In this paper, we propose two contributions to neural network based denoising. First, we propose applying separate convolutional layers to each sub-band of discrete wavelet transform (DWT) as opposed to the common usage of DWT which concatenates all sub-bands and applies a single convolution layer. We show that our approach to using DWT in neural networks improves the accuracy notably, due to keeping the sub-band order uncorrupted prior to inverse DWT. Our second contribution is a denoising loss based on top k-percent of errors in frequency domain. A neural network trained with this loss, adaptively focuses on frequencies that it fails to recover the most in each iteration. We show that this loss results into better perceptual quality by providing an image that is more balanced in terms of the errors in frequency components.

Binarization of neural networks' activations may be a requirement for some applications. A typical example is end-to-end learned deep image compression systems where the encoder's output is requred to be a binary vector. Binarization is non-differentiable, therefore one needs to approximate it in order to train neural networks with stochastic gradient descent. In this paper, we investigate these training strategies and provide improvements over baselines. We find that during training, constraining the activations in a region that is far away from binary points leads to a better performance at test-time. The above finding provides a counter-intuitive result and leads to re-thinking the binarization approximation problem in neural networks.

There is an emerging interest aiming at defining principles for signals on general graphs, which are analogous to the basic principles in traditional signal processing. One example is the Graph Fourier Transform which aims at decomposing a graph signal into its components based on a set of basis functions with corresponding graph frequencies. It has been observed that most of the important information of a graph signal is contained inside the low frequency band, which leads to several applications such as denoising, compression, etc. In this paper, we show that the low frequency basis functions span the salient regions in an image, which can also be considered as important regions. Motivated by this, we present a novel simple and unsupervised method to utilize a number of low-energy basis functions and show that it improves the performance of seven state-of-the-art salient object detection methods in five datasets under four different evaluation criteria, with only minor exceptions.

Clustering is essential to many tasks in pattern recognition and computer vision. With the advent of deep learning, there is an increasing interest in learning deep unsupervised representations for clustering analysis. Many works on this domain rely on variants of auto-encoders and use the encoder outputs as representations/features for clustering. In this paper, we show that an l2normalization constraint on these representations during auto-encoder training, makes the representations more separable and compact in the Euclidean space after training. This greatly improves the clustering accuracy when k-means clustering is employed on the representations. We also propose a clustering based unsupervised anomaly detection method using l2normalized deep auto-encoder representations. We show the effect of l2normalization on anomaly detection accuracy. We further show that the proposed anomaly detection method greatly improves accuracy compared to previously proposed deep methods such as reconstruction error based anomaly detection.

Computer vision algorithms with pixel-wise labeling tasks, such as semantic segmentation and salient object detection, have gone through a significant accuracy increase with the incorporation of deep learning. Deep segmentation methods slightly modify and fine-tune pre-trained networks that have hundreds of millions of parameters. In this work, we question the need of having such memory demanding networks for a reasonable performance in salient object segmentation. To this end, we propose a way to learn a memory-efficient network from scratch by training it only on salient object detection datasets. Our method encodes images to gridized superpixels that preserve both the object boundaries and the connectivity rules of regular pixels. This representation allows us to use convolutional neural networks that operate on regular grids. By using these encoded images, we train a memory-efficient network using only 0.048% of the number of parameters that a majority of other deep salient object detection networks have. Our method shows comparable accuracy with the state-of-the-art deep salient object detection methods and provides a much more memory-efficient alternative to them. Due to its easy deployment and small size, such a network is preferable for applications in memory limited IoT devices.

In this paper, we provide a novel data set designed for Camera-independent color constancy research. Camera independence corresponds to the robustness of an algorithm's performance when it runs on images of the same scene taken by different cameras. Accordingly, the images in our database correspond to several laboratory and field scenes each of which is captured by three different cameras with minimal registration errors. The laboratory scenes are also captured under five different illuminations. The spectral responses of cameras and the spectral power distributions of the laboratory light sources are also provided, as they may prove beneficial for training future algorithms to achieve color constancy. For a fair evaluation of future methods, we provide guidelines for supervised methods with indicated training, validation, and testing partitions. Accordingly, we evaluate two recently proposed convolutional neural network-based color constancy algorithms as baselines for future research. As a side contribution, this data set also includes images taken by a mobile camera with color shading corrected and uncorrected results. This allows research on the effect of color shading as well.

We present a novel approach for spatiotemporal saliency detection by optimizing a unified criterion of color contrast, motion contrast, appearance, and background cues. To this end, we first abstract the video by temporal superpixels. Second, we propose a novel graph structure exploiting the saliency cues to assign the edge weights. The salient segments are then extracted by applying a spectral foreground detection method, quantum cuts, on this graph. We evaluate our approach on several public datasets for video saliency and activity localization to demonstrate the favorable performance of the proposed video quantum cuts compared to the state of the art.

