Application of Deep Convolutional Neural Network Under Region Proposal Network in Patent Graphic Recognition and Retrieval

Abstract

To improve the security of product technology and design efficiency, the recognition and retrieval methods of patent graphics are explored. First, for low detection accuracy of images in traditional detection methods, the deep convolutional neural network (CNN) with stronger feature extraction capabilities is selected for feature extraction. Second, the region proposal algorithm is used to improve patent graphic recognition accuracy and reduce the probability of image feature missed detection. Finally, the detection model is compressed by the deep separable CNN, and the convolution kernel performs the separation operation. The results show that the deep CNN has a more vital ability to extract features. The region proposal algorithm can accurately select the regions to be recognized and detected, which reduces the recognition workload. Simultaneously, through the multi-layer feature fusion strategy, the neural network’s ability to express patent graphics is improved, and the probability of missed detection of image features is reduced. The deep separable CNN can allow the convolution kernel to perform the separation operation, effectively reduce the model parameter amount, and shorten the model retrieval time. The investigation finds that the deep CNN based on the region proposal network (RPN) improves the accuracy and precision of the recognition and retrieval of patent graphics, as well as the detection time of the algorithm. The method of this investigation can be applied to the recognition and detection of actual patent graphics, thereby improving the security of patent information and the innovation of patent design.