Research On Object Detection Method Based On YOLOv3

Artificial intelligence and computer vision are rapidly developing as the most cutting-edge computer technology in modern science and technology in the conte xt of the national strategy of ‘strong science and technology'.Object detection is one of the important research directions in the field of computer vision,which is widely used in various fields of society,such as military,medical,traffic and security.Along with the continuous improvement and innovation of artificial intelligence methods,deep learning-based object detection research has been favored by scholars.Deep learning-based object detection methods have the advantages of good generalization performance,adaptability,and robustness,but the current research on object detection methods are only on the detection speed or accuracy,and it is difficult to achieve a balance between the two.How to make the target detector achieve high accuracy and high speed detection is an urgent problem to be solved.To address this problem,this paper presents relevant research work based on the YOLOv3 model in terms of candidate box selection module,backbone network and neck connection.1.The candidate box selection module common to the One-stage and Two-stage framework models in the deep learning model is improved.The characteristics of the non-maximum suppression algorithm are analyzed in the candidate box selection,and the candidate box fusion algorithm is proposed to address the problem that the non-maximum suppression algorithm can only select the re latively optimal candidate boxes and cannot further adjust the candidate box positions.The proposed algorithm aims to select the redundant boxes and selectively fuse the position information of the redundant boxes into the maximized boxes,so as to maximize the amount of object information in the maximized boxes.This method enables the secondary localization of the target after the neural network and obtains a higher index in t he final performance evaluation.2.Improving the model of YOLOv3 backbone network with ResNeXt.We propose a more powerful backbone network,Res?Darknet,to address the shortcomings of the poor feature extraction ability of Darknet-53 backbone network.Firstly,we propose an improvement method based on Res Ne Xt to construct a new res idual block,and then use the improved residual block to redesign a backbone network that meets the structural characteristics of YOLOv3.Finally,the comprehensive performance of the improved model is better compared with other deep learning models.3.The attention mechanism and multi-branch convolution are invoked to optimize the model of YOLOv3 neck connection.The improved YOLOv3 model is proposed to address the shortcomings of the YOLOv3 neck connection which is prone to feature information loss and poor feature expression ability.Firstly,a feature map concatenation module is designed for enhancing the expression of important feature information.Secondly,a feature map mapping module is proposed to replace the simple multi-layer stacked convolution structure in the original model with a multi-branch convolution module with multiple sensory fields,so as to capture richer feature information.Through experimental verification,the improved method can achieve real-time object detection with higher accuracy rate.