On the multi-scale real-time object detection using resnet

Abstract

Real-time target detection and location has important values in video surveillance. Aimed at the low accuracy of existing real-time object detection algorithms, this paper proposes a multi-scale real-time target detection algorithm based on residual convolution neural network. Firstly, the residual convolutional neural network is introduced into the YOLOv3-Tiny algorithm. The jump connection of the low-level and high-level networks forms the residual module in the YOLOv3-Tiny algorithm to effectively prevent network degradation while increasing the depth of the neural network. Secondly, a new prediction layer is added to the neural network to improve the results of small target detection. Finally, the trained model is tested on the Pascal VOC public dataset. The experimental results show that the proposed algorithm achieves 64.6% accuracy on the validation dataset, and the speed of 60FPS in the video detection. The detection accuracy is improved to a higher level at a small cost of a little lower speed still meeting the real-time detection requirements, and small targets in the image can be effectively detected. The algorithm is effective and robust.