator. Then, these differences are fuzzified by a Gaussian function that carries out a weighting from 0 to 1, which indicates the degree of similarity of the neighborhood pixels with respect to the central one. The second stage obtains a background model based on fuzzy local histograms that are made from small neighborhoods in order to provide spatial and gray level information. The third stage updates the background histograms based on the similarity between the current and past histograms. Experimental results indicate that FGDH presents a competitive performance in the Change detection database considering the elimination of shadows and noise in video scenarios from the categories Baseline, Badweather and Shadow. The implementation was attained under an optimization scheme that reduced the computational cost and a parallelizable scheme of the method which allows real time speed.

Image complexity has been estimated by computational algorithms that try to simulate the human criterion to determine complexity. Image complexity has been used to study the behavior of the human brain, as well as in the area of image processing. This paper presents a work to analyze image complexity. A fuzzy inference system using principal component analysis, PCA, is designed to model the human criterion. The PCA characteristics are obtained from the most used features reported in the literature; contrast, correlation, energy, homogeneity, frequency factor, edge density, compression ratio, number of regions, colorfulness, number of colors, and color harmony. The work was achieved with the data based RS1 and RS2. The results obtained by cross validation, demonstrate a correlation level of 0.8566 between the proposed method and the human criterion.

This paper describes the evaluation of the Auto-Adaptive Parallel Neural Network Architecture, AAPNNA, in the SBMnet dataset. AAPNNA is an artificial neural model based on two networks whose neurons represent two different Background models that adapt their parameters at different rates. A very important feature of AAPNNA is its capacity to auto adapt to new scenario conditions as demonstrated with the results of the Illumination Change category. AAPNNA presented good qualitative results in several of the videos available for comparison. The overall quantitative performance with the complete SBMC 2016 dataset can be considered as an acceptable performance with the evaluation of all the videos and the performance interval of each metric.

A depth video background modeling method based on a cascade retinotopic map is presented in this paper. The proposed scheme is intended to be used in a virtual therapy system to attend children with different restricted arm movement. The proposed method involves two retinotopic maps with different background modeling capacities. The neural network scheme was tested under several laboratory and real world therapy scenarios. Regarding the findings, the model achieved an acceptable performance on both types of scenarios, and showed its potential to solve bootstrapping conditions, deal with different patient body positions, and visual interference commonly found in therapy sessions.

Motion detection represents a challenging issue in artificial vision systems. Besides detection of movement in normal scenario conditions robust systems must deal with other non-normal conditions. We propose the improvement of a former neuro-fuzzy motion detection method to face drastic illumination changes, gradual illumination conditions, moving background and scene composition changes. The improvements include adaptive learning rates as well as the inclusion of new fuzzy rules. Experimental findings over several video sequences verify that the improvements outperform the performance of the original method in the non-normal conditions without affecting the performance under normal conditions.

Self Organizing Maps (SOMs) are neural networks that have been widely focused in computer vision applications and simulation of visual cortex areas. From among those applications, there are successful works related to neuroinspired motion processing. In this work, we propose the Retinotopic SOM (RESOM); a neural network based on Self-Organizing Retinotopic Maps, which was applied in dynamic segmentation using background modeling. Every neuron in the network has a set of retinotopic weigths similar to the projections from the retina to primary visual cortex. The hebbian learning of the RESOM makes in every neuron a global modeling of a frame from a video sequence, causing different reconstructions of the frame where the common pattern learned, defined in all neurons, is the video static information. Thus, the background is modeled finding the expected value of all neurons and inhibit the differences in the pattern weights of the neurons. In order to obtain a dynamic segmentation, we use the background subtraction method. Experimental results over real videos taken with stationary cameras showed that the RESOM has good performance to segment dynamic objects in video sequences and it is robust to illumination changes.

