Aimed at the problems that the supervised learning model for image segmentation is long in training time, a large number of training samples is needed to ensure model accuracy requirement, and the labeling takes a lot of work, time and energy, an unsupervised image segmentation method is proposed based on neural network in different color space. Firstly, the images are transformed into different color space models to obtain the color representation of images in different color gamut spaces, and then by using felz and quickshift methods, a coarsegrained cluster is made after the images being transformed to form superpixel results and label each pixel accordingly. Finally, the finegrained image feature discrimination ability of neural networks is utilized for making a fineturning, obtaining the final image segmentation results. The method is validated on publicly available datasets such as COD10K selected, and the experiments prove that the proposed method can segment images reasonably with less inference time consumption and faster speed in comparison with the supervised training for a long time.