Optimized MobileNetV2 Based on Model Pruning for Image Classification

Due to the large memory requirement and a large amount of computation, traditional deep learning networks cannot run on mobile devices as well as embedded devices. In this paper, we propose a new mobile architecture combining MobileNetV2 and pruning, which further decreases the Flops and number of parameters. The performance of MobileNetV2 has been widely demonstrated, and pruning operation can not only allow further model compression but also prevent overfitting. We have done ablation experiments at CIIP Tire Data for different pruning combinations. In addition, we introduced a global hyperparameter to effectively weigh the accuracy and precision. Experiments show that the accuracy of 98.3 % is maintained under the premise that the model size is only 804.5 KB, showing better performance than the baseline method.