Study on Improvement of YOLOv3 Algorithm

In order to optimize the problem of wrong detection and missed detection of small targets in complex environment, a target detection algorithm of YOLOv3-SPP5 was proposed. YOLOv3 in the deep learning algorithm has achieved excellent detection effect in target detection, but it is not perfect in the complex environment. In this paper, YOLO detection heads were added to the overall network of YOLOv3 to improve the extraction of multi-scale information features. In order to avoid the effect of feature extraction being reduced with the increase of network depth, Spatial Pyramid Pooling (SPP) were added to each YOLO detection head to optimize the extraction of deep network features. The optimized algorithm was named YOLOv3-SPP5.With the same setting of YOLOv3-SPP5 and YOLOv3, Under the same setting of YOLOv3-SPP5 and YOLOv3, the experimental results on COCO data set show that the mAP of YOLOv3-SPP5 increases by 7.9, the Inference time and Volume do not increase significantly. The above experimental results show that the optimized YOLOv3-SPP5 algorithm is more suitable for target detection in complex scenes than YOLOv3 algorithm.