Prompt Description

This is a clean, professional neural network architecture diagram of a lightweight encoder-decoder semantic segmentation model. The encoder utilizes a MobileNetV2 backbone to generate multi-scale feature maps. The decoder initiates from the lowest-resolution encoder output, applying a 1x1 convolution for channel reduction, followed by iterative bilinear upsampling. At each upsampling step, skip connections merge intermediate encoder feature maps with decoder features. The decoder exclusively employs 1x1 convolutions and bilinear upsampling, avoiding complex modules like ASPP or attention mechanisms. The final output is a full-resolution segmentation map. The diagram features a flat vector style, clear arrows, labeled blocks, and adheres to an academic paper figure aesthetic with a white background.