Inspection System for Vehicle Headlight Defects Based on Convolutional Neural Network

Round 1

Reviewer 1 Report

This paper presents a method to detect defects in vehicle headlights based on a convolutional neural network. The proposed method comprises two main steps: in the first step an inspection region (ROI) is set after position calibration of an acquired image; and the second step consists in the application of a convolutional neural network to determine the presence of defects within the ROI. Although the results presented are promising, this paper needs several improvements to help clarify its real contributions, and to present the method in a more clear manner. Additionally, there are some parts of the paper that are difficult to read and understand and must be improved. The writing of the paper has to be thoroughly revised because there are several unclear parts, and poorly constructed sentences.

The paper is organized in 5 sections with the proposed method being described in section 3 (core section). The introduction section of the paper is poor and must be significantly improved:

1 – It should clearly describe the contribution of the paper: what is/are the novelty(ies) of the proposed methodology? In the experimental results the proposed methodology is compared with the method described in [17], what are the main contributions of the new method when compared to this one?

2 – The first paragraphs of introduction are devoted to the motivation to develop the proposed methodology, but unfortunately this is performed in a very confusing manner that does not help the reader to understand the objectives that are intended to be achieved with this work.

3 – The related work description is poor and the list of references are insufficient.

4 – The goals of the work should be clearly settled at the Introduction. Clearly explain what a defective headlight is, and clearly explain what your system should do.

Section 2 describes the system configuration, where the authors present the proposed system structure that automatically determines the defects for different types of parts. Concerning this section I have the following remarks:

1 – There should be a match between Figure 2 and the items described at the end of section2: Inspection lighting module; Camera module; Inspector module; PLC module.

2 – Figure 1b) and 1c) show us two headlight images that are very similar. The proposed method must be able to identify the type of headlight and if it is defective or not? Please clarify. Or is there a previous model selection before the inspection process starts? This is very difficult to understand, and Figures 1, 2 and 3 must be improved to help the readers understanding the proposed system and the method.

The authors describe the program structure at the beginning of section 3, however Figure 3 is difficult to understand (the arrows are difficult to follow) and several of its contents are not explained, such as: Start module; motor (what is the role of the motor???),etc…

1 – In the object acquisition and object major axis coordinate search subsection, the process described by the numbers 1 to 6 does not match with Fig. 6c)

2 – In subsection 3.2, the authors describe the image size conversion process required to apply a certain CNN structure. Although the authors mention several CNN structures proposed in the literature they decided to choose VGG19. Why? Were there any experiments using other structures? Is this the best choice?

Author Response

Thank you for your detail comments.

Author Response File: Author Response.pdf

Reviewer 2 Report

The manuscript proposes a convolutional neural network (CNN) to detect the headlight defects' defects to determine if they are defective based on the region of interest (ROI). Compared with the conventional method, the results reported a performance improvement of more than 15.69%, on average.

The vision-based inspection system has been extensively applied to various industries, particularly vehicle parts defect detection linked with the intelligent factory notion. The CNN method enhances the inspecting accuracy and decreases the cost for the human inspection. As such, the method used in the paper is not new, but its application to distinguish defects in vehicle headlights could be considered the contribution of the paper. 

Some minor comments

• Fig 13 and 14 could be improved.
• Reference no 17: 'Inspection of Vehicle Headlight Defects", Journal of the Korea Industrial Information Systems Research', what is the difference between the article, which has already been published and the present article?
• Six out of seventeen references in the article belong to the authors with the same objective set in the manuscript.

Author Response

Thank you for your detail comments.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

This paper presents a method to detect defects in vehicle headlights based on a convolutional neural network. The proposed method comprises two main steps: in the first step an inspection region (ROI) is set after position calibration of an acquired image; and the second step consists in the application of a convolutional neural network to determine the presence of defects within the ROI.

The paper is much improved from previous version, and although the proposed methodology is not new, the proposed application and the obtained results are enough contribution.

Although the paper is much more clear, there is still room to improve paper presentation and writting. For instance, try to improve Figure 3 in order to help the reader getting the whole picture of your system in an easy manner. What is "a" and "b" in Figure 7b) - what is written in the images must match the text .

Author Response

Thank you for your comment.

Author Response File: Author Response.pdf