Master thesis defense: Mauricio Perez and Javier Muñoz

Today the RECODists Mauricio Lisboa Perez and Javier Alvaro Vargas Muñoz have earned their master’s degree. Congratulations!

Mauricio presented the dissertation entitled “Video Pornography Detection through Deep Learning techniques and Motion Information” supervised by Prof. Anderson Rocha, RECOD, and co-supervised by Dr. Vanessa Testoni from Samsung Research Institute Brazil. Javier presented the dissertation entitle “A Soft Computing Approach for Learning to Aggregate Rankings” supervised by Prof. Ricardo Torres, RECOD.

The abstract of both dissertations are:

Video Pornography Detection through Deep Learning techniques and Motion Information

With the exponential growth of video footage available online, human manual moderation of sensitive scenes, e.g., pornography, violence and crowd, became infeasible, increasing the necessity for automated filtering. In this vein, a great number of works has explored the pornographic detection problem, using approaches ranging from skin and nudity detection, to local features and bag-of-visual-words. Yet, these techniques suffer from some ambiguous cases (e.g., beach scenes, wrestling), producing too much false positives. This is possibly related to the fact that these approaches are somewhat outdated, and that few authors have used the motion information present in videos, which could be crucial for the visual disambiguation of these cases. Setting forth to overcome these issues, in this work, we explore deep learning solutions to the problem of pornography detection in videos, taking into account both the static and the motion information available for each questioned video. When incorporating the static and motion complementary features, the proposed method outperforms the existing solutions in the literature. Although Deep Learning approaches, more specifically Convolutional Neural Networks (CNNs), have achieved striking results on other vision-related problems, such promising methods are still not sufficiently explored in pornography detection while incorporating motion information. We also propose novel ways for combining the static and the motion information using CNNs, that have not been explored in pornography detection, nor in other action recognition tasks before. More specifically, we explore two distinct sources of motion information herein: Optical Flow displacement fields, which have been traditionally used for video classification; and MPEG Motion Vectors. Although Motion Vectors have already been used for pornography detection tasks in the literature, in this work, we adapt them, by finding an appropriate visual representation, before feeding a convolution neural network for feature learning and extraction. Our experiments show that although the MPEG Motion Vectors technique has an inferior performance by itself, than when using its Optical Flow counterpart, it yields a similar performance when complementing the static information, with the advantage of being present, by construction, in the video while decoding the frames, avoiding the need for the more expensive Optical Flow calculations. Our best approach outperforms existing methods in the literature when considering different datasets. For the Pornography 800 dataset, it yields a classification accuracy of ~97.9%, an error reduction of 64.4% when compared to the state of the art (~94.1% in this dataset). Finally, considering the more challenging Pornography-2k dataset, our best method yields a classification accuracy of ~96.4%, reducing the classification error in 14.3% when compared to the state of the art (~95.8% in the same dataset).

A Soft Computing Approach for Learning to Aggregate Rankings

This work presents an approach to combine rank aggregation techniques using a soft computing technique — Genetic Programming — in order to improve the results in Information Retrieval tasks. Previous work shows that by combining rank aggregation techniques in an agglomerative way, it is possible to get better results than with individual methods. However, those approaches either combine only a small set of rank aggregation techniques or are performed in a completely ad-hoc way. In order to address these limitations, given a set of ranked lists and a set of rank aggregation techniques, we propose to use a supervised genetic programming approach to search combinations of them that maximize effectiveness in large search spaces. Experimental results conducted using seven datasets among different domains (text retrieval, content based image retrieval, multimodal retrieval) show that our proposed approach reaches top performance yielding superior results than state-of-the-art in learning-to-rank and in the supervised rank aggregation tasks. We also show that our proposed framework is efficient, flexible, and scalable.

This entry was posted in thesis defense and tagged , , , , , . Bookmark the permalink.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s