Research And Application Of Region-based Convolutional Network In Target Detection And Shadow Environment

Image classification is an important task in the field of computer vision.In the development of image classification,many model methods have been proposed and achieved satisfactory classification accuracy.Because the popular classification model methods are driven by large-scale data,which are lacking in many actual scenarios.Therefore,zero sample learning has been proposed and become a hot research topic.At present,many zero shot image classification methods use learning the mapping relationship between semantic information space and image feature space to achieve image classification,but these methods will produce hub and mapping domain drift problems when mapping.Moreover,the uneven distribution of samples between visible and invisible classes leads to the poor performance of many methods in dealing with generalized zero shot tasks.Based on the above problems,this thesis proposes to use the generation model to solve them.The main work is as follows:Firstly,a zero shot image classification model based on image feature reconstruction is proposed.The semantic information of invisible classes is used as a condition to generate image features of invisible classes by using the generation countermeasure network to alleviate the problem of uneven sample allocation.At the same time,in order to make the generated image features more realistic and diverse,a reconstruction network is added to the generator network to reconstruct the generated visual features back to the semantic information,and judge the gap between them to limit the generator to generate more semantic image features.The effect of the previous model in dealing with fine-grained data sets is not ideal,so the attention kernel bilinear network is introduced on the basis of the previous model.The attention mechanism is used to improve the weight of important information in image features and reduce the weight of redundant information.At the same time,the kernel network can obtain nonlinear relations,so that the output image features can be learned more easily.On the basis of the above research,a zero shot image classification system based on visual samples is designed and implemented.The trained model is integrated into the system,which can realize the functions of image classification,viewing the classification accuracy of different data sets and viewing the operation log.After analyzing the requirements of image classification system,each module is designed and coded and tested to develop a fully functional image classification system.