Research On Illegal Building Detection Technology Based On Deep Learning

The phenomenon of illegal construction has seriously affected the sustainable economic development and social stability of our country.Aiming at the problem of low efficiency of manual inspection,this paper is devoted to researching the automatic detection technology of illegal buildings based on deep learning in fixed-point monitoring scenarios.The main research contents of the paper are as follows:(1)A FDA-Deep Lab semantic segmentation algorithm based on the fusion of dual attention mechanism is proposed.Aiming at the problems that Deep Labv3+ is easy to misjudge similar objects,easy to miss small targets,and the prediction output has holes,a feature fusion module combined with dual attention mechanism is designed,and the feature maps are downsampled by 4 rates,8 rates,and 16 rates respectively.This module is used to fuse low-level features to make up for the lack of high-level features.In view of the unbalanced problem of training samples,the loss function is improved by introducing the sample difficulty weight adjustment factor and class weight,and the segmentation accuracy is improved.Experiments show that the algorithm effectively improves the segmentation performance of the model.Under the fixed-point monitoring perspective,the building segmentation achieves 92.8% MIo U.(2)A CA-YOLO target detection algorithm based on hybrid atrous convolution is studied.In view of the low detection accuracy of YOLOv4 in a mixed natural environment,the channel attention is decomposed into two one-dimensional features aggregated in different directions and introduced into the model to enhance the target representation of interest;the feature pyramid structure is added before and after the convolution structure to improve Deep feature extraction ability;using parallel hole convolution to enhance the ability to obtain location information;applying class smooth labels to classification loss to reduce the penalty of negative samples and improve network generalization ability.Experiments show that the algorithm effectively improves the detection performance of the model.From the perspective of fixed-point monitoring,the construction site detection of construction projects reaches 73.53 m AP%.(3)Developed a set of automatic monitoring system for illegal buildings under fixed-point monitoring.Firstly,the color migration of the old and new phase images is carried out to eliminate the local color difference caused by the weather change.Then,the algorithm proposed in Chapter 4 is used to detect and mark the construction site of the construction project,and the algorithm proposed in Chapter 3 is used to segment the building area and keep it;In order to extract the changed area in the building area,the improved pixel difference method and gradient difference method are combined to extract the changed area and mark it.Finally,in view of the lack of fixed-point monitoring at this stage,the corresponding illegal building automatic monitoring system to assist law enforcement problem,in Windows Under the platform,an automatic monitoring system for illegal buildings under fixed-point monitoring is developed.