A Lightweight High-Resolution Network-Based Method for Metal Product Quality Inspection
HAN Junjia
DUAN Chengpu
HUANG Jingtao
Abstract：Traditional manual inspection suffers from issues such as inconsistent standards and susceptibility to fatigue interference.However,existing deep learning methods based on real-time object detection algorithms tend to lose feature details due to downsampling operations and exhibit high computational complexity,making it difficult to meet precise detection requirements.To address these challenges,this paper proposes a lightweight high-resolution network detection method for metal surface defects.This method preserves subtle defect signals through a full-process high-resolution feature preservation architecture,achieves a lightweight design by integrating depthwise separable convolutions,introduces a conditional channel weighting mechanism to optimize multi-resolution feature fusion,and proposes a global spatial feature extraction method to enhance contextual correlation.Experiments show that the network achieves mAP50 scores of 79.6％and 70.9％on the NEU-DET and GC10-DET datasets,respectively,with a parameter count of only 5.3M and 7.1 GFLOPs.When tested on a dual RTX 4090 GPU setup,it reaches a frame rate of 117 FPS.Compared with previous benchmark models,the proposed method demonstrates significant advantages in both detection accuracy and computational efficiency.Ablation experiments verify the effectiveness of each module,providing a high-precision and low-power solution for real-time industrial detection of metal surface defects.
Keywords：intelligent manufacturingpattern recognitionmetal surface defectsurface quality inspectionlightweight model
Publication Date：2025-10-25
Online Publishing Date：2025-11-17(First online date of this platform, not the publication date of the document)
Pages：11( 40-50 )