TaDP-Det: Semi-Supervised Texture-Aware Dynamic Pseudo-Labeling Detector for Industrial Surface Defect Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Industrial Quality Control

Background:

• Surface defect detection is crucial for industrial quality control.
• Acquiring labeled data for defect detection is expensive and time-consuming.
• Semi-supervised object detection (SSOD) uses unlabeled data to reduce labeling costs, but faces challenges with industrial imagery like ambiguous textures and class imbalance.

Purpose of the Study:

• To develop an effective semi-supervised object detection method for industrial surface defect detection.
• To improve the quality of pseudo-labels generated from unlabeled industrial images.
• To address challenges like foreground-background confusion and class-dependent detection difficulty.

Main Methods:

• Proposed TaDP-Det, a semi-supervised detector with dual enhancements for pseudo-label quality.
• Introduced a Texture Enhance Module (TEM) to amplify texture cues in shallow backbone stages.
• Implemented a class-wise dynamic pseudo-label filtering (CDPF) scheme using Gaussian mixture models for adaptive thresholding.

Main Results:

• TaDP-Det demonstrated superior performance over state-of-the-art SSOD baselines on NEU-DET, GC10-DET, and PCB-DEFECT datasets.
• Achieved consistent improvements in mean average precision (mAP) for defect detection.
• Showcased effectiveness with only modest computational overhead.

Conclusions:

• TaDP-Det significantly enhances pseudo-label quality through improved feature representation and adaptive filtering.
• The method is effective for robust semi-supervised defect detection in industrial applications.
• TaDP-Det offers a practical solution for reducing the reliance on expert annotations in quality control.