Aerial image matching based on dynamic convolution and adaptive neighborhood consensus network

Aerial image matching is the process of geometrically aligning two or more images under different environments (time, angle, sensor changes). Aiming at the problems of low matching accuracy and weak robustness of aerial images in matching and alignment tasks, we improve an algorithm for feature extraction combining dynamic convolution and pre-trained residual networks. In the matching stage, an adaptive domain consensus network is adopted to enhance the matching accuracy. Dynamic convolution replaces the convolution in the original Resnet101 residual block and combines it with other pre-trained convolutional layers for feature extraction, which can extract deep fine features. The bidirectional matching network is designed and the adaptive neighborhood consensus network is used to filter the mismatched point pairs, which can obtain higher matching accuracy and lower matching error. Experimental results show that the model is better than other models under multiple evaluation indicators of aerial image matching.