Insulator Semantic Segmentation in Aerial Images Based on Multiscale Feature Fusion

Figure 2

Overview of proposed ED-Net architecture (a). It consists of two parts: Encoder path and Decoder path. The feature maps of C1, C2, C3, C4, and C5 are obtained by the Initial Module (IM) and Bottleneck + CA Module. M2, M3, M4, and M5 are the feature maps of the encoder path obtained through Asymmetric Convolution Module (ASM) or Refinement Boundary Module (RBM). P1–P5 are obtained by the Attention Feature Fusion Module (AFFM), and P1 is used for insulator segmentation. The details of ASM, RBM, and AFFM are illustrated in (b), (c), and (d), respectively. The red and black lines represent the upsample and downsample operations, respectively. The green line does not change the size of the feature map, only the number of channels.