Study On Intelligent Tracking And Classification Method Of Cotton Leaves With Verticillium Wilt Based On Image Series

Cotton is one of the most important economic crops,widely planted all over the world.Meanwhile,in the world’s major cotton areas,because of its extensive transmission route,serious harm and complex infection mechanism,verticillium wilt is regarded as the main disease of cotton production.The cotton verticillium wilt is generally infected by soil fungi including verticillium dahliae,which causes the leaves wilt,fade and fall off.More seriously,the growth and development of cotton will slow down or even wither,which finally results in the decline of cotton quality and yield.Therefore,the accurate evaluation of cotton verticillium wilt is crucial to the cotton disease resistance research.The infected leaves are classified four levels by national standards.Traditional detection method mainly relies on manual work,which is subjective and inefficient.In order to obtain the accurate number and classification of ill leaves,this paper has proposed a tracking and counting method based on the VFNet-Improved,Deep Sort and mask collision mechanism to realize ill cotton leaf measurement.（1）The rotating video capture experiment on 100 cotton varieties was conducted based on a multi-perspective cotton leaf imaging device.The videos were randomly divided into object detection set and multi-object tracking set according to 8:2 ratio.The open source software Darklabel and Labelimg were used to annotate and fine-tune the object detection set.Meanwhile,image augmentation methods were applied.Finally,the annotation files were converted into json format by self-designed script and divided into training set and testing set.（2）The cotton leaf recognition was established based on six deep learning models（Cascade R-CNN,Faster R-CNN,Retina Net,SSD,VFNet and YOLOv5）,in which the VFNet model got the best performance with 0.898 m AP,0.894 AP_ill,0.902 AP_fine and 14.0frames/s.And then,considering the scale difference and overlapping of cotton leaves,especially ill leaves,deformable convolution,deformable Ro I pooling,multi-scale training and Soft NMS optimization methods were adopted to improve the original VFNet detection network.The m AP of VFNet-Improved model was promoted by 1.3%with fps 12.4frames/s.Besides,the images obtained by phenotypic information acquisition platform were detected by VFNet-Improved model.On the 9th day,the prevalence rates of all varieties were lower than 20%.And the prevalence rates of 3 susceptible varieties accelerated from the 10^th day.After the 16^th day,the prevalence rates were all more than80%.The prevalence rates of resistant varieties were lower than 20%on days 9 to 12 and increased slowly from the 12th day to the last day,and the final prevalence rates were all lower than 40%.（3）Because the object detection of a single view could not obtain the grade of all diseased leaves,multi-object tracking method based on VFNet-Improved model was used to realize the match of the same target in different frames through Kalman filter and Hungarian algorithm.The tracking results MOTA of Sort and Deep Sort were 0.654 and0.839 respectively,and Deep Sort increased by 18.5%compared with Sort.（4）In order to solve the ID Switch problem,the mask collision mechanism was developed.Meanwhile,Open CV was applied to conduct feature extraction and cotton leaf classification.The counting results showed that R²,MAE,RMSE and MAPE were 0.902,4.150,4.914 and 14.472%respectively,and the Recall values of SL1,SL2,SL3 and SL4were 0.91,0.82,0.80 and 0.86 respectively.Finally,the leaf tracking and analysis software for cotton verticillium wilt was designed.The software interface was concise and easy to operate,which could be transplanted to other computers to use.