Change Detection Based on Binary Mask Enhancement

Introduction

Detecting changes in a video sequence or still images is a crucial task in the image processing domain, which aims to distinguish moving objects from static ones. This functionality has a lot of applicability in the image processing domain like mostly in the surveillance cameras in organizations both the government and private. This change detection has captured a lot of attention in the recent years due to its use and achievement in the domain.

Over the years, computer vision research has made considerable progress in developing change detection methodologies that can handle complex and challenging scenes. Classical methods such as frame differencing and Gaussian Mixture Models (GMM) have been extensively studied. However, in recent times, there has been a surge of research on novel approaches, resulting in significant advancements. Another promising technique is video object segmentation using deep learning, which has shown promising results.

As there is advancement in the methodologies and the algorithms used, still there is a lacking efficiency in these algorithms and because of that they cannot able to handle the challenging and complex scenes that contain interference and noise in it; for example, some of the situations can be considered here like the varying illumination, dynamic background and disturbances of shadows in the scenes etc. So due to this problem the researchers decided to bring a solution for the problem which can be some of the post processing methods that can enhance the results of the detection by lowering the noise and interference and enhance the contours in the foreground regions more accurate and precise nearly. Some of the methods such as morphological operators like dilation and erosion can perform well but at times, they become unstable and provide poor results.

The argument is that the likelihood of a pixel being part of the foreground is influenced by the quantity of foreground and background pixels present in the surrounding area. This can improve the detection of foreground pixels by calculating the probability of a pixel belonging to the foreground regions and has link to the neighboring pixels. To achieve this, a well-known energy function is defined, and a Markov Random Field model is developed to iteratively compute the posterior probabilities for foreground and background pixels. This process allows for the refinement of the foreground detection through successive iterations. But due to the complex computational process in the semantic image segmentation some of the post processing methods like local range Cumulative Relative Frequency (CRF) is used so that it can increase the credibility and accuracy of the results of enhancement.

Recent advances in image processing study have made advanced image filtering a hot topic. It is frequently used for stereo matching, picture haze removal, and image denoising. Image filtering is the process of separating pixel data from noise and recovering the true structure of the pixels. Typically, change detection produces a coarse binary mask which can be easily improved with sophisticated image filters. An image that acts as a different version of itself can be linked to the same image which can be given to a class of cutting-edge filters while filtering. For filtering the input picture, they consider the guidance image's information to be a trustworthy reference.

As image filtering enhances the change detection results, it is simply performed by some operations in the pixel values of an image. First, the working principle of two types of image filtering i.e., local and global filter are discussed. In the local image filter, the filtering process is done by moving a small window which is a kind of mask like a square matrix that slides the image pixels and captures the fine edge details but it fails to account for the long-range dependencies between the pixels. Whereas, the global filter allows each pixel in the image to change all the pixels in the vicinity that handles the problem of rich textures in the foreground region. But still the global filter lacks the capacity for preserving the edges. Guided filter and spatial tree filter are the examples of local and global filter respectively.