Supplementary_1. Distinguishing PFR, self-channel and self-spatial attention modules.
This figure illustrates the distinctions between our Pixel Feature Refinement (PFR) module and the self-attention mechanisms based on spatial and channel attention modules. We represent features from the i-th and j-th channels as F_{i} and F_{j}, with F_{i, x, y} and F_{j, m, n} indicating specific coordinates within these channels. It's observed that the combination of F_{i, x, y} with F_{j, m, n} enhances the integrity of information. The self-channel attention module (a) calculates an attention metric between each channel, but fails to capture spatial correlation. In contrast, the self-spatial attention module (b) captures spatial correlation within individual channels but does not consider the entire channel set. Our PFR module (c) takes into account pixel-wise correlations across all channels, resulting in a more integrated and comprehensive structural information representation.
Supplementary_2. Visualization of heatmaps through PFR and CPFR modules.
Visualization of heatmaps through PFR and CPFR modules. (a) Input image (top) and Ground Truth (bottom). Channel visualization (b) before and (c) after our PFR module. Our proposed CPFR module uses low-level features (d) and high-level features (e) to generate features (f). The features presented in (c) and (f) demonstrates the ability of PFR and CPFR modules to capture integrity features.
Supplementary_3.
Experiments with Res2Net50 and PVT Backbones
Comparative Analysis of Integrity Among Our Method and Other SOTAs.
Integrity comparisons among various methods at false negative ratios (FNR) are essential, as FNR is a direct indicator of integrity; lower values signify better performance.