In this paper, a computational color constancy method is proposed via estimating the illuminant chromaticity in a scene by pooling from many local estimates. To this end, first, for each image in a dataset, we form an image pyramid consisting of several scales of the original image. Next, local patches of certain size are extracted from each scale in this image pyramid. Then, a convolutional neural network is trained to estimate the illuminant chromaticity per-patch. Finally, two more consecutive trainings are conducted, where the estimation is made per-image via taking the mean (1sttraining) and median (2ndtraining) of local estimates. The proposed method is shown to outperform the state-of-the-art in a widely used color constancy dataset.

Object proposals improve the efficiency of object detection by providing probable locations of objects in an image. Most of the state-of-the-art object proposal methods employ a supervised approach and learn object features from ground truth annotations. We present a novel unsupervised approach for generating object proposals that is based on the human visual system and quantum mechanical principles. Despite of being devoid of any learnt priors pertaining to objects in images, the proposed method is shown to yield competitive results with supervised approaches.

In this paper, we propose a graph affinity learning method for a recently proposed graph-based salient object detection method, namely Extended Quantum Cuts (EQCut). We exploit the fact that the output of EQCut is differentiable with respect to graph affinities, in order to optimize linear combination coefficients and parameters of several differentiable affinity functions by applying error backpropagation. We show that the learnt linear combination of affinities improves the performance over the baseline method and achieves comparable (or even better) performance when compared to the state-of-the-art salient object segmentation methods.

In this study, we propose an unsupervised, state-of-the-art saliency map generation algorithm which is based on a recently proposed link between quantum mechanics and spectral graph clustering, Quantum Cuts. The proposed algorithm forms a graph among superpixels extracted from an image and optimizes a criterion related to the image boundary, local contrast and area information. Furthermore, the effects of the graph connectivity, superpixel shape irregularity, superpixel size and how to determine the affinity between superpixels are analyzed in detail. Furthermore, we introduce a novel approach to propose several saliency maps. Resulting saliency maps consistently achieves a state-of-the-art performance in a large number of publicly available benchmark datasets in this domain, containing around 18k images in total.

In this paper, a real-time railway fastener detection system using a high-speed laser range finder camera is presented. First, an extensive analysis of various methods based on pixel-wise and histogram similarities are conducted on a specific railway route. Then, a fusing stage is introduced which combines least correlated approaches also considering the performance upgrade after fusing. Then, the resulting method is tested on a larger database collected from a different railway route. After observing repeated successes, the method is implemented on NI LabVIEW and run real-time with a high-speed 3-D camera placed under a railway carriage designed for railway quality inspection.

In this study, the link between quantum mechanics and graph-cuts is exploited and a novel saliency map generation and salient object segmentation method is proposed based on the ground state solution of a modified Hamiltonian. First, the graph representation of certain quantum mechanical operators is studied. This reveals strong connections with widely used graph-cut algorithms while quantum mechanical constraints exhibit crucial advantages over the existing graph-cut algorithms. Furthermore, concepts such as potential field helps solving a particular singularity problem related to Laplacian matrices. In the proposed approach, the ground state (wave function) corresponding to a sub-atomic particle of a modified Hamiltonian operator corresponds to a particular optimization problem, the solution of which yields the salient object segmentation in a digital image. This approach provides a parameter-free -hence dataset independent-, unsupervised and fully automatic saliency map generation, which outperforms many existing state-of-the-art algorithms. The results of the proposed salient object extraction method exhibit such a promising accuracy that pushes the frontier in this field to the borders of the input-driven processing only - without the use of "object knowledge" aided by long-term human memory and intelligence. Furthermore, with the novel technologies for measuring a quantum wave function, the proposed method has a unique potential: Salient object segmentation in an actual physical setup in nano-scale. Such an unprece-dendent property will not only produce segmentation results instantaneously, but may be a unique opportunity to achieve accurate object segmentation in real-time for the massive visual repositories of today's "Big Data".

An automatic object extraction method is proposed exploiting the rich mathematical structure of quantum mechanics. First, a novel segmentation method based on the solutions of Schrödinger's equation is proposed. This powerful segmentation method allows us to model complex objects and inherent structures of edge, shape, and texture information along with the grey-level intensity uniformity, all in a single equation. Due to the large amount of segments extracted with the proposed method, the selection of the object segment is performed by maximizing a regularization energy function based on a recently proposed sub-segment analysis indicating the object boundaries. The results of the proposed automatic object extraction method exhibit such a promising accuracy that pushes the frontier in this field to the borders of the input-driven processing only — without the use of “object knowledge” aided by long-term human memory and intelligence.

In aerial images, the performance of the segmentation and object recognition algorithms could suffer due to shadows in the scene. This effort describes a novel shadow restoration algorithm based on atmospheric effects and characteristics of sun light for aerial images. Firstly, shadow regions are detected exploiting the Rayleigh scattering phenomena and the well-known fact related to the low illumination intensity in the shadow regions. After detection, shadow restoration is achieved by first restoring partially occluded shadow areas, as a result of modeling these transition regions with a continuous function that considers shadow formations. Next, fully occluded shadow regions are restored by first segmenting the image into multiple uniformly illuminated regions, then multiplying the intensity values in these regions with a constant, which is determined by the ratio of intensities between each segment and its non-shadow neighborhood. The simulation results indicate improvements over similar work from the literature.