Research Of The Medical Image Retrieval Algorithm Based On Perceptual Hash

With the development of information and medical technology, CT, MRI and other digital medical equipment have been increasingly widely used in clinical work and auxiliary diagnosis. These equipment make medical institutions produce large amounts of medical imaging data every day. Medical images have become an indispensable tool in modern clinical diagnosis and medical research. How quickly and efficiently retrieved the images needed from these massive medical images is becoming an increasingly pressing issue. Development and application of content-based image retrieval technology on the one hand greatly reduce the workload of doctors and improve the efficiency, on the other hand, make more objective diagnostic and improve the accuracy of diagnosis.Currently the extracted features dimension of content-based medical image retrieval are generally high, the extracted features of images require a large amount of computer memory, resulting in retrieval efficiency is not high, and image similarity calculation is more complex, resulting in image retrieval speed is not high, too. The image perceptual hashing technique is a new image processing technology proposed on the basis of the traditional hashing techniques, the robustness and good compression of image perceptual hash provide efficient and accurate technical support for medical image retrieval. The image perceptual hash technique can make the extracted features compressed into a binary sequence of dozens or hundreds of bits, and use the hamming distance to calculate similarity, which greatly reduces the amount of storage of medical image features, while the complexity of similarity calculation is also greatly reduced.In this thesis, image perceptual hash technology is used in medical image retrieval, and two new medical image retrieval algorithm based on perceptual hash are proposed. (1) Proposed a medical image retrieval algorithm which is based on DCT and LBP perceptual hash. Firstly, the medical image is divided into several blocks, for each image block, use the low-frequency coefficients which get from discrete cosine transform to represent the frequency domain features, and then extract LBP features for each image block, then fuse the two features, use the principal component analysis to compress and quantize the getting features, and ultimately get the binary feature sequences, use the hamming distance to calculate similarity between images. (2) Proposed another medical image retrieval algorithm which is based on NMF and Zernike perceptual hash. Disassemble the medical image with NMF and get local features, at the same time, calculate Zernike moments of the image and get global features, these two features are fused to get the final binary feature sequences, use the hamming distance to calculate similarity between images. Experimental results show that both the proposed algorithm have good retrieval effect.