In order to study a suitable intelligent root cutting device for garlic combined harvesting, a non-contact bulb root cutting method with machine vision was proposed, and a garlic root cutting test bench based on a deep convolutional neural network was designed afterwards. Specially, the test bench adopted a deep learning theory to perform target detection on the collected images, through using the APP software in Matlab to complete the human-computer interaction. Then, the results presented that the deep convolutional neural network could determine the cutting position of the garlic root, and the motor control system could adjust the position of the double disc cutting automatically, ensuring the root cutting process completed by the root knife. Target comparison tests showed that bulb (availability rate was 94.79%, confidence score was 0.97697) was suitable for detecting, among the three kinds of bulb, root plate and garlic root. Comparison tests of detection models performed with ten models based on Faster R-CNN, SSD, YOLO v2, YOLO v3 and YOLO v4. The improved YOLO v2 model combined the detection speed and accuracy (the detection time in the test program was 0.0523s, and the confidence score was 0.96849), where ResNet50 was selected as the feature extraction network;by using the improved YOLO v2 model, the root cutting test took bulbs as the targets (the confidence score was 0.97099, the availability rate was 96.67%, the qualified rate of cutting roots was 95.33%, and the detection time in the APP was 0.0887s), can meet the requirements of garlic combined harvesting and cutting roots.