Object Classification And Detection Based On Attention Mechanism And Knowledge Distillation

Natural scene understanding is one of the hottest research directions in computer vision at present,which mainly includes subproblems of target detection,image clas-sification,video understanding and image segmentation.In recent years,convolutional neural networks have developed rapidly,bringing new methods to the study of natural scene understanding.This paper mainly focuses on how to obtain the semantic infor-mation of various targets in the natural scene,especially the category label and location information of the target.Therefore,this paper selects image fine-grained classification and target detection as the research content.The fine-grained classification of images acquires the category label information of the target,and the target detection acquires the position information of the target.The relevant research has important theoretical significance and application value.It is a difficult problem to obtain the whole and local information of the object in the weakly supervised image fine granularity classification.To solve this problem,this paper proposes a new weak supervised learning method based on attention mechanism and mixed links.In the problem of target detection,aiming at the slow detection speed caused by multiple ROI Pooling,this paper improves Cascade R-CNN network based on knowledge distillation,and proposes a new two-stage detection network based on knowledge distillation.The specific work is as follows:1.For obtain the weak supervision and learning in the whole and partial difficult problem,this article,based on the analysis,using the receptive field of receptive field changes over the depth of the network characteristics,through the network character-istic figure weighted way to strengthen the network to study the features of object area and the local area,and introduced in the mixed link attention mechanism,this paper proposes a based on convolutional neural network used for image classification of fine-grained weak supervision and learning network-attention mixed link network.The method is based on the different levels of feature weighting,realize from the whole image to the object,to the local focus on refining process,different from other layers or module network,this paper directly to the attention mechanism into the mixed links in the network for image classification,using convolution with layers into wild focus refinement process and scale change.In the open dataset of Stanford Dogs,the clas-sification accuracy of this method is 86.2%,reaching the leading level in this field.In addition,experiments were carried out on small self-built data sets of natural disasters and ice-wind disasters,and the classification accuracy was over 90%.2.In view of the slow detection speed caused by multiple ROI Pooling,based on the Cascade R-CNN network architecture for target detection,this paper proposes a two-stage target detection network architecture with faster speed and fewer parameters-two-stage distillation network.Based on the full connection module instead of a Cascade of R-CNN after two ROI Pooling to reduce the calculation,draw lessons from the thought of "distillation of knowledge".This paper designs a loss function that includes improved network prediction results and sample real results and improved information between the network and Cascade R-CNN,and uses the results of Cascade R-CNN second and third-stage detector regression classification as "soft targets".Add to the loss function to learn how Cascade R-CNN continues to refine the regression results.In this paper,the performance of COCO public data set was comparable to that of Cascade R-CNN model,but the speed was faster and the mAP beyond one point was higher than that of Faster R-CNN,which was also a two-stage detector.The experiments on the open data set and the self-built data set show that the method proposed in this paper achieves ideal experimental results on image fine par-ticle classification and target detection,which verifies the effectiveness of the above method.