Research On Detection And Classification Of Corn Kernels Based On Deep Learning

Food is essential for human production and life,however,during the storage of food,it may be contaminated by mildew and insects,causing loss of food during storage.Therefore,detecting insects and mildews has important practical significance for the infection of food in storage.When traditional image methods are used to detect food kernels,they rely on the accuracy of feature extraction,the extracted features have a greater impact on the detection results.For the detection of corn kernels,this paper studies the object detection method based on deep learning,Proposed improved Faster RCNN(Region-based Convolution Neural Networks)object detection network model,and applied to the detection and classification of corn kernels.The improved Faster R-CNN object detection model can extract the characteristics of corn particles in the image automatically,detect and classify the corn kernels,and achieve good detection and classification results.The main contents of this paper include the following sections:The image dataset of corn kernels set was collected,including mixed images of undamaged corn kernels,mildewed-damaged corn kernels and insect-damaged,and the images were uniformly pre-processed and information annotated according to the corn kernels detection requirements,a dataset containing the coordinates and types of corn kernels in the image was obtained.This paper proposes an improved Faster R-CNN object detection network model to identify corn kernels,the performance of the network model is improved by using dilated convolution,transposed convolution,Squeeze-andExcitation(SE)modules.The dilated convolution does not need to introduce additional parameters.without increasing the calculation amount of the network model,the receptive field of the feature map can be increased.The resolution of the feature map extracted by the feature extraction network can be increased to more accurately locate the object and improve the detection accuracy rate of corn kernels.The introduction of transposed convolution in the Feature Pyramid Network(FPN)can increase the semantic information obtained by the feature pyramid and build a feature pyramid that is more effective for the small object detection.Introducing the SE module in the feature extraction network can weighted the channels of the feature map extracted by the network model,enhance the effective information,suppress the invalid information,and make the extracted features more obvious.This paper also introduces Region of Interest Align(RoIAlign)when the candidate box is mapped to the feature map in Faster R-CNN model,introduces Soft Non-Maximum Suppression(Soft-NMS)in the candidate box deduplication stage,and the accuracy of the network model was further improved.This paper adopts Intersection-over-Union(IoU)balanced sampling,guided anchoring and key point detection methods to improve the Faster R-CNN model.IoU balanced sampling performs difficult sample mining on the anchor points generated in the network model,improves the training efficiency of the network model,improves the training speed of the network model and detects mAP;guided anchoring guides the number and location of anchor points by judging the probability of generating anchor points at each position in the feature map,reducing the amount of calculation when extracting anchor points;the key point detection method adopts a new anchor point generation strategy,using dot matrix instead of anchor points,to better fit the position of the object in the image.This paper also combines these three improved methods with hollow convolution,transposed convolution,and SE modules to improve the Faster RCNN model to further improve the network model's detection of corn kernels.Detection mAP reached 97.7%.The improved Faster R-CNN model proposed in this paper can effectively detect and classify corn kernels,greatly improve the detection efficiency,and also provide a reference for the detection of other grain kernels.