Study On Methods Of Color Image Segmentation Based On Threshold And Clustering

Image segmentation is the process of extracting the interesting targets of theimage. It applies the certain similarity criterion, and divides the image into multipleregions. Image segmentation is one of the most basic and important tasks in imageengineering. Compared with the traditional gray image segmentation, color imagesegmentation can make full use of the information such as color, space and texture ofthe image, and also can improve the segmentation results. Thus, people today pay moreattention to the study of color image segmentation method. There are many scholarswho are trying to study color image segmentation with diverse methods, so as tooptimize the use of the image segmentation.The threshold segmentation method appears earlier among the segmentationmethods, which is used most commonly today.. This method applied in imagesegmentation always has optimum segmentation results, but still has the disadvantage,which is sensitive to noise, has higher complexity when calculated and makes itimpossible to meet the requirements of real-time segmentation. Segmentation methodis based on clustering and becoming a hot topic in image processing research. And thefuzzy C-means (FCM) clustering method is one of the most representative ones. Thismethod improves the clustering effect compared with hard cluster method, but theproblem of over-reliance the initial cluster centers have not yet been resolved. Thispaper aims at improving the efficiency of multi-threshold segmentation, solving theproblem of FCM initialization, and makes a deep study of color image segmentation.The main work and achievements are as follows:①The paper proposes a fast multi-thresholding method based on the local areaof the images whose color histogram has significant peaks and valleys. This methodexecutes simple clustering based on color component histogram, and then analyzesthe significant peaks and valleys, and study the optimal threshold in the local areaswhich are divided by peaks, which successfully transforms multi-thresholdsegmentation into single threshold segmentation for each local region. It can be aneffective solution to the existing issues such as multiple iterations and highcomputational complexity of traditional multi-thresholding segmentation.②Aiming at the images which have no significant peaks and valleys in the colorhistogram, and when we use FCM method, we will meet the problem of over-reliance on the initial cluster centers, so an improved FCM method is needed. Firstly,preprocess the image by Wiener filter to reduce the effect of noise on segmentation;secondly, use the secondary watershed method to obtain regions which have clearmargins, and then use the centroid to represent the characterizing features of eachregion, which can reduce the computation in the process of centroid filtering and themongering. The method can obtain more accurate initial cluster centers, which not onlyreduces the operation time effectively, but also obtain more accurate segmentationresults.③Compare the methods mentioned above with the current methods, whichindicate that the new methods can significantly improve the effect of the segmentationresults and processing efficiency. In addition, form a color image segmentationprototyping system by two methods, which can choose different segmentation methodsbased on the histogram of the input image. If there are obvious peaks and valleys in thehistogram, choose fast multi-threshold segmentation method; otherwise choose theimproved FCM method. The experimental comparison shows that the two newmethods have different performance characteristics.