Background determination is crucial to visual intelligent surveillance systems. Although several methods have been proposed in the literature, research on this topic is still a paramount objective in the surveillance system community. High performance and low computational cost in a video segmentation model are some of the characteristics of the segmentation model presented in this paper. The model is designed to work with semi-static backgrounds. The segmentation model is based on a SOM like architecture. Weights neuron updates are performed in the fly to provide dynamic background actualization. The model keeps simplicity but it is tolerant to background variations like illumination, shadows, and slow moving background regions. The method was tested in several scenarios, including daytime and night situations, as well as interior and exterior scenarios. Qualitative and quantitative results of the model show high performance for normal backgrounds, and acceptable performance on high dynamic backgrounds, compared with complex models reported in the literature.

In this paper a fuzzy clustering approach for the classification of cosmetic defects is presented. The paper investigates the solution of this classification problem with the Gustafson-Kessel (GK), and Geth-Geva (GG) with Abonyi-Szeifert (AS) fuzzy algorithms. The clustering process is achieved on multidimensional feature vectors that represent the cosmetic defects. The performance of the GK algorithm may be considered similar to a human inspector which is between 85% and 90% approximately. However, the fuzzy clustering technique has the advantage to be very consistent, contrary to a human inspector that can change her/his mind due to subjective influences. The paper also presents the comparison between the fuzzy approach and the artificial neural network approach. The problem faced in this work also helped to compare the performance of FC algorithms with ANN in real world applications.

Image segmentation has attracted the attention of researcher for many decades. Different approaches have been developed in order to find the solution in many different segmentation situations. In this paper we propose a novel edge detection approach aimed to generate useful information to achieve segmentation. The proposed method is based on analysis of the information provided by the time matrix generated from a pulse coupled neural network, PCNN. This information represents gray level differences among the pixel images. Two different schemes for edge detection are presented. The first scheme is developed to generate edges from coarse images and the second one to deal with more detailed edges. Similarity of this method with a previous developed method based on fuzzy edge level detection is also covered in the paper. Final results show that the proposed method may be used as a new alternative to define image edges of different levels for further analysis.

This paper is concerned with the design of a classification system based on artificial neural networks to distinguish between natural and non-natural cosmetic defects found in ophthalmic lenses. Natural cosmetic defects are related to small cotton fabrics, and non-natural defects are formed during the fabrication process. A set of geometric, morphology and topologic features are defined in order to represent these defects. The recognition problem of theses defects is faced with feedforward and SOM artificial neural networks paradigms. The performance of the feedforward and SOM networks turned to be similar, 92.35% of correct classification. The performance of these neural networks is acceptable compared against the performance of a human inspector considering that a human inspector reaches a performance between 85% and 90%. Besides, the ANN approach is completely free of changes in its decision, contrary to a human inspector that can change his/her mind due to subjective influences.

Computational intelligence theories offer, individually, different potentials to solve real world problems. However, fusion of these potentials provides opportunities to generate more real world robust systems. Cosmetic inspection of possible non-uniform surfaces found in manufacturing is a challenge to human inspectors. This paper deals with the proposal of a new hybrid methodology to segment color images in order to detect non-uniform regions that may appear in manufactured goods. The hybrid methodology combines two fuzzy clustering algorithms, the FCM and the GG, and a SOM ANN. Because of its properties the FCM is used to find the optimal number of clusters of a sample population of non-uniform surfaces. This value is then used to initialize the GG algorithm to determine the best centroids that represents the color population. Finally a SOM is trained with the results of the GG to perform the segmentation. Findings show that the proposed methodology generates color regions in accordance to a quality inspection criterion. The proposed methodology is also compared against the performance of the FCM to show the advantages of the hybrid methodology.

This paper presents a case of study where two of the most used fuzzy clustering algorithms in pattern recognition tasks are analyzed under a classification problem that involves a high degree of subjectivity. The problem consists on the classification of seven types of wood defects called knots. The algorithms are the Abonyi-Szeifert modification of the Gath-Geva algorithm, GGAS, and the Gustafson-Keseel, GK. An improvement to the GK algorithm, GKM, is also proposed and analyzed. Besides the analysis of the algorithms, three different techniques are proposed to generate the design set of samples and the testing set of samples. Results of the study show that the GGAS and the GKM algorithms have a performance close to human performance.

