Object Detection Based on Improved YOLOv3-tiny
Hua Gong, Hui Li, Ke Xu, Yong Zhang · 2019
Deep learning has gradually become the mainstream object detection algorithm because of its powerful feature extraction ability and adaptive ability. However, how to guarantee the accuracy and speed is still a huge challenge in the field of object detection. This paper proposes an improved YOLOv3-tiny for object detection based on the idea of feature fusion. YOLOv3-tiny is chosen as the basic network framework to ensure the identification speed. In order to improve the poor detection accuracy of YOLOv3-tiny network, feature fusion is carried out based on Feature Pyramid Network. The last DBL (Darknetconv2d_ Batch Normalization_Leaky) output of the second scale output layer is fused with the fourth DBL output of the network. And a 52$\times$ 52 scale output is added on the basis of the original network. Experiments were carried out on our dataset. The experimental results show that compared with YOLOv3-tiny, the accuracy of the improved network structure is increased by 6.3%, and the detection speed is 31.8fps, ensuring the requirements of real-time detection.