This paper presents a block-wise semi-fragile watermarking algorithm for image content authentication, with tamper region localization and recovery capability. A halftone image is generated by the error diffusion halftoning method and embedded using the Quantization Index Modulation (QIM) method in the Discrete Cosine Transform (DCT) domain of the original image. The proposed method is robust to JPEG compression, because the halftone image is embedded as a watermark sequence in the middle frequencies of the DCT coefficients using QIM. Also to improve the recovered image quality, Multilayer Perceptron neural network (MLP) is used in inverse halftoning process. Data in the tampered region is estimated with gray-scale data obtained from the MLP, using the embedded halftone as input. The experimental results show desirable performance of the proposed algorithm, such as watermark imperceptibility, robustness to JPEG compression, detection accuracy of tampered region and high quality of recovered image.