Abstract:Seed immersion is an important pre-sowing seed enhancement technology in maize production, and efficient detection of cracks during seed immersion is the basis for analyzing the change pattern of endosperm cracks during seed immersion, which is one of the keys to the selection and breeding of good varieties of traits, and there are still difficulties such as internal endosperm cracks are not visible and the degree of automation is not high. Based on CT scanning technology, a rotating target detection network named YOLO v5-OBB was designed based on YOLO v5n detection network, where OBB used rotating rectangular box instead of normal rectangular box and added CA model in the Backbone part. The network used a rotating rectangular box instead of a normal rectangular box, and added CA model in the Backbone part, and also used Skew-NMS for non-maximal suppression to obtain the final prediction box, so as to achieve the detection of corn endosperm cracks with relatively large length and width and different directions. After 300 iterations of training, the model had a precision of 94.2%, a recall of 81.7%, and an average precision of 88.2% on the test set, with model size of 4.21MB and average detection time of 0.01s for a single image, which improved the AP value by 15.0, 16.9, and 7.0 percentage points compared with the SASM, S2A-Net, and ReDet models, respectively, and the average detection time of single image was reduced by 0.19s, 0.22s, and 0.46s, respectively, while the YOLO v5-OBB model size was 1.50%, 1.43%, and 1.73% of the SASM, S2A-Net, and ReDet models, respectively, with an increase in AP value of 0.6 percentage points, a decrease in model size of 0.19MB and an unchanged average detection time of 0.01s for a single image compared with that of the YOLO v5 network with horizontal rectangular box labeling. Comparing the crack length information obtained from the YOLO v5-OBB network after obtaining the crack target frame coordinate information with the real length of the crack obtained in DragonflyEZ software, the absolute error of both was 0.04mm and the relative error was 0.93%. The results on the detection of corn endosperm cracks with different CT gray value distributions showed that the model had P values of 100%, 100%, and 93.3%, R values of 100%, 82.4%, and 79.8%, and AP values of 99.5%, 91.2%, and 86.8% for the three types of corn endosperm crack images with smaller gray values, larger gray values, and mixed gray values, respectively. The results showed that the designed model can effectively achieve the detection of corn endosperm cracks, and at the same time, the model was highly robust and took up little storage, which can provide necessary technical support for the automatic monitoring of corn endosperm cracks during seed dipping.
