Research On Object Detection Algorithm Based On Feature Fusion And Anti-occlusion Network

Object detection is a challenging research work in the field of computer vision.It has important applications in 3D object detection,instance segmentation and intelligent video surveillance.However,there are lots of prospects or shade the small targets with similar background in actual test scenario,when these scenes are detected,the detection accuracy is not high,and the problem is more prominent when the size and shape of the target are different,and the number and location of the target are uncertain.In order to enhance the performance of the target detection algorithm based on deep learning in detecting small targets with occlusion or low resolution,this paper makes corresponding improvements to the Faster R-CNN algorithm.The main research contents are as follows:(1)In view of the mismatch between semantic information and location information in the feature images extracted by the convolutional extraction network in the Fast-R-CNN algorithm,the RFPN feature fusion algorithm is proposed to fuse the feature images of different scales extracted by the convolutional extraction network.Firstly,by adding RPN,the ability of the algorithm to select the target candidate box is improved.Secondly,on the basis of RPN,the FPN network is added to solve the multi-scale problem in the process of target detection by the fusion methods of bottom-up,top-down and horizontal connection.Finally,the data set is enhanced to improve the generalization ability of the network model,and then the accuracy of target detection is improved.(2)Aiming at the problem that FAST R-CNN detector has poor classification ability for low-resolution targets with occlusion,a target detection algorithm based on RFPN and anti-occlusion network is proposed.Based on the RFPN algorithm,the FAST R-CNN detector was innovated and improved,and the adversarial learning strategy and feature mapping are adopted to improve the classification ability of FAST R-CNN detector.Firstly,through antagonistic learning between the generator and FAST R-CNN detector,the detection performance of the target detection algorithm based on RFPN and antagonistic occlusion network for small targets with occlusion is improved without increasing the training samples.Secondly,the ROI pooled low-resolution feature map is mapped into high-resolution feature map by feature mapping,so as to achieve higher detection accuracy of low resolution small target and high resolution large target by target detection algorithm based on RFPN and anti-occlusion network.(3)Through Pascal VOC and MSCOCO data sets,the target detection algorithm based on RFPN and anti-occlusion network is verified and analyzed for the AP and MAP values of small targets with low resolution and occlusion,and it is found that the AP and MAP values are both improved.The algorithm based on VGG-16 achieves 75.80%on VOC2007 and 73.90%on VOC2012.The mAP value of the based on ResNet-101 algorithm on VOC2007 is 80.71%,and the mAP value on VOC2012 is 77.72%.In the MSCOCO data,AP50 reached 58.6%and AP 75 reached 40.6%.Figure[55]Table[13]Reference[75]...