Traditional supervised learning methods have limitations in labeling speed, scene adaptability and accuracy. Unsupervised learning methods do not require labeling data and can automatically extract the laws, which provides a new idea for image segmentation, especially in medical diagnosis, automatic driving and other highprecision requirements of the scene has an important application value. This study explores unsupervised learningbased image segmentation methods in computer vision, focusing on the improved kernel fuzzy C-mean clustering algorithm (KFCM). The study constructs an image segmentation algorithm with noise robustness by introducing a kernel function instead of Euclidean distance and combining it with super-pixel segmentation technique. The experiments are validated on synthetic, natural and medical images and compared with various classical algorithms. The results show that when 30% Gaussian noise is added to the synthetic image, the segmentation accuracy of the KFCM algorithm reaches 99.8%, which is 12.6% higher than that of the traditional FCM; in the segmentation of the natural image with the addition of mixed noise, the average value of the segmentation coefficient of the KFCM reaches 96.45%, which is 17.47% higher than that of the FCM, and the segmentation entropy is reduced by 34.57%; and in the segmentation of the medical cell image, the KFCM algorithm shows good edge keeping ability in complex noise environment. The study shows that the improved KFCM algorithm significantly improves the image segmentation accuracy and anti-noise performance through the adaptive neighborhood information and kernel mapping, and provides an effective solution for unsupervised image segmentation, which is of practical application value for medical diagnosis, automatic driving and other fields.
