Research On Tomato Disease Detection Based On Convolutional Neural Network

Tomato is an important cash crop in China.It has a wide planting area and high cultivation benefits,and occupies an important position in vegetable production and vegetable trade.But disease is one of the key problems that restrict tomato output and quality.Traditional tomato disease recognition methods are artificial recognition methods,but they are difficult to meet the requirements for efficient production in modern agricultural development.The symptoms of tomato diseases usually appear on leaves.The color,texture and shape of diseased leaves are abnormal,and different types of diseased leaves have different external characteristics.So the accurate analysis of tomato leaves is helpful to know the health status of tomatoes in the planting process,and is a precondition for real-time diseases monitoring and control.As for tomato fruits,physiological diseases are important factors affecting their health.Tomato fruits with physiological diseases will show obvious external characteristics,and the external characteristics are important indicators for evaluating fruit quality.Therefore,realizing efficient and automatic detection of tomato leaf diseases and tomato fruit physiological diseases can promote tomato growth monitoring and disease treatment,and helps to strengthen tomato fruit quality management and automatic screeningWith the advent of deep learning,its representative algorithm,convolutional neural network(CNN),has developed rapidly after the 21st century and has been widely used in image recognition and object detection.As a high-performance image recognition technology,convolutional neural network provides new ideas for tomato disease detection.However,most existing studies of tomato disease detection based on convolutional neural networks often used two-stage object detection methods,and such methods have shortcomings in detection speed and are difficult to meet real-time detection needs.In addition,most of the studies only focused on diseased tomatoes,and lacked the researches about the impact of healthy tomatoes under real conditions.And in the studies on diseased tomatoes,tomato leaves were usually used as research objects,and tomato fruit were often ignoredAccording to the above problems,this thesis selected YOLO algorithms,the representative algorithms of one-stage object detection methods,as the experimental models to do researches of tomato disease detection(1)In view of the excellent detection speed and detection performance of YOLOv2,we conducted a research on the detection of tomato fruit common physiological diseases based on YOLOv2.We used k-means clustering algorithm to cluster the appropriate anchor boxes of the tomato fruit training dataset,and applied the improved YOLOv2 model to the detection of seven types of tomato fruits.The results showed that the improved model effectively improved the detection performance.In addition,we explored the impact of image data type and image data size on detection.The comparison experiment showed that the RGB complete images under the state of balanced data were more suitable for detecting tomato fruit common physiological diseases(2)In view of the excellent detection performance of YOLOv3 and the nuance and similarity of tomato leaf diseases,we conducted a research on the detection of tomato leaf diseases based on YOLOv3.After clustering the appropriate anchor boxes of the tomato leaf training dataset with k-means clustering algorithm,we reduced the original three feature detection scales to two.In addition,the DenseNet was appropriately combined to construct a new model for tomato leaf disease detection——Dense-YOLOv3,and the model was applied to the detection of ten types of tomato leaves.The comparison experiment showed that the Dense-YOLOv3 model effectively improved the performance of tomato leaf disease detection.