Research On Bean Seed Detection Method Based On Compressed YOLOv3 Mode

The kidney bean is one of the world most widely consumed grains,containing many vitamins and rich in nutrients.Kidney bean seeds are the source of kidney bean growth,however,they are prone to mix-ups during the cultivation process.It is crucial to digitize and automate the management of agriculture to quickly and accurately classify different kidney bean seed varieties.It could promote the development of the kidney seed cultivation production and marketing industries.Traditional seed variety identification methods rely heavily on manual identification and expensive instrumentation.To accomplish rapid and accurate identification of seed varieties using a low-cost platform,the main elements of the research in this paper are as follows:Firstly,1292 images were captured by an image acquisition device,and the images were annotated to create a dataset of kidney bean seeds.The dataset was divided into training,validation and test set with a ratio of 8:1:1.To complete the identification of multiple kidney bean seed varieties on a single image,the dataset was fed into the singlestage target detection networks YOLOv3,YOLOv3-tiny,and Mobilenetv2-SSD for training.The test images were tested to compare accuracy and speed.A method of compressing the model was proposed to address the redundancy problem of the parameters of the YOLOv3 bean seed recognition model.The model was first trained to be sparse,and the appropriate sparse factor could be determined by the variation of the scaling factor,and the loss of sparse training.The channels and layers were then pruned to obtain a more concise and compact model.The m AP of the pruned model slipped,and the knowledge distillation algorithm was introduced in the fine-tuning to bring the model m AP back further.The performance of the compressed model was compared with YOLOv3,YOLOv3-tiny and Mobilenetv2-SSD.Considering the requirements of cost,power consumption,and workspace in the recognition process,the embedded platform Jetson TX2 was used to perform variety recognition of rape bean seeds.The tensor and convolutional layers of the model were merged by the TensorRT framework.The network trained by the Pytorch framework was first converted to the ONNX format.Then,the ONNX format converted to an engine file.The 32-bit floating-point model was improved to a half-precision 16-bit floating-point model to inference on the variety of kidney bean seeds.Finally,the accuracy and speed of the deployed models were compared.