This paper describes the implementation of a system based on Pulse Coupled Neural Networks (PCNNs) and Field Programmable Gate Arrays (FPGAs). The PCNN implemented is oriented to the industrial application of segmentation in sequences of images. The work went through several real physical stages of implementation and optimization to achieve the needed performance. The greatest performance achieved by the digital system was of 250M pixels per second, enough to process a sequence of images in real time. Details of these stages about the neuron implementation with different Altera's FPGAs families are presented. Furthermore, the implementation is compared with previous implemented schemes based on floating point DSP microprocessor.

This paper presents the solution of a specific classification problem classification of cosmetic defect found in ophthalmic lenses - using artificial neural networks (ANN). In an ordinary industrial inspection process cosmetic defect classification involves a lot of subjectivity because the test depends on the appreciation of the cosmetic defect by a human inspector. Therefore, a machine classifier is of great help in order to reduce the effect of the subjectivity. The neural network classifier described in this paper is of relevance because it demonstrates the applicability of ANN to solve real world problems. The classifier also helped to discover a bias criteria during the inspection performed by human inspectors. Besides, the performance of the neural network impacts positively in the cost of incorrect decisions during cosmetic lens inspection. The paper shows the performance of several ANN classifiers designed. The best classifier turned out to be a multilayer perceptron trained with the backpropagation algorithm. This classifier has 94% of correct classification and solves the controversies between the production and quality departments due to subjectivity of the inspection.

This paper describes a novel scheme for image processing based on the PCNN time matrix. The time matrix is a mapping from the spatial image information to time information generated by a PCNN. The time matrix contains useful information related to spatial information of the image that is under processing. Based on this matrix, image processing operations may be defined; image segmentation, edge detection, and selective filters. This paper slates this new image processing scheme based on the time matrix and illustrates the application of the scheme.

This paper describes a new dynamic ATR, automatic target recognition, scheme based on a pulse coupled neural network, PCNN. The PCNN is used to generate candidate regions that may contain a license car plate. Candidate regions are generated from pulsed images, output of the PCNN network. Statistics of the Fourier transform are used to determine if a candidate region contains a license plate. If the license plate is not located in the set of candidate regions, the PCNN network parameters are adjusted to generate a new pulsed image and new candidate regions are extracted and analyzed. The proposed system is robust in the sense that it does not restrict the location of the plate, neither the illumination conditions when the image is acquired. The purpose of no restrictions is to evaluate the proposed system under conditions where a person can succeed. The system performance was 85% on a set of images. One advantage of the proposed scheme is that it may be adapted to solve different ATR problems.

This paper presents a novel fuzzy method to define edges of interest in images under analysis. It is an adaptive method that automatically defines which edges will be shown according to the complexity of the image. Complexity is considered as a subjective characteristic that is represented by a fuzzy interpretation of the percentage of edges in an image. The edges that appear in the final result are selected through an integration of edge levels using fuzzy rules. These fuzzy rules were derived from the analysis of edge level components over a set of images. Results over these images indicate that the behavior of the method is adaptive according to the complexity of the image as it was stated in our theory and they turn to be better than other methods as shown in the paper. These results also prove that integration of information, edges, according to what a vision system sees can be used for subsequence stages of analysis making a vision system more robust.

This paper describes a fuzzy neural network classifier to perform document region classification using features obtained from human visual perception theories. The foundations of the classifier are derived from human visual perception theories. The theories analyzed are texture discrimination based on textons, and perceptual grouping. Based on these theories, the classification task is stated as a texture discrimination problem and is implemented as a preattentive process. Engineering techniques are then developed to extract features for deciding the class of information contained in the regions. The feature derived from the human visual perception theories is a measurement of periodicity of the blobs of the text regions. This feature is used to design a fuzzy neural network classifier. The results of this work may be a good support for the assertion that visual perception theories may be incorporated into engineering techniques to produce better results.