In the study, four different models were used to understand the effect of light on the data set. These models will be discussed in turn and it will be analyzed in which dataset each of them gives the best results. In addition, the best success rate of each model will be presented in the form of a table at the end of the section. The parameters such as iteration number, learning rate, batch size, and subdivision values of were selected as 12000, 0.001, 64, 24 during the training of each model, respectively. The performance of each model is given in Fig. 11.
As shown in Fig. 11, the brightness ratios in the datasets have an effect on the success rates of the models. In the experimental study on which brightness improves the model positively, it has been observed that all models give better results in 10% illuminated datasets. The YOLOv4 model takes 416x416 images, but the dataset consists of 200x200 images. After the input images are converted to 416x416 size, it goes to the training phase Among five different datasets with changed light values, the best success rate of the dataset with 20% darkening is 67%, and the best success rate of the dataset with a blackening rate of 10% is 68%. It has been observed that the darkening to be made in the dataset for YOLOv4 almost do not affect the success of the model. The success of the original dataset is 67%. It is observed that the success rate changes positively in the illuminated datasets. The 10% illuminated dataset gives the best results with 71% mAP. Although the dataset with 20% illumination gave better results than the original dataset with dimming, it gave worse results than the dataset with 10% illumination. The most successful dataset is the 10% applied dataset. The YOLOv5 model, which came out after the YOLOv4 algorithm is a more advanced model. Using the same parameters, five different trainings were applied to the datasets. In these training, the best success rate of the dataset with 20% darkening is 75.9%, and the best success rate of the dataset with a darkening rate of 10% is 77.1%. It has been observed that darkening in the dataset for YOLOv5 has a negative effect on the success of the model, as in YOLOv4. The success of the original dataset is 76.4%. It is observed that the success rate in illuminated datasets increases positively as in YOLOv4.
The 10% illuminated dataset gives the best results with 79.7% mAP. Although the dataset with 20% illumination gives better results than the original dataset with darkening, it gives worse results than the dataset with 10% illumination. The most successful dataset is the 10% applied dataset. The YOLOX model used in the study was run as the default parameters specified in [33]. Although it achieved higher success rates in a shorter time compared to other models, it did not achieve as much success as the YOLOv5. The last model YOLOR in the study was run in 500 iterations. Since there was no improvement in performance after 500 iterations, all algorithms were run 500 iterations. As in the other models, the parameters are set to be equivalent, but it did not produce as successful results as the YOLOv5. When evaluated with the lighting element, the dataset with 20% darkening achieved 72.7% success and gave the most unsuccessful training result as in the other models. Figure 12 shows the detection results of the best performing YoloV5. As shown in Fig. 12, it can be seen that different defects were successfully detected on the same image. The original dataset and the success rate after 10% illumination are given in Table 1. In Table 1, information about the performance ratios of the models is also given.
Table 1
Detection of best result of the dataset.
Model Name
|
Size
|
Parameters
(M)
|
Layers
|
GFLOPs
|
Latency
(ms)
|
FPS
|
Best Dataset
mAP (%)
|
Improvement Rate (%)
|
YOLOv4
|
608x608
|
27.6
|
137
|
52
|
44
|
22.7
|
71.00
|
4.0
|
YOLOX
|
640x640
|
8.94
|
286
|
26.65
|
33
|
30.3
|
73.20
|
2.7
|
YOLOR
|
416x416
|
36.8
|
665
|
80.44
|
22
|
46
|
75.90
|
0.6
|
YOLOv5
|
416x416
|
7.3
|
232
|
16.8
|
17
|
58.8
|
79.70
|
3.3
|
As shown in Table 1, there are differences in performance and success rate between models. This study was carried out on a server with a Tesla P100 graphics card and 16 GB of RAM. The results show that the YOLOv5 architecture performs better and is more successful than the latest YOLOR and YOLOX architectures, which have not been used in the field of metal surfaces before. When YOLOX and YOLOR architectures give better results than YOLOv4, it is seen as an output of experimental results. While the YOLOR model showed the least improvement rate on the illuminated dataset, the YOLOv4 model showed the highest difference, however, the YOLOv5 model is the most suitable model for detecting defects on steel surfaces among the YOLO models. When the recent defect detection studies on the NEU dataset are examined, no previous studies have used YOLOX, YOLOR and YOLOv5 models, and no studies have been found to examine the effect of lighting on success. In addition, the study has the uniqueness of the first metal surface defect detection study using YOLOX and YOLOR models. In 2022, Tian et al. [39] proposed the DCC-CenterNet model to get fast and accurate results on metal surfaces. ResNet50 was preferred as the backbone structure of the model and DFEM (Dilated feature enhancement model) head structure was used. In 2021, Kou et al. [40] developed a YOLOv3-based model for damage detection on the NEU dataset. Studies that applied feature selection to improve the model achieved 72.2% mAP. In 2021, Cheng and Yu [41] used the DEA_RetinaNet model on metal surfaces. This model, which basically consists of five parts, consists of feature extraction network, DE-block, FPN, adaptive spatial feature coupling (ASFF) module and prediction network modules. The authors obtained 78.25% mAP as a result of the study. In 2021, Song et al. [42] used the multiscale adversarial and weighted gradient-domain adaptive network (MWDAN) model to detect the defects of metal surfaces. This model is obtained by applying HRNet [43] as a backbone to the Faster R-CNN model. Researchers achieved a success rate of 76.2% in their study. The comparison results are given in Table 2.
