Enhancement of GUI Display Error Detection Using Improved Faster R-CNN and Multi-Scale Attention Mechanism

Round 1

Reviewer 1 Report

1.      In this research work, authors have addressed four display error detection such as component occlusion, image loss, text overlap, and empty values by developing Faster R-CNN

2.      Using ResNet-50 the Feature Pyramid Network and the Enhanced Multi-Scale Attention algorithm are developed. To improve the generalization capability ROI-Align is used. An algorithm is proposed to generate screenshots based on the Rico dataset for training the models.

3.      Related work can be enhanced by categorizing and elaborating research work on each of the four display rendering problems. Earlier data generation work can be included.

4.      Algorithm 1: Heuristic-based training data auto-generation can be improved by using generic constants instead of actual values, also ‘past block on’ can be replaced by mathematical operation. Use function or mathematical operation for ‘Generate block image with [width, height] and background’, ‘resize icon to’ ,’ draw ’null’ string on (x1, y1) with font and background’, ‘random uniform between 0.8*width and 0.85*width’, ‘random uniform between 0 and 1;’, ‘crop image by ()’

5.      Provide more examples of four GUI errors. At least provide 4 examples for each.

6.      Include the discussion section and provide the justification for selecting the particular layers, in the feature extraction network. What are the benefits of the proposed network structure in the Faster R-CNN? Provide elaboration in the discussion section. Also, mention reasons why the proposed architecture works better on the given data samples and where it does not perform well.

7.      The Fig. 2 and Fig 4 using the colour combination as used in Fig. 5.

8. Table 3 provides a comparison of only three methods. It is suggested to include the performance of relevant methods which are mentioned in the related work section.

Can be improved. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

I have several questions fur authors 

1. There is no mention on computational complexity of the method, so it is difficult to assess the usefulness of the approach. Coukd you provide some time related information?

2. You mention testing industrial applications, but all the data you use is from Android app which are rather everyday apps, could ypunexplain?

3. Is the meghod transferable, for example as you trsin it on Anfroid examples could you test Winfows based GUI?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

The authors have answered all my issues, so I think the paper is now ready for publication.