Abstract:

The rapid and accurate identification of feeding targets is a crucial guarantee for intelligent feeding assistance robots. Balancing segmentation accuracy and operational efficiency is a key aspect of ensuring the comprehensive performance of algorithms and a significant challenge for recognition methods. To address the issue of matching segmentation accuracy and efficiency in existing methods for identifying cow feeding targets, this paper proposes a real-time feeding target instance segmentation (RTFIS) based on Segment Anything Model (SAM) optimization. Building on the SAM-det architecture, the method introduces lightweight parameter designs for the image encoder and object detector, along with a parallelized buffer queue approach to balance the operational efficiency of each module, significantly improving inference speed. The use of HQ-token enhances feature space decoding capability, optimizes the design of the mask decoder, and employs a phased training strategy tailored to feeding targets, thereby improving segmentation accuracy. Research and experimental results show that the proposed method ensures segmentation efficiency while enhancing segmentation accuracy. In the task of cow feeding target recognition, the method achieves a segmentation accuracy of 98.7% for cows, 96.4% for feed, and a processing speed of 52.9 FPS, meeting the application requirements for cow feeding target recognition in complex environments.

Key words: segment anything model, cow feeding, target recognition