Cascade network with detection and inpainting for low-quality phase images

Because of discontinuous surface or environmental noises, phase images from real applications in digital speckle pattern interferometry are usually damaged with isolated stains, or missing fringe boundaries. These anomalies will seriously destroy the subsequent phase unwrapping. To improve the robustness, efficiency and accuracy of phase analysis for the low-quality phase images, a cascade network combining anomaly detection and inpainting over these abnormal areas is proposed. Firstly, a U-Net anomaly detection network (LIMT-U-Net) is built with long-distance attention, island attention mechanisms, and multi-scale feature fusion. This network is capable of both accurately identifying diverse defect morphologies and producing pixel-level masks. Secondly, a revised Generative Adversarial Network (GLM-GAN) is designed for inpainting. It takes the detection masks as spatial constraints and employs dilated convolutions of different rates to extract fine noisy textures of local defects and global phase fringe features. Through adversarial training and perceptual loss functions, the discriminator evaluates the authenticity of inpainting results, which in turn provides refined optimization gradients to guide the generator's optimization. The proposed cascade framework tightly integrates the detection and inpainting modules, ensuring accurate anomaly localization while maintaining correctness and visual consistency with the original images. Experimental results show that this cascade network exhibits significant advantages in both anomaly detection accuracy and inpainting quality.