Work

Deep texture exploration and multi-scale interaction network for surface defect detection

Year: 2025

Type: article

Abstract: Although texture features inherently contain rich contextual and structural information, they have rarely been explicitly considered in mainstream object detection tasks for surface defect identification. In industrial scenarios, low contrast between defects and complex background textures severely impedes accurate localization and recognition, leading to inadequate modeling of multi-scale textures and fine-grained details, and ultimately degrading detection accuracy. To address these challenges, we propose a novel Deep Texture Exploration and Multi-Scale Interaction Network (TexMS-YOLO), specifically designed to enhance texture-aware and scale-sensitive defect detection. First, we develop the Texture-Aware Feature Fusion (TAFF) module, which fuses high- and low-resolution features across spatial and frequency domains. Through adaptive parameter optimization, TAFF efficiently extracts discriminative texture representations at multiple scales, thereby enhancing the model’s sensitivity to texture-related defect features. Additionally, a learnable high-pass filtering mechanism is introduced to preserve edge structures and fine detail variations essential for identifying subtle defects. Furthermore, we introduce a Multi-Scale Interaction (MSI) module that employs gated units and spatially adaptive structures to dynamically select scale-relevant features. A bottom-up pathway aggregates local and global cues along channels, effectively enhancing the detection of scale-variant defects. Lastly, to mitigate semantic inconsistency in traditional pyramidal frameworks, we adopt the Triple Feature Encoding (TFE) module, which applies dual-modal pooling, zeroth-order interpolation, and dynamic weighting for enhanced multi-scale representation. Extensive experiments on a self-collected relay surface defect dataset and the public NEU-DET benchmark demonstrate that TexMS-YOLO significantly outperforms existing state-of-the-art detection methods across multiple evaluation metrics, confirming its robustness and practical effectiveness in real-world industrial applications.

Source: Journal of King Saud University - Computer and Information Sciences

DOI: 10.1007/s44443-025-00389-5

Authors: Yuxin Ma, Enrui Gong, Feng Huang

Institutions: Zhejiang University of Science and Technology, Zhejiang University

Language: English

Topic: Industrial Vision Systems and Defect Detection

Subfield: Industrial and Manufacturing Engineering

Field: Engineering

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $1,990

Funder: Zhejiang University of Science and Technology