Efficient Residual Network Compression for Optimizing the Accuracy-Complexity Tradeoff

Image recognition algorithms based on deep learning techniques have played an important role in the military, medical, industrial, and many other applications. However, most existing deep neural networks consume excessive computational resources which are unaffordable for the widely used edge devices, such as mobile phones. In this paper, we propose a lightweight network CResNet based on ResNet-50 by combining efficient channel pruning with depthwise decomposition. Ablation experiments are carried out based on the Animals-10 dataset for measuring the impact of each adopted technique. Great compression performance of the model parameters can be achieved at the price of slightly lower accuracy. Eventually, CResNet results in 4.08 M parameters, which is only one-fifth of the parameter size of the original ResNet-50, sufficiently reducing resource consumption. Approximately 90.2% Top-1 classification accuracy estimated on Animals-10 can be achieved by our lightweight CResNet. Compared to ResNet-50 and many existing lightweight networks, this work achieves a better tradeoff between segmentation accuracy and computing complexity by optimizing the computational efficiency, resulting in a small model size and a decent accuracy.