A Fast Detection Algorithm of Small Targets Based on YOLOv3

Abstract

Robots, endowed with computer vision function by on-board cameras and embedded systems, have become popular in the modem power industry. In order to realize automatic image analysis at the data acquisition part, a fast detection algorithm especially for small target based on YOLOv3 was proposed. It is proposed to up sample by 2 X feature map which is down sampled by 8 X of the previous network, and concatenate it with the output of the second Res block unit, and the last Res block unit of the previous network was abandoned. To obtain more features of the small target, add two Resnet units in the second Res block unit of Darknet53 in YOLO V3 network structure. Aiming at deploying the algorithm to the hardware, model pruning was applied to decrease model size and improve detection speed. The experimental results demonstrate that the mAP (Mean Average Precision) reaches 0.85, and the proposed algorithm is capable of real-time detection as faster than 15 frames per second (fps).