Figure 2.
D M Net is enclosed within a dashed boundary and is divided into theta D and theta E sections. The theta D section functions as an encoder and contains sequential depthwise separable convolution 2 d layers with dimensions 36 by 64, 64 by 128, 128 by 256, 256 by 512, 512 by 1024, 1024 by 512, 512 by 256, 256 by 128, and 128 by 64. The theta E section functions as a decoder and contains alternating depthwise separable convolution 2 d and convolution transpose 2 d layers with dimensions 64 by 128, 128 by 64, 64 by 256, 256 by 128, 128 by 512, 512 by 256, 256 by 256, 256 by 128, 128 by 64, and 64 by 36. A skip connection links the input to a plus operation at the output, where features are summed to produce the final result.The detailed architecture of the D-MNet module. The module is composed of a sub-encoder () and a sub-decoder (). The sub-encoder utilizes a sequence of depthwise separable convolution blocks to create a multiscale feature representation, progressively expanding the channel dimension from 36- to a 1024-dimensional bottleneck. The sub-decoder then uses a combination of depthwise separable convolutions and convolution transpose layers to reconstruct the features. A global skip connection sums the module’s output with its input, enabling the network to learn the residual mapping