Research On Color Image Clustering Segmentation Algorithm Based On Superpixel And Transfer Learning

As the first step of image processing,image segmentation is the basis of image analysis and pattern recognition.With the development of science and technology,the image information that grayscale images can provide is more and more difficult to meet the needs of people.Therefore,people gather their eyes more and more on color images.The FCM algorithm and k-means algorithm are the most classical clustering segmentation techniques,which are simple and can obtain the characteristics of effective clustering results.However,due to the complexity of color images and the interference of noise and other factors,there still exist some defects in the actual segmentation tasks,such as inaccurate clustering results.Aiming at the shortcomings of FCM clustering algorithm in the color image segmentation tasks,which is susceptible to noise interference and inaccurate segmentation,this paper proposes a segmentation method that combines the improved SLIC algorithm and transfer learning to optimize the clustering performance of FCM algorithm.The algorithm takes the clustering center of the original image obtained by FCM algorithm as the historical knowledge,and takes the image processed by SLIC algorithm as the target domain,and completes the segmentation task by using the FCM algorithm which combines the transfer mechanism.In addition,aiming at the defect that the SLIC algorithm does not fit at the edge of color image,this algorithm is improved by improving the proportion of luminance in color distance measurement.The improved SLIC algorithm has a more suitable edge than the original algorithm at the edge blur.The experimental results show that compared with the classical FCM algorithm and FCM algorithm combined with SLIC superpixel,the proposed algorithm shows better clustering performance and more accurate image segmentation results.K-means++ is an algorithm that has been improved by an initialization strategy.This algorithm provides a better selection mechanism for initial clustering centers,which avoids the negative effects of artificial selection of initial clustering center randomness.However,the algorithm still has the shortcoming of inaccurate segmentation results,and even the phenomenon of miss-segmentation.Therefore,this paper combines the k-means++ algorithm with the superpixel algorithm and the principle of transfer learning to achieve the purpose of image segmentation.The superpixel algorithm can make full use of the spatial organization relationship between pixels.The transfer learning part can add historical knowledge to the current segmentation task to modify the segmentation results.In this paper,a transfer mechanism suitable for the current situation is proposed based on the clustering algorithm.Due to the similarity of geometric distribution between the source and target data,based on the geometric similarity strategy,this paper adds penalty items to the k-means++ algorithm to make the clustering algorithm have the ability of knowledge migration.The experimental results show that the algorithm can effectively overcome the miss-segmentation of the k-means++ algorithm and improve the accuracy of segmentation.