A lightweight adaptive image deblurring framework using dynamic convolutional neural networks

Abstract

Image deblurring remains a fundamental challenge in computer vision, particularly for Lightweight models facing Limited input adaptability and inadequate global context modeling. This paper proposes a Lightweight adaptive image deblurring framework based on dynamic convolutional neural networks, featuring three core modules to enhance adaptability, global context modeling, and multi-scale feature fusion: 1) The Shallow Adaptive Feature Module (SAFM) employs dynamic convolution to adjust kernel weights according to input characteristics, improving adaptability to diverse blur patterns; 2) The Attention Feature Conditioning Module (AFCM) incorporates a Simple Spatial Attention (SSA) mechanism, which captures global context via 1D spatial pooling while preserving spatial location information, enhancing the model’s capability to model long-range dependencies; 3) The Multi-Scale Attention Fusion (MAF) module dynamically weights cross-level features using global attention, enabling efficient hierarchical feature aggregation. Experiments show that the proposed framework achieves competitive PSNR and SSIM performance compared to other lightweight models on the GoPro and HIDE datasets, while maintaining relatively low computational complexity, thus offering a practical solution for intelligent applications.