Intelligent Coarse Dropout and Anti-ICD: Saliency-Guided Masking Augmentation for Visual Classifiers

Visual classifiers trained with standard augmentation pipelines tend to exploit spurious correlations:
background statistics, texture patterns, or scene-level cues that happen to co-occur with class labels
but do not reflect object structure. Conventional augmentation strategies, including random crops,
colour jitter, and policy-based methods such as RandAugment and AutoAugment, are constructed
without reference to what the model has actually learned, and so offer no mechanism for targeting the
specific shortcuts a given classifier has acquired.
We present Intelligent Coarse Dropout (ICD) and its complement Anti-ICD (AICD), two data
augmentation methods that condition spatial masking on the model’s own class activation maps.
Standard coarse dropout (cutout) hides rectangular regions chosen at random, erasing background as
readily as object. ICD and AICD make this saliency-aware: they split the image into a grid of tiles,
score each tile by its mean saliency, and selectively mask tiles by a percentile threshold. ICD masks
the highest-saliency tiles, the regions the model already relies on, so that the classifier must recruit
additional cues. AICD masks the lowest-saliency tiles, perturbing the background while leaving the
discriminative region intact. Masked tiles are not filled with a hard black box but with one of several
soft strategies (blurred original, local mean, noise, or a constant), which preserve different aspects of
local context.
We describe the formal construction of both masks, the family of fill strategies, the role of the
tile size and threshold hyperparameters, and a reference implementation in the open-source BNNR
library. This paper introduces the methods; a quantitative evaluation against standard augmentation
baselines is left to a companion study.