CISC-YOLO: A Lightweight Network for Micron-Level Defect Detection on Wafers via Efficient Cross-Scale Feature Fusion

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper proposes a detection network for wafer surface micro-defects based on the YOLOv8 framework. Its lightweight structure provides a high-precision, low-latency real-time defect detection solution for the semiconductor industry. The following are some comments and suggestions that may be considered:

• The overall structure of the paper is relatively complete, but some paragraphs are long and the formula derivations are redundant, which can easily cause reading burden.
• The paper constructs a WSDD dataset, but it doesn't yet specify whether it will be publicly available or limited to laboratory use. If the dataset isn't publicly available, the reproducibility and generalizability of the model will be limited. The authors are advised to provide additional information on dataset acquisition, annotation consistency, and future plans for open access.
• While the article proposes modules such as IGC, CCFF-ISC, and DyHeadv3, it lacks sufficient evidence to demonstrate their essential differences and advantages over existing lightweight YOLO improvements. The authors are advised to provide a more in-depth analysis of the motivations behind each module's design in the method introduction, and clearly explain its uniqueness compared to existing improvements.

Author Response

Thank you very much for the reviewers' comments. Please find the detailed responses in the PDF file attached below.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

While this manuscript demonstrates solid technical contributions, it contains significant methodological weaknesses and inconsistencies in conclusions that require substantial revision and justification:

1. To maximize reader comprehension, every visual element in a diagram should have specific and easily interpretable meaning. In Figure 1, the use of various types and colors of arrows (e.g., green arrows, orange arrows, and black dashed lines) lacks accompanying explanation. This creates ambiguity regarding the relationships or processes intended to be illustrated. It is strongly recommended to add a legend or enhance the figure caption to explicitly define the meaning of each arrow type.
2. The data augmentation methodology in this manuscript has a fundamental flaw. The authors performed augmentation on the entire dataset before splitting it into training, validation, and test sets. This procedure is methodologically incorrect as it causes data leakage, where variants of the same image can contaminate the validation and test sets. This renders the evaluation results unreliable, as the reported performance metrics are likely inflated and do not reflect the model's true generalization capability.
3. The claim of "significant breakthroughs" is overstated. The manuscript is limited to YOLO-based comparisons. A more comprehensive comparison with state-of-the-art methods is critically needed. The authors should compare against more lightweight detection methods and include detailed trade-off analysis (RT-DETR-MobileNetV4, MDD-DETR, EfficientViT, ShuffleNet V2, etc.) to substantiate their claims.
4. The claim of "excellent generalization ability" is excessive. Validation on datasets with potential data leakage risks (see comment 1) is insufficient to support such strong generalization claims. Moreover, the limited dataset requires additional validation or stronger justification. The lack of statistical analysis undermines the credibility of performance improvement claims. The authors must add statistical analysis including confidence intervals, significance testing, and error analysis.
5. Include comprehensive ablation study results with systematic ablation scenarios: YOLOv8n; YOLOv8n+IGC; YOLOv8n+IGC+CCFF_ISC_neck; YOLOv8n+IGC+CCFF_ISC_neck+DyHeadv3; YOLOv8n+CCFF_ISC_neck+DyHeadv3; YOLOv8n+CCFF_ISC_neck; YOLOv8n+DyHeadv3.

Author Response

Thank you very much for the reviewers' comments. Please find the detailed responses in the PDF file attached below.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The response solved all my questions. It's ready for publication.

Reviewer 2 Report

Comments and Suggestions for Authors

The answers were satisfactory, but the manuscript's similarity level should be below 5%. Then, it can be published without hesitation