Order Matters Beyond Accuracy: Paired Prediction Analysis of Compound Image Corruptions

Classification accuracy under compound corruption depends on the order in which degradations are applied. We evaluate all six permutations of Gaussian noise (N), Gaussian blur (B), and JPEG compression on 3,925 Imagenette-v2 validation images using four pretrained classifiers: ResNet-50, ConvNeXt-Tiny, ViT-Base/16, and Swin-Tiny. Under common nominal settings, the model-specific six-order accuracy ranges span 17.13 to 21.55 percentage points. Two orders that both end in JPEG differ by about 16 percentage points in across-model mean accuracy. A second parameter regime selected toward shared PSNR and SSIM targets reduces the range of the six across-model mean accuracies from 18.43 to 6.79 percentage points, a 63.18% descriptive contraction. A paired five-state analysis shows that predicted-label churn is 40.78% under nominal settings and 34.74% under the PSNR/SSIM-targeted regime, compared with correctness-state churn of 16.93% and 12.34%. Many of the additional label changes occur between two incorrect classes, changes that an accuracy-only analysis cannot show.