A Light-Weight Neural Network Using Multiscale Hybrid Attention for Building Change Detection

Abstract

The study of high-precision building change detection is essential for the sustainable development of land resources. However, remote sensing imaging illumination variation and alignment errors have a large impact on the accuracy of building change detection. A novel lightweight Siamese neural network building change detection model is proposed for the error detection problem caused by non-real changes in high-resolution remote sensing images. The lightweight feature extraction module in the model acquires local contextual information at different scales, allowing it to fully learn local and global features. The hybrid attention module consisting of the channel and spatial attention can make full use of the rich spatiotemporal semantic information around the building to achieve accurate extraction of changing buildings. For the problems of large span of changing building scales, which easily lead to rough extraction of building edge details and missed detection of small-scale buildings, the multi-scale concept is introduced to divide the extracted feature maps into multiple sub-regions and introduce the hybrid attention module separately, and finally, the output features of different scales are weighted and fused to enhance the edge detail extraction capability. The model was experimented on the WHU-CD and LEVIR-CD public data sets and achieved F1 scores of 87.8% and 88.1%, respectively, which have higher change detection accuracy than the six comparison models, and only cost 9.15 G MACs and 3.20 M parameters. The results show that our model can achieve higher accuracy while significantly reducing the number of model parameters.

A Vehicle Detection Method Based on an Improved U-YOLO Network for High-Resolution Remote-Sensing Images

“…where Precision is the percentage of positive samples among the target detection samples, Recall is the percentage of correctly identified targets among all targets, AP is the area of the P-R curve, and mAP is used to represent the average of the mean accuracy of multiple categories in the dataset. It is calculated as shown in Equation (10).…”

“…Zou et al [9] improved the detection accuracy of traffic signage via a bidirectional feature pyramid to extract the context information in the feature layer. Hua et al [10] improved the ability of the model to recognize fine details by constructing a multi-scale hybrid attention module to aggregate contextual information of the input image. Yadav et al [11] designed a multi-scale feature fusion module by analyzing three characteristics of grayscale images, which makes full use of the contextual information of the features and improves detection accuracy.…”