Abstract: Object detection algorithms based on convolutional neural networks are difficult to deploy onboard due to their high storage complexity and computational complexity. To address the problems a structured strip pruning algorithm to achieve object detection model compression is proposed in this paper. First, the strip pruning algorithm is used to obtain the convolutional kernel skeleton matrix. Then the elements in the skeleton matrix are sorted and the smaller elements in the corresponding skeleton matrix are pruned according to the pruning ratio to make the convolutional kernel structured. Finally, a mixed precision training method is used to obtain the structured strip pruning model. The proposed structured strip pruning algorithm is evaluated on the NWPU VHR 10 dataset and our dataset respectively. The algorithm can achieve a parameter compression rate of 1.97 times and a speedup ratio of 1.68 times, and the mAP decreases only 0.9% on the NWPU VHR 10 dataset and 1.7% on our dataset. Experimental results show that the structured pruning method can effectively achieve the pruning of the object detection model.

Key words: object detection, convolution neural network, deep learning, model compression, pruning