The Wavelet-based Saliency Model And Its Application In The Lumbar Disc Segmentation

The visual attention of human visual system enables people to select a portion of relevant information from a large of information to perform further processing. With the development of multimedia technology, image and video becomes the main information carrier instead of the text gradually. If the computer can imitate the human visual system to process the information selectively, it not only saves large amounts of computing resources, but also improves the processing efficiency. This makes the research of visual attention become a research hotspot. In this thesis, the visual saliency model and its application in lumbar disc segmentation are researched. The main contents of this thesis are as follows:(1) The wavelet-based feature probability evaluation model is proposed after summarized the inferiors and superiors of the existing models. In the proposed model, the visual saliency of a pixel is defined as the self-information of the local features. The model applies the multi-channel structure based on the YCbCr color space. The feature vector is obtained using the wavelet decomposition, and the joint probability density of feature vectors is evaluated with the ICA method. The experimental results demonstrate that the model has the superior ability to detect saliency, can provide the accurate and fast evaluation of saliency. In the experiment compared with existing algorithms, the proposed model not only demonstrates the certain superiority in psychology image test, but also obtains the results accord with human’s visual attention in the natural image test.(2) The saliency map of lumbar CT image is computed by the wavelet-based feature probability evaluation model. In the algorithm flow, the image is processed by boundary remove, multi-scale image enhancement, edge detection with orientation information measure, anisotropic diffusion. Finally, the saliency of filter result is computed by the proposed saliency model. The saliency of lumbar disc region is higher than other regions, therefore the lumbar disc region image can be extracted based on the saliency.(3) The lumber disc region is extracted roughly based on saliency, then the center point of lumbar disc is detected by Hough forest. Consider this point as the center and use the vertical direction to the line of two disc center as the inclination angle to extract a fixed size region as the disc region image which is to be segmented. Finally, this disc region image is segmented in order to get the contour profile of the lumbar disc. The method not only takes advantage of the superiority of CT in bone imaging, but also removes the influence of soft tissue area in the segmentation. The results show that the disc can be segmented precisely from the lumbar image through the saliency computation, the disc center detection and threshold segmentation.