Research And Implement Of The Design Patent Image Classification Based On Image Caption

Strengthening the protection of design patents is an important part of improving the intellectual property protection system,and retrieval is a key means of protecting design patents.The design patent retrieval system is mainly a content-based image retrieval system.There is a defeat that a semantic gap exists in the content-based image retrieval system,and the description information of the design patent image content in the text data of the design patent application file is very limited.Therefore,the design patent image retrieval system has low retrieval efficiency and cannot satisfy the user's diverse semantic retrieval needs.Image classification is one of the important methods to cross the semantic gap,which can improve the retrieval efficiency of the system and also meet the needs of users' semantic retrieval.However,it's a large number of the design patent image data,and the general image classification model can only realizes the automatic recognition and determination of image categories,and it is difficult to express more content of the semantic information in the image,thus there is a limit of the general image classification model for improving the searching efficiency of design patent image system.The image caption model can automatically generate content descriptions for a given image and can express richer content information in the image.Therefore,it is necessary to make a research of image caption.Aiming at the above problems,this article applies image caption to design patent images,and proposes a classification method for design patent images based on image caption.In order to verify the feasibility of the method proposed in the text,the main content of this article is as follows:By analyzing the classification basis of design patents and the search features used to improve the searching efficiency of design patent images,annotate design patent images and construct a design patent image caption data set.During the process of obtaining the design patent image caption model encode network,this paper will use two convolutional neural networks(VGG16 and inception V3)pre-trained on the Image Net dataset to implement the design patent image classification model.Through training and fine-tuning,two design patent image classification models based on pre-trained convolutional neural networks are obtained,and the accuracy rate of design patent image classification is higher than the experimental comparison model.At the same time,this paper has analyzed the capability of image feature extraction of coding network.By training the design patent image caption model based on the encoder-decoder structure on the design patent image caption data set,an automatic description generation of the design patent image by the model is achieved.The feasibility of the method proposed in this paper is verified by analyzing the accuracy of the description results generated by the model on the classification features of design patent image.At the same time,the text analyzes the impact of different encode networks on the performance of the design patent image caption model under the same conditions,and the effect of different training sample quantity on the classification performance of the design patent image.