Aiming at the problem that parts packaging boxes composed of manual sorting often have missing parts, an intelligent sorting system integrating training, recognition and sorting is developed. The system can detect the preset objects appearing on the screen through one shot of the product in the box, and compare it with the standard settings to determine whether the product in the box is short of material or has too much material, so that qualified and unqualified boxes can be selected. In the design process, an improved Yolov3 algorithm based on deep learning was proposed. In view of practical factors such as industrial site lighting, industrial part shape and texture, the training and detection of the Yolo algorithm were improved. In the case where the objects placed overlap each other within 20%, the accuracy rate of object detection is 98.2%, and the recall rate is 99.5%. Through the improved algorithm proposed in this paper, the designed detection system can work normally in a complex industrial field environment, and can accurately detect the integrity of the packaging.