Automatic Classification of Web Images as UML Static Diagrams Using Machine Learning Techniques

Round 1

Reviewer 1 Report

This paper proposes a method and a tool to classify web images in different formats based on UML static diagrams.

The authors show that the proposed methodology outperform the methods proposed earlier, which are also based on UML diagrams.

However, in this work, there is completely no comparison of methods implemented on UML diagrams with methods that already exist and are actively used both scientific and commercial purposes, namely, algorithms that directly extract features from images, and there is no comparison with neural networks, which is de facto standard tool in computer vision.

Accordingly, several questions arise.

Why do you need to use the UML transformation, then to use standard classification procedures such as SVM, Random Forest and so on?  If, first of all, there are ready-made algorithms for extracting features from images. Secondly, it well known that such methods work quite effectively.

In general, it is generally not clear why to use the UML approach? if there are ready-made algorithms based on neural networks that work very well.

 

What should be done?

The table shows the tests for different algorithms where UML technology was used. But tests with non UML diagrams algorithms are not presented. It is necessary to compare the results of image classification, where features were obtained in different ways. One feature extraction based on UML diagrams? Second one based on machine learning algorithm (for example, see the OpenCV package).

In general, in machine learning, there is a field of computer vision. Within this area, there are a number of algorithms that can outperform UML-based algorithms.

Based on this, it is necessary to add a comparison of the UML approach with existing algorithms (outside the UML approach) and show this difference. If the UML approach is comparable in quality and works faster, then this article can be printed, because results are significant. If not, then the UML diagrams approach is a dead end branch in the field of image classification.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

1. Too little methodology, especially about how to use machine learning techniques.
2. Table 4, Summary of the results of the classifiers obtained with the 10 experiments; but there was no detailed information about these available data sets in this paper.
3. The title of Figure 10 was not on the same page as the figure itself.
4. Lines 880, 881 (References 8) must correct text layouts.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The main problem of the article is that the authors did not compare models based on the UML approach with models where the selection of features is based on standard methods in the field of computer vision.

About interpretability of UML in comparison with neural networks.

This is a rather weak argument, because, firstly, there are classifiers whose work is well interpreted, for example, naïve byes or KNN, and secondly, there is a good tradition in the field of machine learning (for journal with high ratings) where it is necessary to compare different approaches.

Moreover, if the algorithm is very good at classifying, it does not matter what the essence of the algorithm itself is. In this case, quality is main thing.

Thus, authors simply need to add one or two algorithms as a base line, where feature extraction is implemented not based on UML. There are ready-made packages in python that are in the public domain.

Author Response

We have made the following modifications and additions:

At the beginning of section 4 (lines 313-336), an explanation of traditional techniques in computer vision, in order to compare them with our own ad-hoc approach that lowers the computational complexity and achieves better image processing times.

In section 6 (lines 626-627), a notice that our approach (rule induction) will be compared with neural networks approach in section 7.

In section 7 (lines 703-709), results of experiments performed with a neural networks approach (Table 4), showing there is not a statistically significant difference with the rule induction approach.

In section 9 (lines 748-751), a paragraph to recapitulate these new results.

Four new references 27-30 (lines 861-868).

Round 3

Reviewer 1 Report

Now it is ok.

Thank you for your patience

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.

Round 1

Reviewer 1 Report

The idea and the goal behind the paper seem to be well-motivated, and the authors spend a lot of space to explain the motivations behind the work. I believe the ideas behind the papers are interesting. However, the authors did not seem to have spent enough time making the presentation quality sufficiently good for a journal submission. This is a bit disappointing as there seem to be a lot of work behind this paper. As such, I think the technical part of the paper is ok, while the abstract (which should also contain the main conclusion), experimental result and conclusion sections are too brief. Especially, the experimental results should be more elaborated and better analysed (as promised in the abstract).
Overall, my general recommendation is to rewrite this paper and resubmit after it has been improved. When working on the improvement, I recommend the authors to spend less time introducing the concept, and more time explaining the goals and contributions of the work instead.

Author Response

see attachment

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors present an algorithm that is able to detect UML diagrams in images stating that the presented algorithm is able to perform better than other algorithms.

Some important comments that can be made are related to "non formal" language used to several parts of the research work. Despite the fact that may make it easier to understand it seems that there is a need for alteration to some sentences.

A second comment is related to the limitations of the current work that are surpassed with the statement that they will be researched at a later research work. Without arguing about the results presented it seems that some parts of the research are omitted for a "later stage". To me, a better clarification on why the soundness of the current scientific work remains despite the fact that several aspects are omitted will be useful.

Finally, and the most important. As the authors state they have a free online tool as a result of their research work where one can check the algorithm generated and presented. Actually it was very easy to produce false positives by trying images deriving from a search result on "rectangles". The weirdest was false positives on piece of art of Mondrian and his rectangles. So, to me, it seems that a large number of false positives can be produced from images that have simple rectangles - even on rectangles with common vertical lines that one could argue strongly that this is a reason for not being UML.

At the end, and after trying several images (at least 10) that are pretty clear that they are negatives as they have all the aspects to be selected as negative and the tool states that they are UML diagrams, it started being odd... More clarification on the results and the reason for all the false positives is needed in order to understand the soundness of the results.

 

Author Response

see attachment

Author Response File: Author Response.pdf

Reviewer 3 Report

The paper is well written and sounds technically. Authors are aware of related work.

The topic is already covered by a large number of authors in scientific community. Overall, it is not clear: what is the added value of this paper compared to the others? The paper provides none or only very small contributions to scientific community.

The paper is split in 11 sections. That is too much.

Sec. 1 and 2 are introduction sections and should be summarized to maximal 2 pages in 1 section.

Sect 2 contains many useless phrases e.g. “to improve not only precision, but also recall.” It is a ground truth in information retrieval that you always have to improve both. Delete these phrases.

In Sect 3 literature should be provided with impact on the provided methodology. Do not give an overall view on literature.

Sect 4 to 8 contains the methodology. But this is more a work description than a chapter in a scientific article. Summarize this to one section ‘methodology’ where you presented the methodology. In Sect 4 – 8 there are parts of several methodologies described but it remains unclear: what is the methodology you provide to the scientific community?

Sect9to11 is the experiment chapter and belong together. Note that your IT-Tool is just a tool for getting data evaluated. It is not of scientific value. The experiment itself is described very poor and it is processed also in a very poor way.

Sec 11 also contains conclusions. Write them in a separate section. The given conclusions are also very poor.

From my point of view, this paper needs a complete rewriting.

Author Response

see attachment

Author Response File: Author Response.pdf