Aiming at the problem of many noisy points and blurred boundaries of highly adherent and heavily stacked fresh pellet images, a method for edge detecting and segmenting of pellets is proposed by a deep learning image semantic segmentation model (CBAM-Unet) that incorporates attention mechanism. The method steps:1) sample data set preparation, collecting 30 images with different states including depth separation, high adhesion to serious stacking, etc., and forming a sample data set of 900 images by data augmentation;2) model building, and incorporating the Convolution Block Attention Module (CBAM) into the Unet network model to form the CBAM-Unet model for fresh pellet edge detection and segmentation. The results show that, compared with the Unet method, the similarity coefficient (e_Dice), the category-average pixel accuracy (e_mPA), and the average intersection ratio (e_mIoU) of CBAM-Unet method are improved by 9.8%, 9.7%, and 14.7%, respectively. The recognition model proposed in this paper can improve the accuracy of pellets segmentation and lay a foundation for practical applications.