Study On Unopened Cotton Boll Detection And Counting Methods Based On Deep Learning

With the rapid development of agricultural robotics and automation technology systems in recent years,the studies of computer vision-based crop detection and counting have increasingly become a research hotspot.Accurate detection and counting of unopened cotton bolls at the early stage of cotton maturation is an effective way to develop crop load management and harvesting strategies in advance,which can help monitor growth information dynamically.However,robust and effective detection and counting of unopened bolls under complicated field conditions is a challenging task.The current research has three main problems:first,there are few publicly available datasets of unopened cotton bolls taken in real scenarios,and related work on unopened boll detection is scarce;second,the detection accuracy is low due to the smaller size of unopened bolls compared to opened bolls and their similarity to numerous background disturbances;and third,most of the existing algorithms use deep neural networks,but with the increase of convolutional layers,there are disadvantages such as large memory occupation and poor real-time performance,which cannot meet the demand for realtime monitoring of cotton bolls in real field environment.Therefore,in this thesis,the following two detection and counting methods are proposed to solve the above problems by taking an unopened cotton boll as the research object.(1)A deep learning method based on YOLOX for multi-receptive field extraction module is proposed for detecting and counting small targets in order to solve the problem of high missed detection rate due to the small proportion of unopened cotton bolls in the image.In the target detection part,firstly,multi-scale residual blocks and SE attention modules are introduced to enhance the extraction of cotton boll feature details.Secondly,a multi-receptive field extraction module is added to reduce the loss of small target features in the deep network.Finally,a smalltarget detection layer is integrated to improve the detection accuracy.Compared with existing methods,the model substantially improves the missed detection of unopened cotton bolls while maintaining a high processing speed,and achieves accurate detection of small target cotton bolls under real planting conditions.(2)Further analysis of the causes of false detection of unopened cotton bolls revealed that it was caused by the color characteristics of the surrounding leaves being too similar to the cotton bolls.There is a high similarity between the unopened cotton bolls and its background,which has similar features to the camouflaged object.Therefore,a deep learning method based on YOLOv7 is proposed.By introducing an edge-aware extraction module which is in the concept of camouflage detection,the edge information in the low-level feature map is deeply mined.In order to meet the real-time requirements,a lightweight feature extraction module is used to replace the original multi-branch stacking block to further reduce the number of parameters of the model.The structure of the feature extraction network is also adjusted,a depthwise separable convolution replaces the standard convolution and a lightweight upsampling operator is introduced.The above improvements reduce the false detection rate of unopened cotton bolls in the case of sunlight reflection,and at the same time reduce the volume and calculation amount of the model,making the model lighter and meeting the real-time monitoring needs of cotton bolls in real field environments.This work was validated on the unopened cotton boll dataset.Image acquisition for the dataset was collected from Aodu Water Control Farm in Kashgar,Xinjiang.Data annotation and data enhancement operators were performed on the original images according to specific needs.For detection and counting tasks,ablation experiments and comparison experiments are conducted in this thesis to verify the effectiveness of each part of the improved model and its advantages over other methods.