Research On Object Detection Algorithm Based On Anchor Free

Since the explosion of deep learning technology in 2012,algorithms in the field of object detection have also seen a huge innovation.Object detection algorithms on deep learning technology emerge in an endless stream,and the detection accuracy is getting higher and higher and the speed meets the real-time requirements.Meanwhile,the detection algorithms become diversified.In order to explore the detection performance of the Anchor Free algorithm in the field of visible light images and thermal infrared images,CenterNet,a representative algorithm in this field,is selected as the baseline in this paper.The algorithm has clever ideas,good performance and simple network structure.After its publication,many researchers optimize and apply it.Through a series of improvement measures,the algorithm achieves a better level in two different image fields and is superior to some new algorithms proposed at the same time or later.Specifically,this paper mainly made improvements in the following two aspects:(1)Focusing on the problem of low utilization of object features and inaccurate detection results in CenterNet,an improved algorithm of double branch feature fusion is proposed in the paper.One branch of the algorithm includes feature pyramid enhancement module and feature fusion module to fuse the multi-layer features output from the backbone network.Meanwhile,in order to utilize more semantic information,the upsampling branch of the baseline is preserved.Secondly,a frequency-based channel attention mechanism is added to the backbone network to enhance the richness of extracted features.Finally,the features of the two branches are concatenated and convoluted.The detection accuracy on PASCAL VOC dataset is 82.3%,which is 3.6%higher than CenterNet,and the detection accuracy on KITTI dataset is 6%higher than CenterNet.The detection speed meets the realtime requirements.The double branch feature fusion method is proposed to process the features of different layers,which can better integrate the spatial and semantic information in different layers of the network,and improves the detection performance of the algorithm.(2)Focusing on the problems of CenterNet in infrared images,such as feature loss and insufficient information utilization,an improved algorithm based on spatial feature enhancement is proposed.Firstly,a frequency-space enhancement module is used to enhance the details of the target region.Secondly,a module that can count global information is introduced into the backbone network to model the feature graph globally.Finally,in the case of no increase in computation and complexity,the residual mechanism is adopted to redesign the overall structure of the algorithm,which strengthens the feature interaction simply and efficiently.Experimental tests are carried out on the self-established infrared object detection dataset G-TIR and public infrared object detection dataset FLIR.The proposed algorithm improves the accuracy of the baseline by 8.4%and 15.3%respectively,and is better than many mainstream object detection algorithms in recent years.Meanwhile,the detection speed reaches 72 FPS,which balances the detection accuracy and speed well,and meet the real-time requirements.