With the widespread deployment of mobile imaging sensors and smart devices, the risk of image privacy leakage is increasing daily. Protecting sensitive information in captured images has become increasingly critical. Existing image privacy protection measures usually rely on manual blurring and occlusion, which are inefficient, prone to omitting privacy information, and have an irreversible impact on the usability and quality of images. To address these challenges, this paper proposes TSLEPS (Two-Stage Localization and Erasure method for Privacy protection in Sensor-captured images). TSLEPS adopts a two-stage framework comprising a privacy target detection sub-model and a privacy text erasure sub-model. This method can accurately locate and erase the private text areas in images while maintaining the visual integrity of the images. In the stage of detecting privacy targets, an inverted residual attention mechanism is designed and combined with a generalized efficient aggregation layer network, significantly improving privacy target detection accuracy. In the stage of privacy text erasure, a texture-enhanced feature attention mechanism is proposed with an adversarial generative network for the erasure task to achieve efficient erasure of privacy texts. Moreover, we introduce the half-instance normalization block to reduce the computational load and inference time so that it can be deployed on resource-constrained mobile devices. Extensive experiments on multiple public real-world privacy datasets demonstrate outstanding performance, with privacy target detection achieving 97.5% accuracy and 96.4% recall, while privacy text erasure reaches 38.2140 dB PSNR and 0.9607 SSIM. TSLEPS not only effectively solves the privacy protection challenges in sensor-captured images through its two-stage framework, but also achieves breakthrough improvements in detection accuracy, erasure quality, and computational efficiency for resource-constrained devices.
Li et al. (Wed,) studied this question.