Color Image Segmentation Based On Dependency Tree And Neighborhood Approximation

Nowadays,with the improvement of the quality and production of computer hardware and the increasing demand for computing,color image segmentation has been more valuable than gray image segmentation.Color image segmentation technology has been applied to various fields,and a variety of novel algorithms have been put forward.However,the color image segmentation algorithm based on K nearest neighbor classification model has many shortcomings such as high time complexity,high spatial complexity,very sensitive to unbalance of samples,poor segmentation effect,mediocre distance calculation method,difficult to segment objects with complex background image,etc.In order to solve these difficulties,a color image coarse segmentation algorithm named CSDT based on CLT,basic decision tree CART and Bayesian Inference,and a color image fine segmentation algorithm named INARF based on neighborhood approximate random forest are proposed,and the combination method of the two is also described.First,the image is converted from RGB to HSI color space,only the H component is extracted,then the H component image is cut into a large overlap patches,and the Haar feature and the LBP feature are extracted.The coarse segmentation contour is obtained by the coarse segmentation algorithm,and we use cover Ratio and Fineness to show the effect.Then the H component image is denoised with median filter and divided into smaller overlapped patches.According to the Ground Truth image distribution,the training data are formed by stratified sampling of these patches,extract the internal color features and external neighborhood features,calculate the semantic distance,combine the coarse segmentation data,and apply the fine segmentation algorithm and the k nearest neighbor classification model to obtain the final segmentation result.The algorithm is verified on the color image segmentation open data sets of Carnegie Mellon-Cornell University named icoseg.The Io U of most of the test data is over 90%,and the MIo U is 81.6%.The experimental results show that CSDTINARF has a better segmentation precision compared to the traditional threshold segmentation algorithm Otsu,the KNN algorithm based on the KD tree,and the ANN algorithm based on the KD tree.