# Enhanced Multi-Scale Defect Detection in Steel Surfaces via Innovative Deep Learning Architecture

## TL;DR

A new deep learning model called CTG-YOLO improves steel surface defect detection by combining advanced modules for better accuracy and efficiency.

## Contribution

The paper introduces CTG-YOLO, an innovative deep learning architecture with a CBY backbone, TFF-PANet fusion network, and GS-Head for edge deployment.

## Key findings

- The CBY parallel backbone improves fine-grained defect recognition and model efficiency.
- The TFF-PANet architecture enhances multi-scale feature fusion and detail capture.
- The GS-Head enables lightweight deployment on edge devices while preserving feature quality.

## Abstract

What are the main findings?
A CBY parallel backbone module is proposed in the feature extraction stage to enhance the network’s ability to understand input data and improve its fine-grained feature extraction capability for minor defects.In the feature fusion network, a novel TFF-PANet feature extraction architecture is designed to optimize multi-scale feature interaction, enhance heterogeneous feature fusion, and improve detail-capturing capabilities.The GS-Head (Grained Structure Head) detection head is designed for deployment on edge devices. This design preserves feature channel connections, enhances image texture, and compresses the model.

A CBY parallel backbone module is proposed in the feature extraction stage to enhance the network’s ability to understand input data and improve its fine-grained feature extraction capability for minor defects.

In the feature fusion network, a novel TFF-PANet feature extraction architecture is designed to optimize multi-scale feature interaction, enhance heterogeneous feature fusion, and improve detail-capturing capabilities.

The GS-Head (Grained Structure Head) detection head is designed for deployment on edge devices. This design preserves feature channel connections, enhances image texture, and compresses the model.

What is the implication of the main finding?
The proposed CBY parallel backbone module improves the model’s performance in recognizing fine-grained defects, making it more efficient in defect detection.The TFF-PANet architecture optimizes multi-scale feature interaction and enhances feature fusion, improving the model’s ability to capture detailed and diverse features.The GS-Head design ensures that the model is lightweight enough for deployment on edge devices, while still preserving critical feature information and enhancing image texture.The algorithm’s superior performance on the NEU-DET and GC10-DET datasets highlights its robustness and generalization capability, making it an effective solution for real-world industrial defect detection applications.

The proposed CBY parallel backbone module improves the model’s performance in recognizing fine-grained defects, making it more efficient in defect detection.

The TFF-PANet architecture optimizes multi-scale feature interaction and enhances feature fusion, improving the model’s ability to capture detailed and diverse features.

The GS-Head design ensures that the model is lightweight enough for deployment on edge devices, while still preserving critical feature information and enhancing image texture.

The algorithm’s superior performance on the NEU-DET and GC10-DET datasets highlights its robustness and generalization capability, making it an effective solution for real-world industrial defect detection applications.

Steel surface defects significantly impact product quality and safety in industrial settings. Traditional defect detection methods suffer from inefficiencies and limitations. This study introduces an innovative deep learning architecture, CTG-YOLO, designed to enhance multi-scale defect detection accuracy on steel surfaces. By integrating a CBY parallel network structure, a TFF-PANet neck network, and a GS-Head detection head, our model achieves superior feature extraction and fusion capabilities. Experimental results on the NEU-DET and GC10-DET datasets demonstrate significant improvements, with mean Average Precision (mAP) scores of 76.55% and 69.94%, respectively, outperforming the original YOLOv8s by 3.72% and 3.14%. This research provides a robust foundation for industrial defect detection applications.