Real-Time Industrial Surface Defect Detection Based on Lightweight Deep Learning Network

July 31, 2026

Industrial surface defect detection often requires both accurate localization and high inference speed, especially in real-time inspection systems.This thesis focuses on lightweight detection for small-scale industrial defects and weak-texture defects.These defects are often difficult to distinguish from complex backgrounds in industrial production ,which may lead to missed detections.Based on RTMDet-Tiny,this thesis designs a lightweight detection model with detail enhancement.The aim is to improve the model’s ability to represent the features of small-scale industrial defects and weak-texture defects.The framework combines three main components.First,P3 Detail Shortcut strengthens shallow high-resolution features before they are fused into the P3 layer,so that local texture and boundary cues can be better retained.Second,Head Lite simplifies the detection head to reduce redundant computation in the main detector.Finally,a Sparse Local Refiner is used for selected weak texture candidates,providing local verification and conservative box adjustment when the main detector gives uncertain predictions.Experiments are mainly conducted on NEU-DET,with additional validation on DeepPCB and MagneticTile.Compared with the controlled RTMDet-Tiny +CROP(0 .75) baseline ,the final pipeline improves mAP@0 .5 from 0.721 to 0.8107.In the controlled ablation route, the final main detector reduces parameters from 4.875M to 4.727M and FLOPs from 8.029G to 6.760G.The cross-dataset results further indicate that the proposed modules show transferability to other industrial defect scenarios,although the improvements remain category-dependent.Overall,the proposed framework provides a practical accuracy-complexity trade-off for lightweight industrial surface defect detection.

Autores da comunidade :

• 

  Z

  Zhenyu Ge

  ist1115690

Orientadores desta instituição:

• 

  A

  António José Castelo Branco Rodrigues

  ist12288

electrical-engineering-electronic-engineering-information-engineering - Engenharia Eletrotécnica, Eletrónica e Informática

por - Português

Instituto Superior Técnico