Wheat Kernel Quality Analysis Based On Deep Learning

Food safety has become the focus of public attention.The health and safety of wheat kernels is an important part of food safety.Rapid and accurate detection of wheat kernel quality is an important part of flour processing process.The analysis of wheat kernel quality only by artificial experience has problems such as slow detection speed,low accuracy and poor stability,which can’t meet the current demand.With the development of intelligence in the food processing industry,some wheat kernel analysis methods based on near-infrared,hyperspectrum and convolutional neural network classification models have been proposed,but it cannot meet the requirements of high detection accuracy and fast detection speed at the same time.In order to solve the problem of rapid and accurate analysis of wheat kernel quality,this paper uses the wheat kernel quality analysis method based on computer vision.This paper designs a machine vision image acquisition system and establishes a wheat kernel image dataset containing 1746 images,which contains a total of 7844 wheat kernels including germinated grains,gibberella grains,mildewed grains and normal grains.This paper fully studies the application of the deep learning object detection algorithm in wheat kernel quality analysis and carries out research on wheat kernel quality analysis based on improved YOLOX and improved YOLOv5 on this basis:(1)This paper proposes a wheat kernel quality analysis algorithm based on YOLOX_ns.Compared with the method based on YOLOv5_nc,it is more suitable for scenes with large changes in the size of wheat kernels.This algorithm improves the CSPDarknet structure and the F-PAN structure of the anchor-free algorithm YOLOX according to the characteristics of the wheat kernel detection scene,such as distinct grain characteristics and the simple background.At the same time,the width of the model is reduced,and the attention mechanism is integrated into the model to strengthen the detection of local features of wheat grains.The YOLOX_ns model is obtained.The experimental results show that the classification accuracy of the wheat kernel quality analysis algorithm based on YOLOX_ns reaches 97.9%,and the mAP reaches 98.4%.The detection speed on the GTX1050 is 49.3FPS,and the detection time of 100 grams of 2500 wheat kernels is 13s.The algorithm meets the accuracy requirements,bounding box regression requirements and speed requirements of wheat kernel quality analysis.In complex scenes with large changes in wheat kernel size,the classification accuracy of the algorithm based on YOLOX_ns reaches 95.1%,and the classification accuracy of the algorithm based on YOLOv5_nc can only reach 91.3%.Therefore,the wheat kernel quality analysis algorithm based on YOLOX_ns is more suitable for detecting complex scenes with large changes in wheat kernel size.(2)This paper proposes a wheat kernel quality analysis algorithm based on YOLOv5_nc.Compared with the algorithm based on YOLOX_ns,this method is suitable for deployment in lightweight scenes with small changes in wheat kernel size and rapid detection.In view of the deficiency that the YOLOX_ns model cannot be further improved in light weight,this algorithm improves the Backbone structure of the anchor-based algorithm YOLOv5,reduces the width of the model,and redesigns the Neck structure.The YOLOv5_nc model is obtained.The redesigned Neck structure can reduce the size of the model,accelerate the convergence of the model,and increase the detection speed while retaining the detection of the wheat kernels,which is more applicable in the scene of wheat kernel quality analysis.The experimental results show that the classification accuracy of the wheat kernel quality analysis algorithm based on YOLOX_ns reaches 98.4%,and the mAP reaches 98.5%.The detection speed on the GTX1050 is 57.8FPS,and the detection time of 100 grams of 2500 wheat kernels is lls.The algorithm meets the accuracy requirements,bounding box regression requirements and speed requirements of wheat kernel quality analysis.YOLOv5_nc has only 3.5GFLOPs,which is 25%of the YOLOX_ns algorithm.The wheat kernel quality analysis algorithm based on YOLOv5_nc not only has a better detection effect,but is also lighter and more suitable for deployment in low computing power environments.Comparative experiments show that the wheat kernel quality analysis algorithms based on YOLOX_ns and based on YOLOv5_nc have the advantages of high detection accuracy and fast detection speed compared with similar algorithms.