Cultural Relics Image Content Understanding Based On Deep Learning

In the modern society with complex information on the Internet, the visual information, such as images, videos spread fast and wide. How to extract useful information from the huge information database, and use that information to create value is a problem which is worthy of study. Image content understanding means to get the text which can be the semantic description of the image through image recognition. The significance of image content understanding is to identify violence, unhealthy image, to prevent the harmful information spread and harmful behavior occurring. This paper focuses on image recognition in the application of the image content understanding of cultural relics On the other hand, image recognition and image retrieval is closely related. We want to make more accurate and efficient image retrievals through the research on image recognition method. Since the image content identification has wide range of applications, it also makes more sense to do the research in this field.Image feature extraction is an important part of the image content recognition. Unlike some traditional feature extraction method, which need to manually set the extracted features, to build a network of deep learning method is proved to have a great advantage because they can learn the features all by themselves. In this paper, Convolutional neural network, which is suitable for such a multi-dimensional image information identification, is used for image feature extraction, image classification, image content information acquirement. The computational complexity is reduced through sharing the weight. While data dimensionality reduction operation is introduced to the deep learning model which make the learning efficiency of the model improved. As the full linked layer is added before the output layer of the network, the image retrieval speed is improved with the feature matching algorithm improved. Multiple sets of comparative experiments show that the deep learning methods used in this paper make the image recognition task more accurate and efficient.