Task-Driven learned image compression with explainability preservation for image classification

Learned image compression for image classification tasks seeks an effective balance between bitrate and classification accuracy; for applications serving both machines and humans, it must also maintain reconstruction quality. As demands for trustworthy artificial intelligence increase, compressed representations should further preserve the visual evidence that supports model decisions. In this paper, we first show that standard rate-distortion (R-D) objectives in learned image compression can significantly degrade explanation fidelity: class activation maps drift toward irrelevant regions–especially at low bitrates–even when predictions remain correct, potentially undermining trust in downstream deployment. To address this issue, we introduce an explainability-aware R-D regularization that jointly preserves information critical for classification performance, reconstruction quality, and explainability. Focusing on Class Activation Map (CAM)-based explanations, our method incorporates knowledge distillation to retain explanation-relevant cues during training. Experiments demonstrate that the proposed approach delivers consistent R-D improvements while substantially enhancing the fidelity and localization of visual explanations.