Context-aware feature map calibration for CNN generalization improvement

Robust object recognition remains a challenge when visual models rely on background context that co-occurs with target objects. Such dependence can severely degrade performance under distribution shifts. In this work, we examine the degree to which CNNs rely on background cues, and propose a training-time strategy to reduce such reliance. Our method combines Grad-CAM with segmentation masks to quantify the spatial focus of individual feature channels. Channels that consistently activate on background regions are selectively suppressed using an adaptive dropout mechanism, encouraging the model to prioritize object-centric representations while preserving regularization benefits. Experiments under standard and background-altered conditions show improved robustness, with accuracy rising from 82.45% to 89.92% for ResNet- 18, 81.76% to 84.80% for ShuffleNet, and 80.42% to 86.96% for VGG16BN.