Pest Region Segmentation Of Aerial Photography Image In Forest Region Based On Deep Convolution Neural Networks

The traditional image segmentation methods applied in pest region segmentation in aerial images of forest areas feature low accuracy and speed,centering on which the thesis puts forth a set of technical schemes for image segmentation and evaluation of forest pests,including the data set of pest region segmentation,the high-precision pest image segmentation algorithm suitable for the ground end which is beneficial to improve the segmentation accuracy to the greatest extent,the real-time pest segmentation algorithm for UAV with the aim to improve the speed of frontal model segmentation while guaranteeing low accuracy loss,and the operation platform of pest region segmentation realizing the practicability of the algorithm.The detail is as follows:1.The data set of pest region segmentation in UAV aerial image is established.Drawing on the 8-Rotor UAV pest image acquisition platform,167 images of forest pests have been collected and 800 images of 1000X 1000 pixels have been obtained by clipping.Besides,all images are labeled at the pixel level in order to provide reliable materials for the segmentation algorithm research.2.A high precision pest region segmentation algorithm based on fully convolutional networks is proposed.Aiming at the poor generalization ability of traditional recognition methods for aerial forest pest images and the irregularity of pest areas,the proposed algorithm enhances model generalization ability via transfer learning,strengthens model accuracy by virtue of jumping structure and puts forward five full convolution networks.Experiments show that for forest pest images,FCN-2s has the highest recognition accuracy,with a mean intersection over union of 79.49%,a pixel accuracy of 97.86%,and a single segmentation time of 4.31s.Compared with K-means,pulse coupled neural network and composite gradient watershed algorithm,the pixel accuracy of this algorithm is respectively 44.9,20.73 and 6.04 percentage points higher,and the segmentation time of a single image is reduced by 47.54s,19.70s and 11.39s respectively.3.Mobile-BiSeNet,a real-time pest image segmentation model for UAV terminal operation is proposed.In the thesis,targeting FCN and other high-precision pest image segmentation models that contain such problems as the large parameters,the slow segmentation speed and the difficulties in deploying to UAV platform,the backbone model of BiSeNet context path is replaced by Mobilenet V2,attention refinement module and feature fusion module are designed,and the Mobile-BiSeNet segmentation algorithm is constructed,so that the parameters of the model are reduced and the operation speed is increased.Experiments show that Mobile-BiSeNet only has 1/89 of the parameters of FCN-2s,the segmentation time of the model is increased by 50.95 times,the segmentation time of a single image only needs 0.0846s,while the pixel accuracy is only lost by 0.96%,and the mean intersection over union is lost by 8.98%.This algorithm can realize fast pest segmentation based on UAV platform.4.A multi-functional UAV terminal platform is built,which integrates image acquisition,remote transmission,image segmentation and evaluation,and realizes the practicality of pest image segmentation algorithm.