Detailed architecture of the feature extraction backbone. The network consists of four hierarchical stages. An initial feature extraction block maps the input from 3 to 64 channels and integrates a wavelet transformer and a PSF-guided pinwheel convolution (PSFConv) to preserve multi-scale frequency information and PSF-aligned morphology. Each stage is built from residual bottleneck blocks with CBAM attention for adaptive feature refinement. The PSFConv module applies four directional convolutions (two horizontal and two vertical) followed by fusion to model the cross-shaped PSF structure of Lobster-Eye optics. The wavelet transformer decomposes features into four subbands, processes them via grouped convolution, and reconstructs the spatial representation via inverse wavelet transform, enabling frequency-aware feature enhancement before entering deeper layers.