Micro-YOLO: Exploring Efficient Methods to Compress CNN based Object Detection Model

Deep learning models have made significant breakthroughs in the performance of object detection. However, in the traditional models, such as Faster R-CNN and YOLO, the size of these networks make it too difficult to be deployed on embedded mobile devices due to limited computation resources and tight power budgets. Hence, we propose a new light-weight CNN based object detection model, Micro-YOLO based on YOLOv3-Tiny, which achieves a signification reduction in the number of parameters and computation cost while maintaining the detection performance. We propose to replace convolutional layers in the YOLOv3-tiny network with the Depth-wise Separable convolution (DSConv) and the mobile inverted bottleneck convolution with squeeze and excitation block (MBConv), and design a progressive channel-level pruning algorithm to minimize the number of parameters and maximize the detection performance. Hence, the proposed MicroYOLO network reduces the number of parameters by 3.46× and multiply-accumulate operation (MAC) by 2.55× while slightly decreases the mAP evaluated on the COCO dataset by 0.7%, compared to the original YOLOv3-tiny network.