Hierarchical Retrieval Model Of Ancient Chinese Character Images Based On Convolutional Neural Network

The ancient Chinese characters have a complex glyph structure and a large number,and there are a great of variants,homographs and written variants,which makes it difficult to realize the image retrieval of ancient Chinese characters.The existing image retrieval technology was not effective when applied to ancient Chinese characters.Aiming at the characteristics of ancient Chinese characters,the research of image retrieval technology applicable to ancient Chinese characters will assist in the research of ancient Chinese characters and help to improve the efficiency of ancient Chinese characters research.Faced with the actual needs of researchers of ancient Chinese characters,based on the analysis and induction of the characteristics of ancient Chinese characters,the convolution neural network was introduced into the research of ancient Chinese character image retrieval,and an ancient Chinese character image feature extraction network and retrieval model were constructed.The main contents of this thesis are as follows:1.ACCINet,a feature extraction network for ancient Chinese character image based on improved VGG-16When extracting features of ancient Chinese character images,the VGG-16 network model paid too much attention to secondary features such as pixel distribution and stroke thickness,while ignoring major features such as glyph structure.To solve this problem,the structure of VGG-16 was adjusted on the ancient Chinese character image data set.Then,we used deformable convolution to improve the geometric deformation adaptability of the feature extraction network,so that it could more accurately extract the ancient Chinese character image features.The experimental results show that the ACCINet model has stronger adaptability to ancient Chinese character images and can extract more accurate ancient Chinese character image features.2.ACCIHNet model for feature hashing of ancient Chinese character imagesBy using a hash layer composed of a segmented encoding layer,an activation layer and a binary encoding layer instead of the fully connected layer of the ACCINet model,the ancient Chinese character image features were hashed.Segmented coding layer was partially connected with the previous layer,and each sub segment generated a bit of hash code independently,which reduced the correlation among bits of hash code,thereby reduced the redundant information in the hash code,and makes the hash code more concise and efficient.In addition,using segmented connections instead of fully connected layers also greatly reduced the number of parameters need to be trained and reduced the risk of over fitting.The experimental results show that the ACCIHNet model can generate simple and efficient hash codes of ancient Chinese characters.3.A hierarchical retrieval algorithm for ancient Chinese character imagesThe high-dimensional depth feature has high retrieval accuracy but long retrieval time,while the hash code has high retrieval speed and low retrieval accuracy.In order to solve this problem,a hierarchical retrieval method from coarse to fine was proposed.In the process of ancient Chinese characters image retrieval,the binary hash code provided by the ACCIHNet model was used for coarse retrieval to determine a image candidate pool with similar features,and then the full connection layer dimension reduction features of the ACCINet model was used for fine retrieval to further obtain images with higher similarity in the image candidate pool.The experimental results show that the method in this thesis can shorten the average retrieval time and ensure the retrieval accuracy.