Binary Neural Networks With Feature Information Retention for Efficient Image Classification

Abstract

Although binary neural networks (BNNs) enjoy extreme compression ratios, there are significant accuracy gap compared with full-precision models. Previous works propose various strategies to reduce the information loss induced by the binarization process, improving the performance of binary neural networks to some extent. However, in this letter, we argue that few studies try to alleviate this problem from the structure perspective, resulting in inferior performance. To this end, we propose a novel Feature Information Retention Network named FIRNet, which incorporates an extra path to propagate the untouched informative feature maps. Specifically, the FIRNet splits the input feature maps into two groups, one of which is fed into the normal layers and another kept untouched for information retention. Then we utilize the concatenation, shuffle and pooling operations to process these features with 64× memory saving. Finally, with only a 1.7% complexity increase, a FIR fusion layer is proposed to aggregate the features from two branches. Experimental results demonstrate that our proposed method achieves 1.0% Top-1 accuracy improvement over the baseline model and outperforms other state-of-the-art BNNs on the ImageNet dataset.