Abstract:Pork belly,as a high-value product,is required to be automatically and accurately graded and inspected under random postures on the automatic segmentation production line,so that customized classification and processing at different grades can be subsequently performed.Traditional graded and inspected methods using steel rulers to measure the fat thickness in cross-sections suffer from low efficiency and poor accuracy,failing to meet production demands.A visual measurement method for fat thickness in pork belly cross-sections was proposed based on deep learning （image segmentation）.The research relied on automatic segmentation and feature point extraction of pork belly cross-sections based on deep learning,three-dimensional reconstruction of fat measurement features of cross-sections through binocular vision,and a measurement method for fat thickness in cross-sections based on locally boundary-oriented vector determination to address the challenges of low measurement precision and reliability caused by random postures.On-line tests of fat thickness in crosssections on 20 randomly selected pork belly samples demonstrated an average measurement error of 0.37 mm and a relative error of 1.44%,showing significant improvement over manual methods.