Research On Seedling Rape Plant Counting Algorithm Based On Improved Neural Network

Rapeseed is one of the main oil crops in China,which plays a crucial role in ensuring the security of the domestic supply of edible vegetable oil.However,there exists a problem of poor plant uniformity and unbalanced development of the population in rapeseed cultivation,particularly in the mechanized drilling and broadcasting planting methods.To evaluate the plant uniformity of rapeseed,it is essential to accurately count the number of rapeseed plants.In this regard,low-altitude remote sensing technology was employed to capture the images of the rapeseed plants during the seedling stage.The rapeseed plant counting was achieved through computer vision technology.However,the accuracy of plant counting is greatly challenged due to the unfavorable factors such as dense occlusion in the captured images.In response to this issue,this thesis conducted research on the precise counting of densely distributed rapeseed plants using various deep learning algorithms.The main research contents are as follows:(1)Aerial drone images were collected to capture the images of rapeseed plants during the seedling stage in both the mechanized drilling and broadcasting planting methods.Two rapeseed plant datasets were created,one for object detection and the other for density estimation,each consisting of 600 images and 48,600 data samples.These datasets provided data support for the subsequent research on rapeseed plant counting algorithms.(2)An improved rapeseed plant counting algorithm based on YOLOv5 was proposed.KMeans++ clustering was used to generate anchor parameters that are more suitable for the rapeseed plant dataset in this study.SE and GAM attention modules were employed to enhance the focus on the feature regions of rapeseed plants.The Focal EIOU loss function was used to achieve more precise bounding box regression and reduce false detections caused by excessive attention after the addition of attention modules.The improved rapeseed plant counting model achieved accuracy(P),recall(R),mean average precision(MAP),and coefficient of determination(R2)of 84.5%,79.4%,85.3%,and 0.93116,respectively.Compared with the original YOLOv5 algorithm,the improved algorithm improved the P,R,and MAP by 2.1%,2.3%,and 2.4%,respectively,and the R2 was closer to 1,and also superior to other mainstream object detection models.(3)Although the improved YOLOv5 counting algorithm has achieved higher accuracy in counting rapeseed plants,its detection accuracy still cannot meet the requirements for precise statistical needs.Therefore,this thesis proposes a BCNet rapeseed plant counting algorithm based on the improved dense crowd counting algorithm.Since the original crowd counting algorithm considers the human head as a landmark target,the head features differ significantly from other body parts,whereas the rapeseed plant landmark target region has similar color and irregular shape.Thus,directly applying the original algorithm will result in poor counting accuracy.In this thesis,the first 13 layers of the VGG16 network are used as the front-end network,and dilated convolution is used as the back-end network.Spatial attention modules and channel attention modules are used to enhance the feature information of rapeseed plant images from both the channel and spatial dimensions,improving the expression ability of the entire feature map.Finally,a 1x1 convolution layer is used for further feature extraction,and the absolute number output is used to constrain the distribution of model parameters and reduce counting errors.The mean absolute error(MAE)and mean squared error(MSE)of the improved model reached 3.39 and 5.01,respectively.Compared with the best-performing Bayesian density estimation algorithm,the BCNet MAE decreased by 8.35% and the MSE decreased by 4.03%.This provides a new method for accurate counting of dense rapeseed plants.