Building Recognition of Aerial Images Based on Improved Unet Network

Due to its simple and efficient design, the Unet network has demonstrated outstanding performance in the field of building identification. Compared to other networks, the Unet network has higher flexibility and detailing capabilities when dealing with buildings with complex contours and irregular shapes. In this paper, a novel Unet structure is proposed to improve the accuracy of building recognition. An attention mechanism is added to the lateral skip connections of the Unet network, which combines different levels of information to enhance the signals in the region of interest.The original convolutional layer module in the feature extraction part is re-parameterised to a symmetric convolutional structure, so that the model obtains better feature expression. And a dilated convolutional module of different sizes is added at the end of the feature extraction to enlarge the receptive field. The experimental results show that the proposed method is effective in segmenting building images. The pixel segmentation accuracy reaches 94.5% and the intersection of union is 85.71%.Both of them are better than the traditional Unet network.