Redundancy Removal Module for Reducing the Bitrates of Image Coding for Machines

With the popularity of the Internet of Things (IoT) and surveillance, the amount of image and video information collected by internet has exploded. On the one hand, transmitting the massive information requires plenty of bandwidth. On the other hand, processing these data requires tremendous manpower. In order to solve these two problems at the same time, image coding for machines (ICM) came into being. At present, most ICM technologies combine traditional codecs such as BPG or machine learning codecs such as the hyperprior codec and the coarse-to-fine codec with task networks such as image classification and semantic segmentation. This process requires the complete restoration of the image, which greatly increases the bitrates. Moreover, the restored images are for human eyes and contain a large amount of redundant information, which is not required by the task network. In order to solve this problem, side information driven image coding for machines (SIIC) was proposed, which only needs to input hyperprior information to the image classification network, which greatly reduces the scale of data transmission. Now, we propose a redundancy removal module that can further reduce the usage of bitrates for SIIC. Through this module, the new codec uses 3% to 4% less bitrates than the original SIIC when the bitrates are under 0.06 bpp, and can save up to 10% of bitrates when the bitrates are over 0.1 bpp.