A lightweight super-resolution reconstruction method for phase contrast interferometric microscopy images

Abstract

To enhance the resolution of raw interferometric images and capture fine details of microscopic images, this paper proposes a lightweight convolutional neural network-based interference image super-resolution model (Efficient Interference Image Super-Resolution, EIISR). This model avoids the need for complex optical system modulation, enabling high-resolution quantitative reconstruction of the structure of the sample under test. The model comprises multiple deep feature extraction modules and incorporates several key design components. An adaptive feature integration module is designed to capture non-local information, while the fusion of fused mobile inverted bottleneck convolutions and channel attention mechanisms effectively aggregates local contextual features. Furthermore, the integration of LayerNorm regularization and residual connections contributes to improved image reconstruction quality. The network is trained using reconstructed phase images of $5\mu m$ micro-particles. Experimental results show that, in a 4x super-resolution reconstruction task, the model significantly enhances image detail, with a structural similarity index (SSIM) of 0.8189 and a peak signal-to-noise ratio (PSNR) of 27.43 dB. This study provides an effective reference for improving the quality of interferometric images using deep learning techniques and offers a key technological approach for integrated detection in fields such as microscopy and precision measurement.