Research On Detection Method Of Sesame Tree Pests Based On Deep Learning

Sesame has many nutritional values and functional values,which will produce great economic and commercial value.In modern agriculture,sesame yield is greatly affected by pests,and it is necessary to identify pests in a timely,accurate and effective manner.Based on this direction,this dissertation develops an intelligent pest positioning and identification system.The main work and research results are as follows:1.Screen the public pest data set Pest24 for detection tasks,and finally select 15 sesame pests as new experimental data named Pest15,and analyze the difficulties of the data set pictures from multiple angles.Among them,22,841 images are used as training set,and 2,539 images are used as test set.Four kinds of target detection algorithms are used to conduct a benchmark experiment on this experimental data set.According to the model evaluation index,Yolo V5 s is selected as the benchmark model for subsequent experiments.2.Through the visual analysis of the pest annotation boxes in the Pest15 dataset,according to the phenomenon that the number of small target pests in the dataset is high,using the unimproved Yolo V5 s network training will cause the small target pests to lose features in the deep feature map,and in the deep feature map.There will be too much background noise in detecting small target pests in deep feature maps,resulting in a large number of missed detections and false detections.In view of this situation,an improvement was made on the basis of Yolo V5 s,and the FF-CBAM-Yolo V5 s network model was proposed.In order to retain the characteristics of small target pests in the low-level feature map,the three detection layer modules will be merged with other feature maps of the same size,and with the addition of the CBAM attention mechanism module,the m AP of the final improved network reaches 79.7%,which is 2.8% higher than the m AP value of the benchmark model Yolo V5 s model.Compared with other mainstream target detection algorithms,the method proposed in this dissertation achieves the best detection effect on the Pest15 data set constructed,which further verifies the effectiveness of the FF-CBAM-Yolo V5 s network model on the task of crop pest detection.3.In order to facilitate the operation of agricultural related personnel,the pest identification system is developed based on the Py Charm development platform and the Py Qt5 framework.Finally,various functions and performance tests were carried out for pest identification.The results show that the pest detection and identification method proposed in our work and the visual detection system constructed according to the model can effectively complete the pest detection task and have certain practical significance.