An Improved Quantization Algorithm For Electric Power Inspection

Abstract

Along with the rapid development of deep learning and artificial intelligence technology, machine vision instead of human vision is used in many applications, such as defect detection, face recognition, etc. In the power system, it is very important to use computer vision algorithms to perform intelligent defect detection on images taken by Unmanned Aerial Vehicles, robots and other equipment. Traditional image compression algorithms aim to reduce the pixel distortion based on the visual characteristics of the human eye, which are not suitable for machine vision. Considering the limited communication channel bandwidth, the images generally compressed at low bitrate. Then, the blocking artifact introduced by compression will result in undetectable or false detection of the state of small electrical components, such as insulator strings, anti-vibration hammers, etc. In this paper, a quantization algorithm which is oriented to machine vision is proposed to address these problems. First of all, low-level image features are extracted based on CNN network. Subsequently, according to the extracted features, the original images are divided into key areas and non-key areas. Finally, key areas are encoded with smaller quantization level and non-key areas are encoded with lager quantization level. Experimental results demonstrate that compared with traditional image compression algorithm, the proposed algorithm improves the accuracy of defect detection from 20% to more than 80% at bitrate of 0.24bpp.