Shufflenet-E: Efficient Lightweight Network Design

Currently, neural network domain mainly focus on improvements for large models, with relatively less research on optimizing the optimization of lightweight models. Thus, we propose improving efficiency of lightweight models through optimizing fully connected layers, the performance of activation functions and introducing the Squeeze-and-Excitation block. These optimizations were applied to ShuffleNet V2, and the optimized model was named ShulleNet-E, it achieves up to a 27.47% reduction in parameter count, a 6.74% reduction in latency and a 1.1% improvement in precision. The results also outperform other lightweight models in performance or efficiency.