Image processing method and image processing terminal

Abstract

The present invention discloses an image processing method and terminal. The method includes: (S1) constructing a causal graph to eliminate false associations and using a generative adversarial network (GAN) to generate images retaining only the target object; (S2) extracting multi-scale features via parallel convolutional branches with an attention network predicting fusion weights, enhancing key features and suppressing noise; (S3) designing three self-supervised tasks—image reconstruction, rotation prediction, and contrastive learning—to enforce learning of cross-domain invariant features using mixed-domain data; (S4) integrating a domain classifier after the feature extractor, with a gradient inversion layer to align source and target domains. The adversarial training intensity is dynamically adjusted to balance feature alignment and task classification. This method improves cross-domain robustness and feature discriminability in image understanding applications, enhancing model generalization, accuracy, and interpretability.

Images