<p>Existing single-image deraining methods still struggle with the complex effects of raindrops on transparent surfaces like glass, camera lenses, and other optical components. This paper proposes a novel single-image deraining algorithm using a multistage recursive network to effectively separate the raindrop layer from the background layer on transparent surfaces. The proposed network includes multiple basic layers, each incorporating a revised layer normalization module to standardize multi-scale features across different data distributions, enhancing generalization and adaptability to real-world scenarios. GCE-Net includes an adaptive feature fusion module (SK Fusion) to integrate features from different levels and suppress redundancy, while the multilayer perceptron module enhances nonlinear feature representation. A gating mechanism is introduced to enhance flexibility, and combined with a recursive structure, it captures dependencies among deep features, significantly improving deraining performance. We trained the model on the Raindrop dataset and tested it on both the Raindrop and SPA-Data datasets. The raindrop dataset targets raindrops on transparent surfaces with diverse shapes and complex scenes. GCE-Net shows excellent performance in PSNR (34.14) and SSIM (0.969) metrics, effectively restoring background information while minimizing raindrop impact on image quality. The proposed algorithm provides an efficient and robust baseline for image deraining on transparent surfaces, demonstrating potential for practical applications in real-time video surveillance systems and large-scale image processing pipelines that require HPC infrastructure to handle high-throughput data streams. Leveraging depthwise separable convolutions, the model achieves a runtime of 4.389 ms on high-end GPUs, balancing deraining quality (PSNR 34.14) with real-time efficiency, making it suitable for supercomputing-accelerated edge-cloud collaborative frameworks.</p>