Convolutional Neural Network Target Detection Algorithms And Application In The Field Of Landslide

China has vast territory,complex topography,and frequent landslides.Timely and accurate detection of landslide targets is very important for reducing landslide hazards and preventing secondary landslides caused by rainfall and earthquakes.The research of landslides detection is mainly based on remote sensing image data sets.At present,the detection of landslide target is mainly based on image processing technology combined with manual analysis.Convolutional neural network(CNN)which is widely used in vehicle,pedestrian,medical images and other target detection fields,has achieved a lot of research results,but relatively few in the field of landslide target detection.Based on the analysis of various models and related parameters of CNN,an improved Faster R-CNN network model is proposed according to the characteristics of landslide target detection,and it is applied to landslide target detection with good results.Firstly,aiming at the phenomenon of white field excess in remote sensing images,the training set was pre-processed by defogging.After pretreatment,the color contrast between the target and background of remote sensing image is more distinct,and the edge feature of landslide target is more prominent,which provides an effective guarantee for the CNN to detect landslide target in remote sensing image more effectively.Secondly,in order to retain more useful feature information,an improved Faster R-CNN network model was proposed.The pooling type in the model was improved from max pooling to stochastic pooling,which avoids maximum interference and improves detection efficiency.In addition,because there are often small landslide targets in remote sensing images,in order to effectively identify and locate such targets,this paper also added a number of small-size anchors to the network model,and the increase in the number of anchors made the detection accuracy of the network model improved effectively.Based on the improvement of Faster R-CNN network model,this paper also optimized the parameters of the network model,making it more adaptive to remote sensing images.Compared with the traditional Faster R-CNN network model,the m AP of the improved network model is increased by 5.26%.Finally,the improved Faster R-CNN network model and YOLO,SSD network model were applied to remote sensing image data centralization,and the detection results werecompared and analyzed.The experimental results showed that the improved Faster R-CNN network model has better detection effect than the other three network models.