Tomatodiseaserecognitionresearch Based On Multi-scale Feature Fusion

In tomato disease detection,there are a wide variety of diseases that require artificial detection.However,artificial detection can be time-consuming and labor-intensive,and it is also prone to errors,such as misidentifying the type of disease.With the rapid advancement of computer vision and neural networks,target detection technology has significantly improved in its ability to accurately locate the position of objects and classify its category.In the meanwhile,this technology is now being increasingly applied to the detection of crop diseases.This paper aims to detect small-target tomato disease,in the condition of the diseases area being under shadowed,significant shape variations between different tomato plants,and there being many small target objects,the model will be improved accordingly to reduce the rate of the false detection and missed detection in the task,improving the recognition accuracy of the model.The main work of this paper is as follows:(1)Firstly,Four common target detection models listing Faster RCNN,Cascade RCNN,SSD,and Retina Net were selected to compared under the same situation.The target detection model with the best detection performance is selected.(2)Secondly,Aiming at the problem that similar diseases in the tomato disease dataset are in significant shape variations on the same tomato leaf,the feature extracting network of the optimal object detection model is improved,which introducing switchable dilated convolution,using multiple dilated convolutions with different dilated rates for the same input feature.The improved model can better detect objects of the same type and different scales,by obtaining multi-scale information relatively.What’s more,The global module further adopted that fuses context information can improve the quality of the extracted feature.(3)Also,This paper aims to improve the feature fusion network of the optimal target detection model,in response to the issue of the model’s tendency to overlook the importance of small target features during the feature extraction process,as well as the presence of a large number of small target objects in tomato disease datasets due to factors such as leaf curling and occlusion.A feedback mechanism was added to the feature pyramid network to fuse semantic information from each stage,addressing the problem of insufficient semantic abstraction in shallow features and improving the detection performance for small and occluded targets.To address the problem of false alarms caused by similar colors between the disease areas on some tomato leaves and the ground,the loss function of the optimal object detection model is also improved toincrease the importance of accurate samples and reduce the impact of abnormal samples on the model.Finally,by combining the above improvements,an IMS-Cascade model was proposed and applied to the tomato disease dataset.Through ablation experiments,the effectiveness of each improvement was verified.The m AP was improved by 2.5% and the average accuracy was improved by 1.84% compared to the state-of-the-art object detection model Cascade RCNN.Therefore,it can be inferred that the content studied in this paper has high detection accuracy for small target tomato disease recognition.