Research On Object Detection Methods Based On YOLOv4

As one of the most popular research fields in computer vision,object detection has important application value in many practical scenarios.This thesis studies the object detection method based on deep learning.As the current mainstream one-stage object detection method,YOLOv4 can better balance the detection accuracy and speed,so this thesis selects it as the research object.Aiming at the problems of insufficient feature extraction of deep and shallow information and insufficient utilization of information in YOLOv4,it is improved to ensure the detection speed basically unchanged while as much as possible to improve the accuracy.The main research contents and innovative work of this thesis are as follows:(1)In view of the problem of insufficient extraction of deep and shallow information features of the YOLOv4 object detector,a new improved YOLOv4 object detection method based on dilated coordinate attention is proposed,namely YOLOv4-D.This method uses multiple dilated convolutions with different dilation rates to improve the attention mechanism module,and proposes a dilated coordinate attention module,which is placed before the shallow feature layer of the backbone network,so that the feature maps under different receptive fields can be fused,and the feature extraction ability for shallow networks is significantly improved.At the same time,a multi-scale training strategy is adopted to enhance the robustness.Finally,experiments are carried out on PASCAL VOC2007 and VOC2012,and compared with other advanced object detection methods.The experimental results show that the detection accuracy of the proposed YOLOv4-D is better than YOLOv4 and other object detection methods.(2)Aiming at the problem of insufficient information utilization in the YOLOv4 object detector,a new object detection method based on improved path aggregation and pooling YOLOv4,namely YOLOv4-P,is proposed.In order to make full use of the feature that the path aggregation network can effectively prevent information loss,the path aggregation network of YOLOv4 is improved,and the second residual block of the backbone feature extraction network is used to add a detection layer to strengthen the fusion of shallow feature layers.In addition,the datasets are reprocessed using Kmeans clustering to obtain a suitable prior box size.Also,the size of the receptive field of the image after passing through the backbone feature extraction network is smaller than the theoretical one.In order to increase the receptive field,a pyramid pooling module is added to the back end of the backbone feature extraction network,and four different scales of pyramid pooling are used to introduce features at different scales information.The simulation experiments on PASCAL VOC2007 and VOC2012 show that the proposed YOLOv4-P effectively improves the detection accuracy.(3)In order to further verify the performance of the proposed YOLOv4-D and YOLOv4-P,the two are combined to obtain a new YOLOv4-Z object detection method.Simulation experiments are carried out on the two datasets and the experimental results are analyzed.The m APs of YOLOv4-Z on the two datasets are 2.25% and 2.11% higher than that of YOLOv4,respectively.On the PASCAL VOC2007 dataset,compared with the proposed YOLOv4-D and YOLOv4-P methods,YOLOv4-Z improves by 0.41% and 0.22%,respectively.on the PASCAL VOC2012 dataset,it improves by 0.2%and 0.17%,respectively.Compared with YOLOv4,the proposed two object detection methods,YOLOv4-D and YOLOv4-P,have improved detection accuracy,and YOLOv4-P is better.The YOLOv4-Z proposed by combining the two has a better detection effect.Experiments show that YOLOv4-Z has higher detection accuracy than the pre-combined detections without major changes in detection speed,and has advantages over other advanced object detection methods.