Table 2
Comparison results on NEU dataset of different methods.
Reference
|
Method
|
FPS
|
Parameters (M)
|
mAP(%)
|
[41]
|
DCC-CenterNet
|
71.3
|
32.8
|
79.4
|
[42]
|
YOLO-V3
|
64.5
|
60.7
|
72.2
|
[43]
|
DEA_RetinaNet + ASFF
MWDAN
|
12.0
-
|
28.5
-
|
78.2
76.4
|
This study
|
YOLOv4 Darknet + (10% Illimunated)
YOLOR + (10% Illimunated)
YOLOX + (10% Illimunated)
YOLOv5 + (10% Illimunated)
|
22.7
46.0
30.3
58.8
|
27.6
36.8
8.9
7.3
|
71.0
75.9
73.2
79.7
|
In Table 2 it is observed that the object detectors based on the mentioned YOLO model have less success on the NEU dataset than the YOLOv5 model and have more parameters. In addition, the effects of lighting and darkening were not examined in these studies. The study clearly shows that the image enhancement techniques on the models make an absolute contribution to the success of the model. In studies with another dataset, the steel dataset, the effect of brightness was measured as in the NEU dataset. According to the experimental results, it is fixed according to the results of the two datasets that the images taken with the camera live better performance against the lighting. The proposed approach has been applied on the steel pipe defects dataset. In the experiments on this dataset, 10% lighting gave the best result. In the Table 3, performance, success rate, improvement rate (%), and model details are given for the steel pipe defects dataset under 10% lighting conditions.
The FLOPs mentioned in the Table 3 are the operations of a computer’s floating point unit in one second. One GFLOPs is one billion FLOPS per second, or floating point operations. In the datasets obtained with lighting and darkening, the dataset with 10% lighting achieved the best results. The improvement rate best dataset represents the additional success rate over the original dataset. The performance comparison of the illuminated dataset is shown in Table 4.
Table 4
Performance comparison of steel pipe detection algorithms Object
Object Detection Model
|
Accuracy, precision, [email protected]%
|
GAN + CFM [44]
|
85.90 accuracy
|
OSTU + MSVM-rbf [45]
|
95.23 accuracy
|
Faster R-CNN + ResNet50 [46]
|
78.10 [email protected]
|
YOLOv5x (over 10x image augmentation)[47]
|
97.8 pre
|
Yolov4 (only 10% Illimunated)
|
84.30% [email protected]
|
YOLOX (only 10% Illimunated)
|
86.50% [email protected]
|
YOLOR (only 10% Illimunated)
|
89.80% [email protected]
|
YOLOv5 (only 10% Illimunated)
|
97.60% [email protected]
|
When the results of Table 4 analyzed, it is seen that the YOLOv5 model has been developed to 97.60% success with only 10% illumination. Yan et al. [38] achieved 97.8% success by performing data augmentation on the same data set. In the data augmentatrion process, the data size has been increased 10 times. As the model weights increase because of augmentation of the data set, the training time also increases approximately ten times. In our paper, almost the same success rate was achieved by making 10% illumination with a simple pre-processing without increasing the data, and efficiency was achieved in terms of performance. The following figure shows the YOLOv5 results of the steel dataset.
As shown in Fig. 13, YOLOv5 steel pipe defect can also produce very effective results on dataset. As a result of the analyzes made, it has been observed that the YOLOv5 model can work in the real-time system and achieves the highest accuracy compared to other models. In addition, it has been observed in both datasets that the accuracy can be highly increased with only a simple pre-process lighting, in terms of performance and cost, without the need for data augmentation. Despite the different datasets and different numbers of classes in our study, the YOLOv5 architecture has been identified as the architecture that gives the most effective results in terms of performance, success and cost.
Table 5
Comparation of Illuminated and Darkened Processing on Steel Pipe Dataset Model Name
Models
|
Illuminated Rate
|
Original
|
Darkened Rate
|
10%
|
20%
|
0%
|
10%
|
20%
|
Yolov4
|
82,00%
|
81,80%
|
80,50%
|
80,20%
|
79,70%
|
YOLOX
|
83,80%
|
83,80%
|
83,40%
|
82,60%
|
81,80%
|
YOLOR
|
88,60%
|
86,40%
|
88,10%
|
86,50%
|
85,60%
|
YOLOv5
|
94,90%
|
93,20%
|
93,40%
|
91,30%
|
90,20%
|
Table 6
Comparation of Illuminated and Darkened Processing on NEU Dataset
Models
|
Illuminated Rate
|
Original
|
Darkened Rate
|
10%
|
20%
|
0%
|
10%
|
20%
|
Yolov4
|
71,00%
|
69,00%
|
67,00%
|
68,00%
|
67,00%
|
YOLOX
|
73,20%
|
73,20%
|
70,50%
|
73,20%
|
72,10%
|
YOLOR
|
75,90%
|
75,20%
|
75,30%
|
74,50%
|
72,70%
|
YOLOv5
|
79,70%
|
76,60%
|
76,40%
|
77,10%
|
75,90%
|
In addition, to better highlight and analyze the performances of the YOLO models, the results of the datasets with different illuminated and darkened processes are given in Tables 5 and 6, respectively. As can be seen from the obtained results, the most satisfactory results are obtained with 10% illumination processing.