Research And Application Of Object Detection Method Based On Anchor-free

As an important basic research direction of computer vision,object detection is widely used in various fields.Among all the object dection methods,the anchor-free object detection algorithms are very suitable for mobile devices due to there simple models and easy portability.In this thesis,the anchor-free object detection algorithms are studied from the two aspects of general object detection and detection application.The anchor-free object detection algorithms based on key points has problems such as poor multi-scale object detection and inaccurate matching of detection boxes for dense objects.The detection process of the anchor-free object detection algorithms based on the central domain are simple and usually combined with a lightweight backbone network for detection application deployment.The lightweight model has problems such as insufficient feature extraction and low detection accuracy.Aiming at the above problems,this thesis combines two representative anchor-free object detection algorithms to conduct the following research:Firstly,this thesis selects the Center Net algorithm for optimization for the general object detection direction.Res Net is selected as the backbone network of Center Net,and the bottleneck module and loss function of the algorithm are improved.Because Center Net does not use the anchor mechanism,the center point coordinates of the model regression are not supervised by the anchor,and the center point prediction effect is poor.Therefore,it is necessary to obtain rich semantic information during feature extraction to obtain a good detection effect.A dilation encoder is used for the C5 feature layer of the network to expand the coverage of the feature scale and obtain multi-scale context information.Use feature fusion and interpolation up-sampling instead of transposed convolution to avoid loss of feature information.Optimize the regression loss,add distance constraints to the center point prediction,improve the accuracy of the center point regression.Use the GIo U loss to supervise the detection boxes to improve overlap.By comparing with the original Center Net algorithm on the authoritative object detection datasets,the effectiveness of the improvement is proved.Secondly,the Nano Det algorithm is optimized for detection applications.Nano Det is a lightweight object detection algorithm based on central domain detection.The Nano Det algorithm abandons the convolution operation in the bottleneck module in order to lighten the model,resulting in serious loss of features and large loss of accuracy.In order to improve the detection results of the model without reducing the model inference speed,this thesis uses the Trident module with shared parameters to complete the last down-sampling of the network to capture multi-scale context information.The convolution in the PAN structure of the bottleneck module is implemented by the Ghost module,and a down-sampling layer is added to the PAN structure to improve the performance of the model.The convolution in the PAN structure of the bottleneck module is implemented by the Ghost module,and a down-sampling layer is added to the PAN structure to improve the performance of the model.The depthwise separable convolution in the detection head uses a 5*5 convolution kernel to further improve the receptive field of the network.The vehicle headlight detection is selected as the application direction to verify the algorithm improvement.The main purpose is to judge whether the light of the vehicle works normally in the process of appearance inspection,light inspection and brake inspection Collect images and make a vehicle headlight detection dataset.The Nano Det algorithm is compared and analyzed with the improved algorithm in this thesis.The experimental results on the self-built vehicle headlight detection dataset prove that the improved method in this thesis can significantly improve the average detection accuracy while only slightly reducing the model inference speed.Finally,a web vehicle lamp detection system is designed.The improved model based on the lightweight object detection algorithm Nano Det is applied to the detection module of the systems.In the web system,users can upload pictures or video files,and the front-end interface of the system can feed back the test results or provide the function of downloading the test results.After testing,the vehicle lamp detection systems designed in this thesis has certain accuracy,effectiveness and meets the real-time application requirements.Aiming at the problem of incorrect matching of dense objects and insufficient feature extraction in anchor-free object detection algorithms,this thesis selects two representative anchor-free object detection algorithms for experiments and improvements.Use methods such as dilated convolution and feature fusion to increase the scale range covered by network extraction features and improve network detection accuracy.The improved lightweight detection network is applied to vehicle headlight detection,and a vehicle headlight detection system is designed and implemented.