A Real-time and Efficient Surface Defect Detection Method Based on YOLOv4

Abstract

In order to achieve a better balance between accuracy and speed with limited storage and computing resources in the field of industrial defect detection, a lightweight and fast detection framework Mixed YOLOv4-LITE series is proposed based on You Only Look Once (YOLOv4) in this paper. To reduce the size of model, MobileNet series (MobileNetv1, MobileNetv2, MobileNetv3) and depthwise separable convolutions are employed in the modified network architecture to replace the backbone network CSPdarknet53 and traditional convolution in the neck and head of YOLOv4, respectively. Moreover, we combine the Mosic data enhancement method to enrich the dataset. In the training stage, Transfer Learning is used to accelerate the convergence of network, in which pseudo-convergence is precluded as much as possible by adjusting the learning rate of the cosine annealing scheduler. Finally, we evaluate the proposed methods on both public defect datasets with different types and scales, namely NEU-DET and PCB-DET. On NEU-DET, Mixed YOLOv4-LITEv1, which can detect at a rate of 88 FPS on a single GPU while maintaining the accuracy, achieves an improvement of 214% in detection speed. And Mixed YOLOv4-LITEv3 realizes an outstanding maximum improvement of 200% in detection speed while only losing a mean average precision (mAP) value of 0.11% on PCB-DET. Furthermore, the sizes of our proposed series models are only about one-fifth of the original YOLOv4 model. The extensive test results indicate that our work can provide an efficient scheme with low deployment cost for surface defect detection at different scales in multiple scenarios, which can meet the needs of practical industrial applications.

SDD-YOLO: A Lightweight, High-Generalization Methodology for Real-Time Detection of Strip Surface Defects

“…While these methods have made certain advancements in the field of metal surface defect detection, they are limited by the sensitivity of images to lighting conditions and backgrounds, and the inability of shallowly extracted manually designed features to effectively represent images with complex backgrounds. Therefore, despite the development of various traditional machine learning-based metal surface defect detection models, these models still fail to be effectively applied in practical production [19].…”

Strip Surface Defect Detection Algorithm Based on YOLOv5

“…However, traditional image processing methods usually require complex threshold settings for defect recognition and are sensitive to some environmental factors such as lighting conditions and the background, so they cannot be directly applied in reality. Although researchers have developed a series of target detection models based on various strategies, artificially designed features extracted from shallow layers cannot effectively characterize images with complex backgrounds [ 9 ].…”

Multi-Scale Infrared Military Target Detection Based on 3X-FPN Feature Fusion Network

“…Deep learning-based algorithms are widely used in target detection [10], [11], [12], [13]. The YOLO algorithm has formed the most balanced YOLOv4 algorithm in terms of detection accuracy and speed after introducing the anchor frame mechanism, the use of the feature pyramid network, the activation function changing the ReLU function to the Mish function, and making some other improvements [14], [15], [16], [17], [18]. It's also used in the fields of longdistance traffic, crops, and airborne remote sensing target detection [19], [20], [21].…”