Existing hash-based image tampering detection algorithms primarily rely on traditionally hand-designed features, which results in forensic hashes lacking detailed information. This leads to difficulties in achieving object-level tampering localization and handling various complex tampering types. To address this issue, a hash learning-based image tampering detection model is proposed. The model consists of two main modules: the hash construction module and the tampering localization module. In the hash construction module, a deep learning-based multi-scale feature extraction and fusion module is proposed. This module not only integrates image features from different scales but also constructs compact, information-dense image hashes. In the tampering localization module, differences in hash values between the original and tampered images are compared, resulting in a rough localization of tampering regions. To further improve localization accuracy, a layer-by-layer fusion strategy is employed to decode hash differences and multi-scale features. This strategy effectively combines global information with local details, enhancing the model’s robustness and adaptability to various complex tampering scenarios. Experiments conducted on three datasets and compared with nine state-of-the-art methods show that, on the CASIAv1 dataset, the proposed method achieves a 10.7% improvement in F1 score compared to the second-best model; on the Columbia dataset, the F1 score increases by 1%; and on the COVERAGE dataset, the F1 score increases by 17%. Ablation experiments on the CASIAv1 dataset validate the effectiveness of the proposed modules and their contributions to tampering detection performance, showing that each module significantly enhances overall detection accuracy. To verify the model’s robustness, the proposed method is evaluated under JPEG compression, Gaussian blur, and Gaussian noise on the CASIAv1 dataset. The experimental results demonstrate that the method maintains excellent robustness under various disturbances. The results indicate that the proposed deep learning and hash-based image tampering localization model outperforms existing tampering detection methods and shows strong generalization capability across three public datasets.