Investigating nonlinear point processing in image enhancement

Abstract

Today, the use of convolutional neural network models for deep learning in the field of image processing has become a general trend. However, to obtain satisfying performance, these networks usually stack a dozen of layers or require multiple iterations, which not only increase their complexity but also weaken their interpretability. The reason behind such a phenomenon is that, the key transformations in current networks are mainly realized by modules with limited nonlinearity, such as convolutions plus ReLUs. Some simple nonlinear processing like point processing is highly effective for image enhancement for our visual perception, and histogram equalization is one of the most common methods among them. By studying the principles of the traditional method of histogram equalization and introducing it into CNNs, it may be possible to improve the performance of the latter. In this report, we will first understand the principles and functions of several commonly used image point processing methods. Then, we'll dive into several different histogram equalizations and use them in MATLAB. Finally, we will build a CNN to compensate the results of HE to the desired output. We will see how the image pattern extraction capability of CNN can help HE images to be closer to the target image, and what are the advantages and disadvantages of this approach.