Research On Cotton Boll Detection And Segmentation Method Based On Computer Vision

Applying smart agriculture to the cotton growing process can provide new technical strength and scientific methods for growing and producing cotton crops.Research into boll segmentation and detection can help to detect and prevent pests and diseases early,improve final yields,save time and increase safety,reduce human error and significantly improve the science and accuracy of decision-making.There are several key problems with boll segmentation and detection methods:the detection of mature opened cotton bolls is mainly based on aerial images and work on close up images of cotton plants in real scenarios is scarce;the detection of green unopened bolls is very scarce and has not yet progressed.At the same time,the green unopened cotton bolls has a high similarity to the background,and it is difficult to distinguish them effectively by relying on manual design and extracting features;the color distinction between the white opened cotton bolls and the background is more obvious,but there are more cases of boll overlapping and being separated by branch obscuration.Based on the above problems,this thesis analyzes and proposes methods for unopened green cotton bolls in real cotton fields and mature opened cotton bolls in close-up images,specifically:1.An improved YOLOv5 target detection algorithm is proposed,introducing various more novel and effective methods:replacing the residual unit structure with Dense Block,introducing the SE attention mechanism,replaced the Bi-FPN structure in the feature fusion network part,and adjusted the structure of the feature extraction network to output a larger feature map for detection from a shallower network layer.The above improvements can improve the detection accuracy of unopened green cotton bolls,and at the same time reduce the parameters and volume of the model and decrease the computation,making the model more lightweight and achieving a detection speed of 25ms per image to meet the needs of practical applications.2.A segmentation and detection algorithm for white cotton bolls at maturity is proposed,using S-channel features in HSV color space for pre-segmentation,and the presegmentation results are as mask matrix initialized GrabCut algorithm for further segmentation,and after analyzing and extracting feature parameters for each connected region,a region merging algorithm and a region splitting algorithm based on distance transform and circular threshold segmentation are designed for segmentation and detection,respectively.In this work,image data were collected manually from Aodu Water Control Farm in Kashgar,Xinjiang,and organized and data labeled according to specific needs.For green boll detection,ablation and comparison experiments were conducted to verify the effectiveness of the improvement and its superiority over other methods.For the segmentation and detection of mature opened cotton bolls,the results of each method step are presented after each step,and the effectiveness of the segmentation algorithm is finally verified by comparison experiments.The effectiveness and accuracy of the proposed segmentation method is verified by using the evaluation metric Zeboudj based on regional contrast and designing an evaluation metric F’based on the mean value of color difference in the region.