Picking Up Quantization Steps for Compressed Image Classification

The sensitivity of deep neural networks to compressed images hinders their usage in many real applications, which means classification may fail just after taking a screenshot and saving it as file. In this paper, we argue that neglected disposable coding parameters stored files could be picked up reduce the images. Specifically, resort using one representative parameters, quantization steps, facilitate image classification. Firstly, based on propose novel aware confidence (QAC), is utilized sample weights influence network training. Secondly, utilize steps alleviate variance feature distributions, where batch normalization (QABN) proposed replace networks. Extensive experiments show method significantly improves performance CIFAR-10, CIFAR-100, ImageNet. code released https://github.com/LiMaPKU/QSAM.git