A Checkerboard Corner Detection Method for Infrared Thermal Camera Calibration Based on Physics-Informed Neural Network

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper propose a YOLO model for corner detection and an unsupervised physics-informed training method for accurate subpixel extraction of checkerboard corner points in an infrared camera calibration task. The approach consists of two stages: an improved YOLOv8 model and the integration of physical information. The proposed method has been validated in the calibration of high-magnification infrared cameras and shows potential for extension to more complex optical calibration scenarios. However, the following issues remain:

1. Step 4 of the first stage (Scaling the depth and width of the network) lacks detailed explanation. How are the parameters specifically adjusted? Is it implemented adaptively? It is recommended to provide further clarification.
2. In Section 3.4, the ablation experiment evaluates the accuracy after each step, rather than removing individual improvement steps. This design does not achieve single-variable control. Further improvement is recommended.

3. According to the results, the improvements made to the YOLOv8 model in the first stage have limited impact on performance, while the incorporation of physical information in the second stage significantly enhances accuracy. It is recommended to supplement a comparative experiment using the original YOLOv8 model combined with physical information to verify the effectiveness of the first-stage improvements.

1. The paper only analyzes the computational efficiency of the first-stage YOLOv8 model improvement. It is recommended to include an analysis of the computational time for the second stage as well.

Comments on the Quality of English Language

The overall expression of the paper is fluent, with some minor adjustments that could be made.

Author Response

Please see the attachment which contains the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The submitted manuscript focuses on the development of a robust checkerboard corner detection method tailored for low-quality images characterized by low resolution and contrast. The proposed approach leverages Physics-Informed Neural Network (PINN) to address these challenges. While the study primarily demonstrates the method’s application using infrared (IR) thermal imaging data, the potential scope of application extends significantly beyond this domain, rendering the research topic highly relevant.

The authors have appropriately selected baseline methods for comparative analysis and validation, ensuring the reproducibility of the presented results. The evaluation metrics employed are well-established within the field, and the proposed method demonstrates better accuracy compared to existing approaches, particularly in terms of enhancing camera calibration accuracy.

Despite the overall good quality of the manuscript and importance of the main idea, there are several minor but significant revisions to be made.

Mandatory revisions

1. The manuscript should include detailed specifications of the calibration board dimensions and data acquisition procedures. This should encompass a distance range between camera and calibration board, tilt and rotation angle range.
2. Table 1 requires clarification by adding appropriate units of measurement for all parameters, including focal length (mm or arbitrary units) and image dimensions (pixels).
3. The discrepancy between “digital zoom magnification” (8x, line 152) and “digital zoom ratio” (1.19, Table 1) needs clarification.
4. Introduction section. The authors acknowledge the capability of thermal imaging cameras to detect minute temperature variations. However, it is important to note that the actual temperature contrast captured by radiometric IR cameras in real-world (non-gray) objects is significantly affected by the inherent ambiguity between emissivity and temperature measurements. Furthermore, the research team exploits this ambiguity in their methodology by designing a calibration board composed of materials with varying emissivity values. This approach aims to create a discernible contrast between the «black» and «white» squares while maintaining uniform temperatures across the board. The details associated with the ambiguity mentioned should be clarified.

Recommended improvements

5. The phrase “A set of real infrared thermal camera calibration dataset is constructed and based on this dataset…” (lines 131-132) requires clarification. The distinction between “set” and “dataset” should be explicitly defined to avoid confusion.
6. Figure 7. Figure 7(a) uses corner markers different from other sub-figures. Probably, this is due to the standard figure legend presets implemented in MATLAB. Using the uniform corner markers for all sub-figures may lead to more convenient visual comparison.
7. Figure 7 serves well the main purpose of demonstrating the ability of the developed method to avoid missed corners. But, if there is a significant difference of corner positions between methods, it is barely possible to notice. To highlight the position differences, the enlarged images of single corners may be added to each sub-figure.

Author Response

Please see the attachment which contains the revised manuscript.

Author Response File: Author Response.pdf