Machine learning and pattern recognition models in change detection

Change detection can be roughly defined as the awareness of change within an environment. The ability to detect change is vital in much of our everyday life—for example, noticing an activity change in a heartbeat pulse rate, in a brain EEG, in a vibration part of an electromechanical system, or simply in a highway lane during driving. However, one needs to bear in mind that change detection has a great value because it generates a state of interest—something is happening. Even if change detection is ubiquitous, it still remains a difficult enterprise. In fact, behavioral research suggests that human beings are very poor at detecting change, at least under certain circumstances. In particular, when attention is directed elsewhere a normally obvious change can often go undetected. This failure to detect change can have serious consequences, especially in such circumstances as driving, air traffic control, and medical diagnosis. Thus, machine learning techniques for detection of change can provide invaluable assistance to many human endeavors. Perceiving a change, stating what the change is and pinpointing it (where is it located?) are three activities embedded in the change detection phase. Change detection takes also an essential part of image or video analysis when applied to diverse applications, including remote sensing (e.g., evaluating changes in a forest ecosystems over a long period of time), surveillance (e.g., detecting an abandon objects or a moving object whose behavior deviates from what is normally observed), and medical diagnosis (e.g., inspecting signals from ECG or functional MR images). Detecting changes in a continuous speech or a handwritten script reveal vital discriminative clues that significantly enhance recognition and identification. Our world has never been as highly connected as it is today; cloud computing is emerging as a necessary pathway to information management. Electronic devices such as desktops, smartphones, notebooks, or personal digital assistants and tablets have become necessary and interchangeable means to run human affairs. Securing the information flow exchanged between these computing systems has been a challenge in the past years. Change detection in the information flow contributes significantly to live up to this challenge. However, there are many other scientific areas in which change tracking success is a tremendous achievement. For example, recently epigenetic changes have been proven to be linked to the development and progression of disease such as psychiatric disorders. Detecting these changes will have a profound impact in the prevalence of these diseases. Furthermore, the apprehension of changes in some geographical and physical features allows for a better preparation against natural disasters such as earthquakes and tsunamis. It is just lately that the numerous mathematical paradigms and formalisms attempted to model changes have found an area of agreement. In fact, these techniques started to converge just a while ago because of our deeper understanding of this process. Statistical pattern recognition techniques bring novel and powerful means to address change detection. These techniques allow us to gain insight into the complexity of change detection; their limitations and powers will be thoroughly scrutinized. Time series analysis and identification, model selection, statistical hypothesis testing, statistical approximate inference such as variational Bayesian methods, density estimation techniques, Bayesian networks and graphical models represent invaluable tools to process sequential data and decide on the change detectability. In fact, time series analysis and identification is one of several investigations that will benefit change detection. The main hypotheses relating to these investigations are that the latent parameters characterizing the data may not be subject to changes or are slowly time-varying. Moreover, many practical problems such as change detection in visual cortex for scene perception, quality control, recognition-oriented signal processing, fault detection and monitoring in industrial plants, can be modeled via statistical models whose parameters are liable to abrupt changes at any unknown point in time. The term abrupt change refers to changes in properties that occur very fast with respect to the measurements sampling period. For example, in the case of brainmachine interface, brain stimuli are expected to be heterogeneous, at the same time the device (e.g., wheel chair used by paraplegic) actuation process should be very fast. The mission of this special issue is twofold: (i) it promotes formalisms that exploit machine learning and pattern recognition state-of-the-art models to detect changes, and (ii) it emphasizes applications and contexts in which change detection is deemed necessary. This special issue offers a comprehensive, consolidated, and timely state-of-the-art perspective of the areas that could be of high interest to researchers, practitioners and students. The response to the call for papers was very keen; we have received very high quality submissions (around 20) even if the topic is challenging. The guest editors have finally accepted 9 papers based on a thorough and a comprehensive review process. Because change detection is our inherently omnipresent in our daily life, the spectrum covered in this special issue conveys both theoretical and practical issues. It covers newborn cardiology, ambient-assisted living, wrinkles detection using image morphology and geometric constraints, local image descriptors, polarimetric SAR data change detection in remote sensing, detection of dynamic changes on cyclic time series, video surveillance, video projection and 3D reconstruction. With a bunch of submissions, many established scientists in this field have been involved in the review process mission. We were fortunate to have them aboard and constitute a task force of experts. They have provided detailed, thoughtful, and timely reviews that leaped the