An Efficient and Fast Filter Pruning Method for Object Detection in Embedded Systems

초록

Recently, CNN-based networks have exhibited high performance in computer vision. On the other hand, due to the networks becoming deeper and wider, it is hard to implement the model in real-time embedded environments. To overcome the drawback, filter pruning has been widely studied for neural network compression. Filter pruning does not need any special hardware or software because it removes filters of CNN and accelerates inference without any special software or hardware. In this paper, we proposed efficient and fast filter pruning (EFFP), which focuses on reducing the training computation resources and searching optimal pruned networks. The success stems from two significant improvements upon other pruning methods. (1) Short training time: In the pruning stage, we make redundant filters to zero to make the output feature map the same as a lightweight model, and (2) adjust the change of redundancy using regrowing: It is difficult to get an optimal pruned model by pruning redundant filters at once. Therefore, we use the pruning/regrowing method to gradually remove unimportant filters to avoid permanently pruning important filters to get an optimal lightweight model. Experimental results indicate that EFFP can reduce the FLOPs and parameters more efficiently and faster than other pruning methods on the object detection model. The inference time is measured on NVIDIA Jetson Xavier NX. As a result, we improve mAP and inference time by a maximum of 45 % compared to other pruning methods.

키워드