Construction Waste Recognition From Remote Sensing Images Based On Deep Learning

In recent years,with the acceleration of China’s urbanization,the problem of poor management of urban construction waste is becoming increasingly prominent.Most of China’s construction waste adopts the form of on-site investigation,which has poor timeliness and is difficult to achieve all-round detection.Therefore,the need of rapid detection of construction waste becomes stronger and stronger.The remote sensing image technology which possess the traits of large coverage and short detection cycle is very suitable for detecting the stacking of construction waste and realizing the dynamic monitoring of construction waste.Our research expands the data set of remote sensing images which contain construction waste image to solve the problem of lacking enough sufficient samples.Moreover,we chose one stage target detection model YOLOv3 as well as Faster R-CNN which obtain two steps network to implement construction waste detection research.The YOLOv3 network is selected and improved for the construction waste identification experiment in this treatise,and the number and size of anchor boxes are re-clustered by K-means method;Adding both 1 × 1 convolution and hole convolution to achieve multi-layer feature information integration to gain higher quality feature map.We can learn from the experimental results that the network which we have improved can not only obtain about 78.27% map but also have probably 32 fps detection speed,thus we can get more effective detection accuracy of construction waste in remote sensing images.However,the detection accuracy is still low in the detection of construction waste such as demolition waste without dust net.For the construction waste with poor detection accuracy of the improved YOLOv3,Faster R-CNN network is selected and analyzed and improved: dense connection feature pyramid network is used to extract deeper semantic information;Attention mechanism is introduced to enhance the network’s ability to express multi-scale features;in the regression stage,GVR mechanism is introduced to strengthen the border position learning ability for targets in different directions.The experimental results show that the improved methods selected on Faster RCNN network in this treatise,to some extent,can improve the accuracy of construction waste detection result in remote sensing images.And obtain 83.06%map.In the one stage network model,the types of construction waste with low detection effect,such as house demolition waste,no dust screen and waste disposal yard,have increased by 18.54% and 16.37% in the two stage model,but the detection speed is only 14 fps due to the complexity of the network structure.The two models can be combined to achieve rapid and accurate detection of construction waste.The improved YOLOv3 network can be used to efficiently detect the house demolition waste and construction waste classification points with dust screens,and the improved Faster R-CNN can be used to more accurately detect the house demolition waste,waste disposal yard and suspected stacking points without dust screens.Combine the two models to achieve rapid and accurate detection of construction waste.