Color image retrieval based on improved quaternion convolutional neural network

Nowadays, when performing color image texture feature extraction, most of the color images are converted into grayscale images or the color images are processed by channels. The above-mentioned traditional processing methods will largely lose the effective information of the original image when performing color image retrieval. Aiming at the problem of effective feature description in content-based image retrieval, this paper proposes a color image retrieval method based on improved quaternion convolutional neural network. First, use the quaternion method to characterize the local texture feature LBP value and R, G, and B color information of the image, and use the quaternion discrete cosine transform (QDCT) as the theoretical basis to express the color image as a quaternion matrix form. Secondly, each component of the convolutional neural network is extended to the quaternion domain, and the quaternion matrix obtained above is directly input into the proposed neural network for training and classification. Finally, the result of image retrieval is obtained. Experiments show that this method significantly improves the performance of color image retrieval.