Research On The Algorithm Of Building Detection In The Arable Land

Building detection in the arable land has important application value for illegal building monitoring,change detection,etc.The arable land management department usually obtains building patch information by combining image visual manual delineation and field investigation,which requires a lot of manpower and material resources.Based on the Gaofen 2 image data of Heze City,Shandong Province,this paper constructs a farmland range building detection data set,and performs data augmentation operations such as multi-scale amplification,horizontal and vertical flipping,and image mixup.In this paper,based on the characteristics of the large number of small targets and irregular shapes of some buildings in the building detection data set,a multi-scale feature fusion module is constructed based on the attention mechanism,Carafe upsampling operator and deformable convolution,which improves the accuracy of small objects.The ability to detect buildings and irregular buildings;for the characteristics of some greenhouses and buildings that have similar textures and similar shapes but densely distributed,this paper proposes a ROI Align module that takes into account the contextual features,extracts the features of the context area of the candidate frame and adds it to the feature of the candidate frame area Participate in the follow-up frame regression and classification to reduce the misdetection of the greenhouse.In view of the cross-domain difficulties between the training set and the verification set of remote sensing image data sets,this paper proposes a semi-supervised training mechanism based on image mixing.Based on the improved multi-scale feature fusion module proposed in this paper,the ROI Align module that takes into account contextual features,and the semi-supervised training mechanism based on image mixing,an improved Faster RCNN target detection network suitable for the detection of buildings on cultivated land is constructed.On the basis of the improved Faster RCNN target detection network proposed in the text,in view of the characteristics of high recall rate of building detection results in farmland supervision work,this paper proposes a post-processing mechanism based on softening and weighting NMS,for adjacent high-scoring detection frames Retain,and weight the set of candidate frames whose IoU is higher than the threshold to obtain better detection frames.At the same time,the discrete distribution of detection frames is reserved,which improves the recall rate of the overall network compared with traditional NMS and Soft-NMS.Experiments are conducted on the arable land building detection data set constructed in this paper.The improved Faster RCNN network and soft-weighted NMS post-processing mechanism proposed in this paper have achieved certain improvements in both mAP50 and AR@20 compared to Faster RCNN that indicated the effectiveness of the algorithm proposed in this paper...