Liquid Reservoir Weld Defect Detection Based on Improved YOLOv8s

Comments 1：As described in Section 3.1.1 (“Dataset of liquid reservoir welding seam defects”), the dataset contains only 431 samples. This small sample size raises concerns regarding the generalization and robustness of the model. Although the authors employ data augmentation techniques to improve data diversity and categorical balance, the absolute number of samples remains limited compared to the number of network parameters. It would strengthen the paper to include additional validation results—such as cross-validation or testing on larger dataset—to make the claimed performance improvements stronger.

Response 1：Agree. Thank you for your valuable feedback. We acknowledge that the weld defect dataset of the liquid reservoir is limited in scale and may affect the validation of the model's generalization ability and robustness. To address this, we have introduced a publicly available dataset of steel pipe weld defects as a supplement. This dataset includes 3,408 X-ray images covering eight types of defects. The dataset is split into a training set (2,726 images) and a test set (682 images) in a 4:1 ratio. Comparative experiments with several mainstream models under the same experimental settings demonstrate that our proposed improved model still achieves excellent performance even on this larger dataset, further validating its generalization capability. The relevant content has been added to Section 3.4.7 on page 18.

Comments 2：The authors should better justify using RepGFPN over other widely adopted feature fusion strategies such as BiFPN (EfficientDet), or Transformer-based necks (e.g., Swin-transformer). A short comparison or rationale explaining why RepGFPN was preferred.

Response 2：Thank you for pointing out the limitations in our experiments. We have supplemented relevant experiments in Section 3.4.1 on page 11. To better justify our choice of RepGFPN as the feature fusion strategy, we designed a comparative study in that section, comparing the improved model against several mainstream feature fusion strategies—including YOLOv8s-p2 with BiFPN, AFPN, and Swin transformer. The experimental results show that RepGFPN achieves better performance across multiple evaluation metrics and more effectively balances detection accuracy and inference speed, thereby validating its suitability and advantage for the weld defect detection task.

Comments 3：The heatmaps in Figure 10 are informative but lack a clear scale or color legend. Including one would help readers better interpret the model’s focus regions and attention behavior.

Response 3：Agree. Thank you for your suggestion regarding the figures. Following your recommendation, we have added a corresponding color legend to the heatmap in Figure 10 (now Figure 9 on Page 16). The legend clearly indicates the defect type represented by each colored bounding box, helping readers better understand the model's attention intensity and distribution across different defect regions.

Comments 1：Kindly discuss how you generate the weld defects images. Do you capture  x-ray images? Mention this under 3.1.1.  Also, why steel pipe weld defects dataset? Discuss this under 3.1.2.

Response 1：Agree. Thank you for your valuable feedback. We have made the following additions to the manuscript as suggested:

In lines 245-260 on Page 9, we have added a description of how the liquid reservoir weld defect images were generated: In this study, a fixed line laser profiler scans the weld seam on a uniformly rotating liquid reservoir via a servo-driven mechanism, acquiring 3D contour data at 3 seconds per unit. As this data format is incompatible with standard detection models like YOLOv8, we transform the 3D point cloud into grayscale images by mapping depth to pixel intensity, a process that averages 15–30 ms. This conversion not only reduces data dimensionality and enhances defect visibility but also eliminates interference from color and texture variations. The resulting images enable direct training of deep learning models for efficient weld defect identification, bypassing complex 3D processing.

In lines 469-478 on Page 18, we further elaborate on the rationale for introducing the steel pipe weld defect dataset. The defect types in steel pipe and liquid reservoir welds are highly similar, both covering typical defects, ensuring the validity of the comparison. Although the component shapes differ, the data modality remains consistent, as both are morphology-based grayscale images. This consistency helps the model learn component-agnostic general features. Moreover, this large-scale public dataset alleviates the issue of limited sample size and enables a more comprehensive evaluation of the model’s generalization capability across diverse defect patterns.

Comments 2：Under 3.3, discuss which metric is the key for performance analysis. In results section, your data presents that some models exhibit superiority in terms of precision score while your model performance well in terms of mAP@0.5.

Response 2：Agree. Thank you for your valuable feedback on this matter. In response to your suggestion, we have added a clarification of the key performance evaluation metrics in Section 3.3 on Page 10. For the weld defect detection task in this study, we select mAP@0.5 and mAP@0.5:0.95 as the primary metrics for assessing model performance. By combining these two types of metrics, we are able to evaluate both the practical performance of the model in general industrial scenarios and its comprehensive performance under strict localization criteria. This approach helps prevent potential evaluation bias that could arise from relying on a single metric, thereby providing a more thorough basis for judging the model's reliability in real-world applications.

Comments 3：Figure 9: state that this is for liquid reservoirs dataset.

Response 3：Agree. Thank you for your suggestion regarding the figures. Following your advice, we have revised the title and corresponding description of Figure 9 (now Figure 8 on Page 14) to explicitly state that the chart presents a comparison of AP values for each defect category in the liquid reservoir weld defect dataset. This clarification ensures a clear and unambiguous connection between the visual content and its corresponding dataset.

Comments 4：As of figure 10, isn't YOLOv8s with RepGFPN sufficient? When comparing images (c) and (d) in Figure 10, the results obtained by (c) is sufficiently accurate.  Also, by incorporating RepGFPN-FocalNets-CGA into YOLOv8s, you are increasing the system complexity. Please comment on this.

Response 4：Thank you for your insightful comments. As illustrated in Figure 10 (now Figure 9 on Page 16)(c), while the YOLOv8s model with RepGFPN can initially localize the defect, its attention coverage and response intensity are discernible yet comparatively weaker than those of our full model presented in (d). The heatmaps generated by our full model exhibit a more concentrated and coherent attention distribution, demonstrating its enhanced capability to accurately capture subtle defect features and their structural context.The relevant content has been added to lines 448-458 on page 17.

Although the introduction of RepGFPN, FocalNets, and the CGA module slightly increases model complexity, the ablation studies in lines 388-394 on Page 12 show that these improvements lead to only a marginal growth in parameters while significantly enhancing key metrics such as mAP@0.5 and mAP@0.5:0.95. This demonstrates that the introduced modules effectively improve the model's capability to recognize diverse defects and its generalization performance, resulting in better robustness in practical applications. Therefore, we believe the performance gains achieved within a controllable complexity range hold clear engineering value.

Comments 5：Improve the conclusion by adding statements on performance imporvement achieved. It is advised to add numerical values.

Response 5：Based on your suggestion, we have supplemented the conclusion section on Page 18 with specific performance improvement figures to more clearly demonstrate the enhancements:
Ablation studies and comparative experimental results show that the improved model achieves a 6.3% increase in mAP@0.5 and a 4.3% increase in mAP@0.5:0.95 compared to the baseline model. For key defect types, the AP values for arc craters, porosity, undercuts, and lack of fusion are improved by 3.9%, 13.5%, 5.0%, and 2.5%, respectively, effectively validating the superiority of the proposed improvements. We appreciate your pointing out the lack of specificity in our original conclusion, and have now included these quantitative results to intuitively reflect the model's performance gains.

Comments 6：Line 274 should be "Fig 7(a)" not Fig 6 (a).

Response 6：We sincerely thank the reviewer for their careful reading. The citation in question was indeed a clerical error, and we have corrected it to "Fig. 7(a)" on Page 9. Additionally, we have thoroughly reviewed all figure and table references throughout the manuscript to ensure that such errors do not occur elsewhere.