Micro-YOLO+: Searching Optimal Methods for Compressing Object Detection Model Based on Speed, Size, Cost, and Accuracy

Convolutional neural networks play a great role in solving the problem of object detection. However, conventional object detection models, such as YOLO and SSD, are usually too large to be deployed on embedded devices due to their restricted resources and low power requirements. In this paper, several efficient methods are explored to balance model size, network accuracy, and inference speed. We explore elective lightweight convolutional layers to supplant the convolutional layers (Conv) in the YOLOv3-tiny network, including the depth-wise separable convolution (DSConv), the mobile inverted bottleneck convolution with squeeze and excitation block (MBConv) and the ghost module (GConv). Moreover, we explore the optimal hyper-parameters of the network and use the improved NMS algorithm, Cluster-NMS. Moreover, a new object detection model, Micro-YOLO+, which achieves a signification reduction in the number of parameters and computation cost while maintaining the performance is proposed. Our Micro-YOLO+ network reduces the number of parameters by 3.18 × and multiply-accumulate operation (MAC) by 2.44 × while increases the mAP evaluated on the COCO2014 dataset by 1.6%, compared to the original YOLOv3-tiny network.