Research And Industrial Application Of Fine-grained Image Recognition Based On Deep Learning Under Weak Supervision

With the application of deep learning and other related artificial intelligence technologies in social life,artificial intelligence technology has gradually penetrated into all aspects of human life and production.But there are still many more difficult problems that require more in-depth technology to solve.In the field of basic image recognition,fine-grained recognition has always been a problem that needs to be solved.How to do fine-grained recognition research and application of complex scenes under weak supervision,and solve the problem of fine-grained recognition because of similar categories is easily confused This topic focuses on the improvement of fine-grained identification algorithm and industrial implementation under weak supervision from its essence.The specific contents and results are as follows:Aiming at the problem of effect and performance balance in fine-grained tasks,the neural network structure search method is used to let the model choose the optimal parameters independently,and to build the optimal performance network structure under certain time-consuming conditions,so as to find a model with smaller structure and better effect.Aiming at the problem of recognition accuracy,a new network based on multi-feature fusion is proposed.The global information branch,backbone information branch,and local information branch of the new network are used to obtain the global features,backbone features,and local features of the input image respectively,and then high-order fusion is performed,so that the model obtains more expression information and has better discrimination force.Combined with the most advanced knowledge distillation and model compression technologies in academia,the overall optimization of small models.Aiming at the application problems of complex industrial scenes,combined with the improvement of the above algorithm ideas,a set of fine-grained problem solutions for actual complex scenes landing applications was explored.Experiments show that the model based on neural network structure search algorithm is better than the hand-designed model.Compared with the commonly used model on the mobile terminal,the effect is 2.5%higher while the parameters are similar.The proposed new network structure can give a good fusion of different levels of features,so that the accuracy of the model on the public data set is 2.1%higher than that of the general model using only the backbone network,while reducing the amount of model parameters and time-consuming,Memory consumption and other indicators that are often concerned during actual deployment.Basically solved certain problems related to fine-grained identification.