Towards Improving YOLOv8-Based Defect Detection Methods for Aluminum Profiles

Abstract

In the field of aluminum surface defect detection, current detection models com
monly suffer from insufficient detection accuracy, weak real-time performance,
and excessive parameter size. To address these issues, we propose a novel method
based on the YOLOv8 object detection model. Firstly, an EMA (Efficient Multi
scale Attention) module is incorporated to enhance feature extraction efficiency
and accuracy. Secondly, the conventional convolution is replaced with Repa
rameterized Convolution, which employs a multi-branch structure during the
training phase to strengthen feature extraction capabilities while adopting a
single-branch structure during the validation phase to accelerate detection speed
through parameter re-parameterization. Finally, a composite regression loss func
tion is implemented that combines the CIoU loss and the WIoUv3 loss to enable
dynamic anchor box screening, which effectively resolves inaccurate label assign
ment caused by variations in sensitivity of the IoU in defect types of different
sizes. Experimental results demonstrate that compared with the baseline model,
our method achieves a 4.6% increase in mean average precision (mAP), attains
a processing speed of 94.0 frames per second (FPS), and maintains a compact
model size of merely 3.12 MB. These advancements validate that the proposed
algorithm exhibits superior recognition accuracy, rapid detection speed, and
minimal memory footprint, demonstrating significant potential for industrial
deployment.