Research On Object Detection Method By Parallel Connecting Deep-shallow Layers With Squeeze-and-excitation

Object detection is one of the basic and main technologies in computer vision,and its application is also very extensive,including intelligent security and monitoring system,intelligent transportation and driving system,Internet mobile terminal and military applications.The purpose of object detection is to recognize the category of target from complex images and give the location information of the target in the image.Therefore,the task of object detection includes two important sub-tasks: target classification and target location.Compared with traditional object detection methods,the two-stage object detection method based on R-CNN series eliminates the subjective one-sidedness of artificial feature extraction,and achieves the integration of target feature extraction and classification process.In view of the slow speed of two-stage object detector,the single-stage object detection method based on direct regression does not need the region recommended process,but obtains the category and border of the target through direct regression at multiple locations of the image to complete the object detection.However,the two-stage object detection method based on R-CNN series and the single-stage object detection method based on direct regression both have the problem of low accuracy of small object detection.Therefore,the main idea of this paper is that the model can significantly improve the accuracy of small object detection while maintaining the accuracy of large object detection.At the same time,test speed and convergence speed are used as auxiliary evaluation indicators.Based on the above ideas,the main work of this paper is as follows:1.In this paper,a object detection model DS-CNN based on deep and shallow CNN parallel connection is proposed.Firstly,in order to reduce the loss of small target information and maintain the detection accuracy of large target,this paper designs a shallow neural network suitable for extracting small target information,and redesigns a deep neural network based on extended convolution theory.The deep-shallow network is organically parallel by cross-layer connection to extract the bottom image features and high-level semantic features of the target.Secondly,in order to improve the detection speed and accuracy of target candidate region recommendation mechanism,we uses RPN in Faster R-CNN for reference,and innovates the RPN mechanism of target candidate box generation.Finally,in order to reduce the parameter redundancy caused by full connection layer,a new feature dimension reducer is designed using pooling layer of ROI and convolution layer to speed up object detection.The experimental results show that DS-CNN model not only keeps the accuracy of large object detection,but also significantly improves the accuracy of small object detection,and the speed of object detection is not significantly reduced.2.Based on the DS-CNN model,this paper proposes a sequeeze and excitation detection network SE-CNN based on transfer adversarial spatial dropout network.Firstly,in order to enhance the ability of network to extract image features,SE module is used in the deep and shallow network of DS-CNN to readjust the channel relationship.Secondly,in order to improve the ability of sample detection,we uses the idea of transfer learning to transfer the ASDN which can produce difficult samples to the DS-CNN model.Finally,in order to solve the problem of position deviation caused by ROI Pooling operation,we uses ROI Alignment layer to redesign feature dimension reducer to enhance the detection ability of small objects.The experimental results show that the SE-CNN model improves the detection ability of all categories,especially for small objects,and the convergence speed is improved significantly.