# SD-GASNet: Efficient Dual-Domain Multi-Scale Fusion Network with Self-Distillation for Surface Defect Detection

## TL;DR

This paper introduces SD-GASNet, a new network for detecting surface defects in industrial settings that improves accuracy and efficiency using self-distillation and multi-scale feature fusion.

## Contribution

The novel contribution is the SD-GASNet architecture combining self-distillation with dual-domain multi-scale fusion for efficient and accurate surface defect detection.

## Key findings

- SD-GASNet achieves state-of-the-art performance on three public datasets (NEU-DET, PCB, TILDA).
- The model delivers an inference speed of 180 FPS while maintaining high accuracy.
- The proposed architecture effectively handles subtle defects and sensor-based industrial images.

## Abstract

Surface defect detection is vital in industrial quality control. While deep learning has largely automated inspection, accurately locating defects with large-scale variations or those difficult to distinguish from similar backgrounds remains challenging. Furthermore, achieving high-precision and real-time performance under limited computational resources in deployment environments complicates effective solutions. In this work, we propose SD-GASNet, a network based on a self-distillation model compression strategy. To identify subtle defects, we design an Alignment, Enhancement, and Synchronization Feature Pyramid Network (AES-FPN) fusion network incorporating the Frequency Domain Information Gathering-and-Allocation (FIGA) mechanism and the Channel Synchronization (CS) module for industrial images from different sensors. Specifically, FIGA refines features via the Multi-scale Feature Alignment (MFA) module, then the Frequency-Guided Perception Enhancement Module (FGPEM) extracts high- and low-frequency information to enhance spatial representation. The CS module compensates for information loss during feature fusion. Addressing computational constraints, we adopt self-distillation with an Enhanced KL divergence loss function to boost lightweight model performance. Extensive experiments on three public datasets (NEU-DET, PCB, and TILDA) demonstrate that SD-GASNet achieves state-of-the-art performance with excellent generalization, delivering superior accuracy and a competitive inference speed of 180 FPS, offering a robust and generalizable solution for sensor-based industrial imaging applications.