Figure 3

ICRL framework. The causal intervention module generates augmented images by intervening on non-causal factors. Both the original and augmented image representations are fed into the decomposition module, which applies a decomposition loss to enforce separation between causal and non-causal factors. The disentangled factors are then processed by the Independence module to ensure they follow a normal distribution. Finally, the Adversarial Masking module performs an adversarial task between the generator and the masker, ensuring that the learned representations possess sufficient causal information for classification.