Special issue on computer vision applying pattern recognition techniques

This Special Issue of the Elsevier Journal of Pattern Recognition entitled “Computer Vision applying Pattern Recognition Techniques” presents the ongoing and active research in the Pattern Recognition applied to Computer Vision area, with emphasis on topics around emerging paradigms on computer vision, robotics, and up-to-date applications to related fields (hardware implementations, parallels solutions). The overall goal of this Special Issue is to summarize in a handbook manner, recent discoveries and groundbreaking studies that will account for new research fields and disciplines in the broad area of pattern recognition based approaches to tackle computer vision gaps in knowledge. The call for papers resulted in 28 submissions. At least three reviewers assessed the quality of the papers, and those meeting the top standards were sent to a second round of reviews. Finally, 13 papers were selected for publication. The guest editors hope that the selected papers will provide the readers with interesting examples of current research on the most outstanding theoretical frameworks in the pattern recognition context applied to Computer Vision as well as in the challenging field of applications through practical and efficient algorithms. Let us provide first a general view of this Special Issue (SI). This issue has been compiled to demonstrate the growing interest for robust and innovative pattern recognition techniques to be applied to classical image processing problems, such as image segmentation, image denoising, image classification, camera calibration, edge extraction/detection for image and video analysis. A special feature of this SI is an extended discussion on Bayesian probabilistic schemes and restricted Boltzmann machines subject and its applications to computer vision. A different view, considering the human “in the loop”, as part of complex systems is included in this Special Issue in the form of human–computer interfaces relaying on pattern recognition techniques. Additionally, adaptive learning, data mining and new trends in graph theory and GPU processing, all applied to relevant Computer Vision problems, are also included in this Special Issue. Kiran and Serra open this Special Issue with the work entitled, “Global–Local Optimizations by Hierarchical Cuts and Climbing Energies”. In this manuscript, authors propose a new theoretical scheme dealing with a very relevant matter in Computer Vision: to find optimal cuts in hierarchies of images to practical useful applications such as color image segmentation and texture enhancement. Other important applications are also noted along the paper: detail segmentation and image compression. Another remarkable contribution of this article is the introduction of some new concepts: h-increase, singular and scale increasing energies, which are shown to play a relevant role within the global combinatorial problem of partition selection which results in linear time dynamic programs. The mechanism to ensemble new energies function is also addressed in the article. Along the manuscript, several algorithms map the sound theoretical framework into practical tools. Results on several images show the benefits of the proposal and open new fields of research combining hierarchical cuts with wavelet decomposition. Fischer and Igel, in the paper entitled, “Training Restricted Boltzmann Machines: An Introduction”, present a self-contained tutorial about the very challenging topic of Restricted Boltzmann Machines (RBM) from the probabilistic graphical model perspective, with emphasis on the learning algorithms and their theoretical justification through several theorems. There are not many tutorial papers in Deep Learning/RBMs, and thus the present paper explains recent algorithms, like parallel tempering and persistent contrastive divergence which are of the most interest for researchers on the pattern recognition community. In addition, the experiments section illustrates the behavior of RBMs in practice and shows some applications to notable computer vision problems. Oommen and Thomas propose in the work entitled, “Anti-Bayesian ⎕ Parametric Pattern Classification Using Order Statistics Criteria for Some Members of the Exponential Family” the use of order statistics (OS) for Pattern Recognition and apply the moments of order statistics to some probability density functions belonging to the exponential family of distributions. The new scheme, referred to as classification by moments of order statistics (CMOS), has an accuracy that attains Bayes ′ bound for Symmetric distributions, and is, otherwise, very close to the optimal Bayes′ bound, as has been shown both theoretically and by rigorous experimental testing. The results shown in this article also give a theoretical foundation for the families of border identification (BI) algorithms reported in the related literature. Negri et al. present the paper entitled, “Detecting Pedestrians on a Movement Feature Space” devoted to detecting pedestrians in surveillance video sequences taken from video surveillance-type fixed camera filming outdoors. As it is known, pedestrian detection is considered a key chapter of pattern recognition and computer vision. Given that images may suffer changes in the scene appearance or rapid changes in lighting, this is a challenging problem. The authors propose to divide the problem into two main sub-problems: first one consisting of detecting movement in the scene and then, checking whether the moving object (target) is a pedestrian. The movement detection part is solved through an innovative background model using level lines all embedded into a Color Texton Space. This allows for the retrieval of color transitions and generation Movement Feature Spaces (MFS). Then, several descriptors are assembled and processed by a cascade of boosted classifiers combining generative classification functions with discriminant functions in an efficiently manner to allow its use in real-time video surveillance applications.