Resumen
Due to their rich nutritional value, kidney beans are considered one of the major products of international agricultural trade. The conventional method used for the manual detection of seeds is inefficient and may damage the test object. To locate and classify different kidney bean seeds rapidly and accurately, the Yolov3 network has been improved to realize seed detection in the current paper. Firstly, a dataset of 10 varieties of kidney bean seeds was produced and 1292 images were collected. Then, the dataset was divided into the training, validation, and test sets with the assigned ratio of 8:1:1. The kidney bean seeds dataset was trained using the Yolov3 model. Additionally, the implemented speed needed to be guaranteed while satisfying the detection accuracy. To meet such detection requirements, the Yolov3 model was pruned using the scaling factors of the batch normalization layer as a measure of channel importance, and finally fine-tuned with the aid of knowledge distillation. Then, the Yolov3, Yolov3-tiny, Yolov4, and the improved Yolov3 were used to detect the images in the test set. Subsequently, the performances of these four networks were compared. The results show that the model pruning method can compress the model to a great extent, and the number of model parameters is reduced by 98%. The detection time is shortened by 59%, and the average accuracy reaches 98.33%. Considering the speed and mAP, the improved Yolov3 detected the best results. The experimental results demonstrate that the method can accomplish the rapid and accurate detection of kidney bean seeds. It can provide a solid foundation for the marketing and planting of kidney bean seeds.