Research On Multi-target Stored-grain Pest Identification Based On Deep Learning

Stored grain pests will cause heat and mildew in the process of grain storage,leading to food quality and safety problems,and even affect human health.Through the accurate detection and quantity statistics of stored grain pests,the pest control measures can be implemented more efficiently in the granary.In this paper,the OV2640 sensor is used to take samples of stored grain insect samples,and942 original images including 8210 small targets are established.The training set of the original data set is enhanced by means of target image expansion.Three kinds of datasets are obtained,including original,data-enhanced and mixed.Then,a feature extraction layer structure based on the feature pyramid is proposed,and object standard detection model was optimized.Finally,the parameters of the optimized model are adjusted to further improve the target detection accuracy.To verify the effect of target detection after data enhancement,three kinds of training sets are trained on Faster R-CNN,and the same test set is used for detection.After the test,the detection accuracy is 90.40%,92.61% and 90.36%.The results show that the image of stored grain pests not only has more abundant small target information but also improves the matching degree between candidate frames generated by RPN and insect targets.In the optimization test of the detection model,this paper first improves the feature extraction layer structure of the network model,the high-level feature map generated by the feature extraction network gradually fuses the bottom feature map through down-sampling and generates the feature map suitable for multi-scale target detection.Then trains and tests the stored grain pest image datasets on different detection models,and draws the PR curve and F1 score curve of the detection results.The comparison results of the two curves show that the optimized model is more conducive to the detection of small target stored grain pests,and the average F1 score of the optimized model reaches 92.71%,which is4.91% higher than that of the model before optimization.After the hyperparameter was fine-tuned,the average detection accuracy(m AP)of the model reached 97.49%,indicating that the candidate frames generated by RPN could better match pest targets after fine-tuning,and the optimized model had better detection effect than R-FCN and YOLOv3 models.