Research On Object Detection Algorithm In Shadow Environment Based On Regional Convolutional Neural Network

With the development of image recognition,the effect of various algorithms is getting better and better.At present,in object detection technology,shadow pixels interfere with the effect of the algorithm in complex environment.The existing shadow detection algorithms use attention mechanism to extract cross-channel features and global pixel information is not enough.Another common problem of object detection algorithms is that most of the existing feature extraction networks have good detection effect on large-scale targets,but poor detection effect on small-scale targets.The reason is that the feature information of target subjects with different scales is extracted from bottom to top,which leads to the imbalance of detection effect due to different target scales.Moreover,the anchor frame filtering algorithm used in the current object detection algorithm forces the score of adjacent detection frames with low confidence to zero,which is easy to miss detection in the case of dense image targets.Aiming at the shadow problem,by studying the shadow feature information and combining the design idea of hybrid attention mechanism,this thesis constructs a new network Res-CCNet integrating channel attention and spatial attention,and reuses the ignored features by using dense connection and feature fusion.Experiments were carried out on shadow detection data sets SBU and UCF,and the evaluation indexes SER,NER and BER were used for verification.Experiments show that the shadow detection module based on hybrid attention mechanism is effective,which lays a foundation for the street scene object detection model in the whole shadow environment.Aiming at the imbalance of object detection effect,Mask R-CNN,a object detection and semantic segmentation model based on regional convolution neural network,is adopted,an improved feature extraction module is proposed,and the anchor frame screening method is replaced by duplicate removal network to form the ED-Mask R-CNN detection model.Using the custom Street View dataset cityscapes for testing,the experiment shows that compared with the original model,its detection accuracy for small-scale targets is greatly improved,and it is improved for other scale targets to a certain extent.Based on the above,combined with the shadow detection module,the influence of shadow pixels is reduced on the results of ED-Mask R-CNN model.Experiments show that compared with the ED-Mask R-CNN model,the effect of the object detection model fused with the shadow detection module is further improved.