PEŠKOVÁ, D. Detekce přítomnosti nahoty v obrazových datech [online]. Brno: Vysoké učení technické v Brně. Fakulta informačních technologií. 2024.
V rámci této diplomové práce bylo řešeno téma, ke kterému je poměrně málo dostupných zdrojů a není masivně zkoumáno na vědecké úrovni. Nicméně z hlediska forenzních věd se jedná o jedno z důležitých témat. Celkový počet konzultací měl být vyšší, a práce měla být dokončena s dostatečným předstihem. Při celkovém hodnocení zohledňuji i to, že s finální prací a realizačními výstupy jsem neměl dostatečný čas se seznámit. Celkově jsem se rozhodl hodnotit přístup studentky stupněm uspokojivě (D) .
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Informace k zadání | Cílem práce bylo vytvořit řešení pro detekci nahoty v obraze. Práce se skládala z několika úloh, které zahrnovaly určení obnaženosti těla a detekci vybraných partií. Původně byla práce navázana na spolupráci s Policií České Republiky (PČR) se kterou se definovaly cíle zadání. | ||
| Práce s literaturou | Studentka si opatřila studijní materiály a literaturu svépomocí. | ||
| Aktivita během řešení, konzultace, komunikace | Tato diplomová práce byla řešena dva roky, přičemž první rok práce byl v koordinaci s PČR. Počet konzultací v prvním roce práce byl dostačující. Z hlediska hodnocení aktivity v tomto akademickém roce, ke kterému se vztahuje toto hodnocení, musím konstatovat, že proběhla pouze jediná konzultace (v únoru), což je málo. | ||
| Aktivita při dokončování | Práce byla dokončována blízko termínu odevzdání. S finálním obsahem jsem tak neměl čas se seznámit. V závěru řešení práce by bylo vhodné mít ve větším předstihu další konzultace. | ||
| Publikační činnost, ocenění | Publikační činnost ani ocenění mi nejsou známy. |
The diploma thesis presents several issues, some of which are serious. Notably, it lacks an original solution, has missing citations in the theoretical section, and has unclear authorship of the source code. On the positive side, the creation of a dataset for training neural networks and the trained networks themselves stand out as significant achievements. The thesis has strong potential to be excellent if the formal issues are corrected, the student's contributions are clearly stated, and some original thought is added to the final solution (such as merging the skin and nudity detector or using a mask during training or prediction—there are many options). In its current state, I consider the thesis unsatisfactory.
| Kritérium | Známka | Body | Slovní hodnocení |
|---|---|---|---|
| Náročnost zadání | |||
| Rozsah splnění požadavků zadání | Points 3 and 5 of the assignment have not been fully accomplished. Point 3 required designing an original solution for nudity detection. However, the provided solution is a YOLO neural network without any modifications or improvements. Point 5 required a discussion with a representative of the Police of the Czech Republic, but it is unclear whether this discussion took place and what the outcome was. | ||
| Rozsah technické zprávy | |||
| Prezentační úroveň technické zprávy | 65 | The text is clear and understandable overall, but it lacks structure at some point. It is not clear what the purpose of skin detection is within the nudity detector. The topics of the skin detection and YOLO-based nudity detection interleaves in the text and it becomes confusing. Some figures in the text are unnecessary and could have been omitted, such as Figure 3.13, which represents someone else's skin detection solution or the figures explaining the basics in the neural network theory. There is an unnecessary detailed description of theories, such as Gabor filters, LBP, and wavelets, which are not utilised in the application itself. In the text, YOLOv7 is described and used as the final result, but YOLOv8, despite being used as an optional result and included in the final comparison table, is not described. | |
| Formální úprava technické zprávy | 50 | The report is written in decent English, which I appreciate. However, there are many shortcomings. For instance, the equations are presented in Figure 3.12 instead of being formatted as numbered text within the main content and contain an error. I do not like the equations inside Figure 2.3 describing activation functions either. The code snippets on pages 43, 55, and 57 could have been excluded or, at the very least, separated from the text, numbered, and titled. Many of the images are poor-quality raster images, which reduce the readability of the text (e.g., Figures 2.6, 3.3, 3.4, and several others). There are some minor issues, such as the percentage symbol not being separated from the number, and the distracting and confusing use of bold and italic text. | |
| Práce s literaturou | 40 | The amount and selection of references are acceptable, comprising mainly conference papers, some journal articles, online sources, and books. However, there are issues with citation etiquette. In particular, Chapter 2, from page 6 to page 10, lacks citations, despite discussing the theory of neural networks. This omission makes it unclear where the information originates from. | |
| Realizační výstup | 40 | The practical part of the thesis consists of three outputs, although only one tool was required. The main application should be the socially objectionable nudity detector. Additionally, there is an annotation tool for dataset creation and a skin detector included. The output of the skin detector is not used in the detector, and it is unclear how the police would utilise the skin mask or skin percentage in their work. The nudity detector (NudeDet) is a Python script that scans a directory of images for those potentially containing socially objectionable nudity. It uses a YOLO neural network to identify the inappropriate content. However, the provided solution does not bring any innovation or original approach; it merely uses the original YOLO as is to detect human body parts considered inappropriate. The annotation tool (SkeletonDet) uses YOLOv7 for pose estimation and assists with annotating the images used for training the nudity detector. The creation of the original dataset for training the networks was undoubtedly the most time consuming effort, and this should be emphasised as a positive outcome. The last output is a skin detector (SkinDet), which analyses an image for skin presence, calculates the percentage, and stores the result in a file. This output is not utilised for training, nor is it used to improve the YOLO output or in any other way. It simply detects skin, which may potentially indicate nudity, but it is a separate tool with no benefits for the final nudity detector as such unless integrated in the decision process. Without the support of another algorithm, the skin detection algorithm alone cannot accurately recognise socially objectionable nudity. I would have expected, for example, a merging of the neural network and the skin colour-based segmentation in the final application, or another original approach. The source code of NudeDet is derived from the YOLOv7 repository. Only 5 lines of code and a configuration file have been added. In contrast, SkeletonDet has undergone more modifications, with approximately 300 lines of code added. Interestingly, the most extensive work was done on SkinDet, although it does not contribute directly to the nudity detection functionality. A major issue with the NudeDet code is the lack of clear distinction between the student’s contributions and the original code. This makes it difficult for a reviewer to identify what the student has programmed. The repository used is also unclear, with the only reference being a vague statement: 'This is official YOLOv7 trained on custom data,' which is inadequate. In the case of YOLOv8, only the resulting weights are provided along with instructions on how to install and run it. Similarly, in SkeletonDet, there is no clear information about authorship since the original YOLO files are unsigned. SkinDet appears to be the student's work, although only the main file is signed. | |
| Využitelnost výsledků | The results of this thesis are intended for use by the Police of the Czech Republic for their internal purposes. While the current state of the project is partially usable, a graphical user interface (GUI) needs to be developed for practical application. |
eVSKP id 157162