Abstract: To address problems of low mechanization level and labor-intensive sorting tasks before packaging of hydroponic lettuce, an automatic sorting system for abnormal hydroponic lettuce was designed in combination with the deep learning method. The automatic sorting system was composed of an information perception sub-system, an information processing sub-system, and a sorting action execution sub-system. Hydroponic lettuce classification was based on the difference between abnormal and normal leaves. Three cameras from bottom to top were used to capture images. Real-time processing of hydroponic lettuce images was realized based on semantic segmentation DeepLabV3+. The image segmentation model had mIoU of 83.26%, PA of 99.24% and image processing velocity of（193.4±4） ms/frame. To realize sorting of abnormal hydroponic lettuce, a bracket-type hydroponic lettuce sorting sub-system was designed based on phenotype and harvesting mode of the hydroponic lettuce. Quadratic orthogonal rotational-combinational experiments were designed. Experiments on factoring in horizontal and longitudinal support rod angles and stepping motor speed were conducted to obtain the highest sorting success rate. Regression mathematical models between factors and index were multi-objectively optimized by using Design-Expert software. Optimal combination of parameters was obtained, including the horizontal support rod angle of 146°, the longitudinal support angle of 150°, and the stepping motor speed of 11 r/min. Perform test was carried out according to the optimal combination of parameters. The sorting success rate of the sorting action execution sub-system was 98%, and the sorting success rate of the abnormal hydroponic lettuce automatic sorting system was 95%, which met technical standard requirements of lettuce refrigerated transportation.