Research On Medical Image Classification And Identification

With the continuous development of computer vision technology and medical imaging devices, medical images reveal more information.(classifying)these medical images effectively attracts people's more and more attention. Therefore,(the dissertation) focuses on three aspects,(concluding) medical images preprocessing,feature extraction and selection, classifier design.First, before images are classified and recognized, relative images preprocessing is carried out based on the characteristics of images. 1. wavelet transforms are applied to denoise images to reduce the effect of noise on the classification. 2. GAC model based on level set method is applied to segment images, which separates the target and background region. Thus, it can reduce the effect of background region on the classification.Second, traditional image histograms only allows for the statistical property of pixels, while ignoring the color spatial distribution. Thus, it results that different images may have similar color histograms. Therefore, we calculated the proportion of pixel color value in original image, at first; then, applied weighted processing on pixel gray level to form color histograms that can reflect the number of different color pixel in the original image. At same time, Tamura texture features on aspects of roughness,contrast, and directionality are applied to describe the texture of the image. Then, the color features and texture features of the image are normalized to form data field of describing the image feature. Finally, the image is classified and recognized by SVM based on fuzzy kernel clustering.Third, while the image color feature is extracted, Gabor wavelet transforms are applied to describe texture features of the image and form data field of features. And the capture-control mechanism among PCNN neurons is used to design the classifier,which makes corresponding neurons of same samples activate earlier than ones of other samples at all time, to the effect that internal active items of corresponding neurons of same samples are largest, thus classifying images effectively. Research on biological neurology shows that there are many different neuron functional zones in human's pallium, every of which consists of several neurons that can accomplish the specific functions of corresponding functional zones. In addition, brain memory is nota certain neuron fully corresponding to a model but a group of neurons correspond to a certain model. Whereas, the self-organizational characteristics are achieved by nurture. Research on biological neurology also shows that excitement caused by a thing is not aimed at a certain neuron but all neurons in a zone centered on a certain neuron. Neurons in the center of the zone are most exciting, and the intensity is gradually weak with the distance away from center. Moreover, neurons which are apart from the center are subject to certain limitation. Every neuron within the neuron zone in the same model excites each other, while neuron zones in a different model restrict each other. Therefore, the method of imaging classification based on capture-control mechanism among PCNN neurons is fully consistent with functional zones characteristics of human's brain neurons.