Research on ovine meats intelligent segmentation remains limited because of the low recognition accuracy of target muscle region image caused by preprocessing process. A method of target muscle region recognition in ovine hind leg intelligent segmentation based on R2U-Net and dense atrous convolution algorithm was presented. The traditional U-Net semantic segmentation network was taken as the backbone network and improved. The convolution blocks in the feature encoder and decoder of the original U-Net were replaced with the residual recurrent convolution blocks to avoid the gradient loss of the U-Net and a four branch dense atrous convolutional module was added between the feature encoder and the feature decoder to code multiscale semantic features. On the one hand, aiming at the sartorius muscle region, the ovine hind leg images were collected to build a dataset and the model was trained and tested using the dataset to validate the accuracy and realtime performance of this method; on the other hand, the homogeneous transformation matrix of gripper coordinate system, camera point clouds coordinate system and robot coordinate system was calibrated based on screw theory to calculate the segmentation path, and the robot cutting manipulation was controlled by an active compliant force/position hybrid control method, which validated the feasibility of target muscle segmentation based on the target image obtained by this method. The experimental results showed that when the intersection over union (IOU) was 0.8588, the average precision (AP) of the proposed method was 0.9820, which was better than that of R2U-Net (0.8324, 0.9775); the average time of single sample detection was 82ms, which showed that this method can segment the sartorius image quickly and accurately, which met the realtime requirements of robot autonomous segmentation system, and it was better than U-Net, R2U-Net and AttU-Net algorithms. Finally, based on the image of sartorius muscle obtained by this method, the real robot segmentation experiment was carried out. The average time of robot cutting on five sheep hind legs was 7.9s, the average offset distance was 4.36mm and the maximum offset distance was not more than 5.90mm, which met the accuracy requirements of sheep hind leg boneless segmentation.