Dead Tree Information Detection Based On Convolution Neural Network

Assessing the health status of trees in forests has important implications for biodiversity,forest management and environmental monitoring.Dead trees are a key indicator for assessing forest biodiversity and ecosystem health,so dead tree detection is essential.At present,the detection of dead trees is relatively primitive,relying mainly on manual inspection and observation of forest areas,but manual detection methods can be very labour-intensive and difficult to detect,especially in areas with treacherous terrain,so the traditional manual detection methods are far from adequate to meet the increasingly serious needs of dead tree detection.This thesis addresses the current problems of inefficient and labour-intensive dead tree detection by constructing a dead tree detection method based on convolutional neural networks and unmanned aerial vehicle images,thus providing data support and solution ideas for the automated detection of dead trees.The main research content of the thesis is as follows.1.Create a dataset of images of dead trees.Due to the relative lack of publicly available datasets of dead trees,this experiment went to the field for data collection.In this thesis,an unmanned aerial vehicle was used to collect images in a mountainous forest in a scenic area of southern Liaoning.The images were then screened and the screened images were data enhanced,proposing an improved log-transformed image enhancement method that allows for more flexibility in adjusting the images to select their desired effect by adjusting the parameter V in the improved formula.The data enhanced images are annotated using Label Img to generate a dataset of 10,000 images of dead trees.2.Design of a dead tree detection model based on an improved YOLOv4-tiny model.A dead tree detection method with an improved YOLOv4-tiny model is proposed to address the problems of simple network structure and poor robustness of current deep learning-based data detection models.This method first adds the attention mechanism ECANet to the model so that the network model can automatically highlight the important information in the image and improve the network performance,and changes the Leaky Re LU to ELU activation function in the model,which makes the model more robust to noise during training and improves the model training efficiency.Finally,the SPP structure is added after the Backbone part of the model to fuse local features with global features for the purpose of enriching the feature representation capability.After experimental comparison,the AP value of the improved model reached 93.36%,an improvement of 9.69% compared to the YOLOv4-tiny model,and a better detection result was obtained.3.Design of a dead tree detection model based on an improved YOLOv4 model.In order to further improve the accuracy of the dead tree detection model,a dead tree detection method with an improved YOLOv4 model is proposed.This method modifies the CSPDark Net-53 backbone network in YOLOv4 to a lightweight Mobile Net V3 network and applies deep separable convolution to the enhanced feature extraction network to reduce the number of model parameters while maintaining model performance.In addition,the use of the K-Means++ clustering method to obtain a priori frames that are more suitable for this experiment and have higher accuracy makes the model easier to learn.After experimental comparison,the AP value of the improved model reached 97.33%,which improved the accuracy of the improved YOLOV4-TINY model by 3.97%,resulting in better detection results.