SEMANTIC SEGMENTATION OF REMOTE SENSING IMAGES BASED ON MIXED DEEP CONVOLUTION

As an important part of remote sensing interpretation, semantic segmentation of high-resolution remote sensing images contains a large amount of complex feature information of ground objects, and the size of different ground objects is quite different, which brings some difficulties to semantic segmentation of remote sensing images. To solve this problem, a remote sensing image semantic segmentation model MDU-NET based on hybrid deep convolution is designed and implemented. In this model, a parallel feature extraction module was used in the encoding stage, and the dynamic model topology was realized by dynamically assigning weights to different branches in the module. At the same time, a channel and spatial attention model were introduced to improve the feature fusion effect from encoder to decoder and improve the semantic segmentation effect. The test set accuracy on the ISPRS Validation dataset was 3. 44 Percentage points higher than DeepLabv3 +. Experimental results show that the proposed network achieves good segmentation results in high resolution remote sensing image segmentation.