HRCP : High-ratio channel pruning for real-time object detection on resource-limited platform

Recently, deep learning algorithms have demonstrated remarkable performance in object detection tasks and have applied to many fields such as intelligent networked vehicles and drones. Different from other fields like video surveillance, the fields of intelligent networked vehicles and drones only have resource-limited platform and have high requirements in model storage and real-time inference. To satisfy these requirements, model compression approaches are developed to reduce model parameters and computation. As channel pruning is a coarse-grained hardware-friendly weight pruning method, it is widely used in model compression. Most studies have only focused on utilizing the scale factor of Batch Normalization (BN) layers to determine whether to prune a certain channel. However, the shift factor turns to very important when its value is large. The channel with large shift factor has a great influence on subsequent layers. In our paper, we take shift factor into consideration to protect these channels. We first introduce network building blocks to analyze YOLOv3-SPP3 model and demonstrate the importance of shift factor on the channel saliency in the pruning process. Then we present a new pruning method called high-ratio channel pruning (HRCP) to improve the performance of the pruned model based on definition of the channel saliency with scale factor and shift factor of BN layer. Our experimental results show that HRCP performs better than SlimYOLOv3 when model volumes are reduced by 50%, 88%, and 92% accordingly. In high-ratio pruning, our HRCP can be pruned more 9.7% volumes than SlimYOLOv3 under the same performance condition.