Deep-learning Based Denoising and Enhancement in Video Image of Approach Channel

Abstract

Due to the ubiquitous video sensor technology, it has become possible to build intelligent information system for the supervision of the approach channel. However, at present, there is still a technical problem to be solved, that is, the low-visibility video images taken in low-light environment have low brightness, low contrast and low signal-to-noise ratio. Meanwhile, they are also mixed with a lot of noise, which makes data extraction difficult. We propose a video image denoising and enhancement network based on deep learning, named DeNet, to improve the quality of low-light video images. DeNet consists of two parts: the first part is a deep residual network, which we named D-Net, to denoise the original video image. In the second part, we connect a D-Net network in parallel and add the normalization layer, which is used to fuse the features of the denoised image with the original image, and eventually realize the denoising and enhancement function. Experimental results on synthetic and real port test data sets show that our proposed method is superior to many advanced methods at present, and can meet the requirements of video denoising and enhancement of the approach channel.