文献类型：期刊文献

英文题名：Digital image super-resolution reconstruction method based on stochastic gradient descent algorithm

作者：Yu, Yinghuai[1];Peng, Xiaohong[1];Ye, Xiaoxia[1]

机构：[1]Guangdong Ocean Univ, Coll Math & Comp Sci, Zhanjiang 524088, Guangdong, Peoples R China

年份：2025

卷号：31

外文期刊名：EGYPTIAN INFORMATICS JOURNAL

收录：SCI-EXPANDED(收录号：WOS:001566869700002)、、EI(收录号：20253619112475)、Scopus(收录号：2-s2.0-105014920947)、WOS

基金：1. This work was supported by Research and Application Demon-stration of Key Technologies for the Digitalization of Marine Ranching [grant number: No. 2024R1003] . 2. This work was supported by Guangdong Intelligence Platform of Prawn Modern Seed Industry [grant number: No. 2022GCZX001] . 3. This work was supported by Research and application demon-stration of key technologies for intelligent prawn breeding [grant number: No. 2023ZDZX4012] .

语种：英文

外文关键词：Digital image super-resolution; Stochastic gradient descent; Image reconstruction; Hybrid loss function; Structural similarity index measure

外文摘要：Digital image super-resolution (SR) techniques have gained significant attention in computational imaging for reconstructing high-quality images from low-resolution inputs. Traditional SR methods often struggle with preserving fine details, texture consistency, and edge sharpness while maintaining computational efficiency, limiting their practical applications in real-time systems. The research proposes an Adaptive Dynamic Efficient Parameter Tuning for Super-Resolution (ADEPT-SR) framework based on an optimized stochastic gradient descent (SGD) algorithm. The technique transforms low-resolution (LR) images into high-resolution (HR) counterparts, addressing fundamental limitations in imaging hardware. ADEPT-SR implements an adaptive SGD framework with momentum-based parameter optimization to minimize reconstruction error between predicted and ground-truth HR images. The key innovation in ADEPT-SR lies in a hybrid loss function combining structural similarity index measure (SSIM) and perceptual loss with dynamic weighting that adjusts during training iterations. The approach enables superior edge preservation and texture reconstruction compared to conventional methods. An adaptive learning rate schedule dynamically responds to local optimization landscapes, reducing convergence time by 37 % while avoiding local minima. ADEPT-SR offers significant applications in medical imaging, satellite imagery analysis, surveillance systems, and consumer electronics, where hardware limitations constrain native resolution. Experimental validation across standard benchmark datasets demonstrates that ADEPT-SR achieves a peak signal-to-noise ratio (PSNR) improvement of 1.8 dB over standard bicubic interpolation and 0.7 dB over recent deep learning approaches for a 4 x upscaling factor. The method reduces computational complexity by 43 % compared to deep learning methods while maintaining visual quality improvement.