Feature fusion method for remote sensing image change detection

The invention discloses a feature fusion method for remote sensing image change detection. The method comprises the following steps: step 1, carrying out regularization processing on four different features; step 2, convolution dimension reduction is carried out on the top layer features; 3, giving an expansion factor, and superposing a corresponding dilated convolutional layer; 4, the top layer features pass through a cavity pooling layer, and the features of all channels are extracted; step 5, carrying out superposition and convolution on the results of the cavity convolution and the cavity pooling; and step 6, carrying out top-down fusion on the four different scale features. Step 7, performing transverse connection on the four fused features; and step 8, fusing the transversely connected features to form an output feature map. According to the method, the information loss is reduced, the calculation amount is reduced, the receptive field is increased, the perception ability of the model to buildings with