PaperlayerView exact revision on arXiv

Source-linked AI summary

Combating the Elsagate phenomenon: Deep learning architectures for disturbing cartoons

Akari Ishikawa, Edson Bollis, Sandra Avila

arXiv:1904.08910v1cs.CV

TL;DR

Research specifically addressing disturbing content in cartoons such as Elsagate is lacking, despite the challenge of distinguishing it from non-sensitive cartoons. The paper proposes a deep-learning video-detection pipeline using static and motion information. Its final reported result is 92.6% accuracy, with NASNet transferred from ImageNet and finetuned to Elsagate as the final model.

  • Problem

    Research specifically addressing Elsagate is lacking, while distinguishing it from non-sensitive cartoons is challenging.

    Source: §I · p. 1, §II · p. 2

  • Method

    The paper adapts deep-learning video detection using static frames and motion information, evaluating architectures and fusing the two sources.

    Source: §I · p. 1, §III · p. 2, 2) Motion Information · p. 2

  • Results

    92.6% accuracy was achieved, and NASNet transferred from ImageNet and finetuned to Elsagate was selected as the final model.

    Source: C. Testing · p. 5, §VI · p. 6

  • Takeaways & Limitations

    The paper presents a first solution for disturbing-cartoon detection and brings discussion of the Elsagate phenomenon to the literature.

    Source: §VI · p. 6

    The authors identify the lack of a formal definition of Elsagate as a challenge to the problem.

    Source: §VI · p. 6

Abstract

from arXiv · show

Watching cartoons can be useful for children's intellectual, social and emotional development. However, the most popular video sharing platform today provides many videos with Elsagate content. Elsagate is a phenomenon that depicts childhood characters in disturbing circumstances (e.g., gore, toilet humor, drinking urine, stealing). Even with this threat easily available for children, there is no work in the literature addressing the problem. As the first to explore disturbing content in cartoons, we proceed from the most recent pornography detection literature applying deep convolutional neural networks combined with static and motion information of the video. Our solution is compatible with mobile platforms and achieved 92.6% of accuracy. Our goal is not only to introduce the first solution but also to bring up the discussion around Elsagate.

I. INTRODUCTION

Elsagate videos place familiar childhood characters in disturbing scenes, yet the paper identifies little research focused on this phenomenon and emphasizes the difficulty of distinguishing such videos from non-sensitive cartoons. It proposes a deep-learning detection pipeline and introduces a cartoon-only dataset.

I. INTRODUCTION

  • Elsagate depicts popular childhood characters in disturbing scenes, including violence, sexual situations, theft, and toilet-related content.Source: §I · p. 1
  • The authors identify a research gap: Elsagate classification is challenging because disturbing cartoons can resemble non-sensitive cartoons, unlike the focus of much sensitive-content research on real-life videos.Source: §I · p. 1, §II · p. 2
  • The paper proposes an end-to-end deep-neural-network pipeline for Elsagate detection in videos, evaluates several mobile-platform DNNs, and introduces a 285-hour dataset with 1,396 Elsagate and 1,898 non-sensitive videos.Source: §I · p. 1

II. RELATED WORK

  • Alghowinem’s multimodal YouTube Kids pilot analyzed three videos, none of which were cartoons.It combined image, audio, and transcribed-text features using individual classifiers and threshold-based decisions.Source: §II · p. 1
  • Earlier cartoon-violence studies used color features or a keyframe-and-SIFT pipeline, and their datasets were not publicly available.Kahlil et al. also noted that color features lacked sufficient contextual information for content analysis.Source: §II · p. 2
  • Papadamou et al. reported 82.8% accuracy using video metadata and statistics, but did not analyze video content such as frames.The authors note that uploaders could disguise the analyzed metadata features; that dataset was not publicly available.Source: §II · p. 2

III. METHODOLOGY

The paper adapts a flexible deep-learning video-detection pipeline for Elsagate, combining static and motion information with a design aimed at mobile-platform efficiency.

III. METHODOLOGY

  • The method adapts a flexible deep-learning video-detection pipeline to Elsagate, allowing changes across its stages.The overview presents the proposed Elsagate detection method.Source: §III · p. 2, Figure 1 · p. 2
  • For efficiency and a mobile-platform target, the authors combine raw frames with computationally cheap MPEG motion vectors rather than optical flow.Source: §III · p. 2
  • The methodology evaluates transfer from related versus unrelated datasets, compares mobile-oriented architectures, and adopts a fusion strategy.Source: §III · p. 2

A. Features

The pipeline represents video with sampled raw frames and MPEG motion vectors, pairing visual appearance with motion information while targeting computational efficiency.

1) Static Information:

  • Static information comes from raw video frames sampled at 1 fps, centered, resized to 224×224 pixels, and cropped while maintaining aspect ratio.Source: 1) Static Information · p. 2

2) Motion Information:

  • Motion information comes from MPEG motion vectors, which encode macroblock movement between reference and current frames and can be decoded at low computational cost.The vectors are translation offsets from the reference frame to the target frame; macroblocks are usually 8 or 16 pixels in each dimension.Source: 2) Motion Information · p. 2

B. Deep Learning Architectures

The paper evaluates four deep-learning architectures for Elsagate classification, including GoogLeNet, SqueezeNet, MobileNetV2, and NASNet. It describes GoogLeNet’s inception design as a departure from simply stacking deeper convolution layers.

B. Deep Learning Architectures

1) GoogLeNet:

The described architectures reduce model or data-flow costs through compact parameterization, bottleneck blocks, and depthwise separable convolutions.

1) GoogLeNet:

  • SqueezeNet uses 1.2 million parameters and a fire module that squeezes feature-map depth before expanding it with 1×1 and 3×3 filters.It contains eight fire modules.Source: 1) GoogLeNet · p. 3

2) SqueezeNet:

  • The evaluated SqueezeNet setup adds a two-class fully connected layer and uses its 1000-neuron layer as the feature vector because its other layers were unsuitable.The cited passage attributes this to layer size or sparsity.Source: 2) SqueezeNet · p. 3
  • MobileNetV2 uses depthwise separable convolutions and bottleneck blocks to reduce computational cost and the amount of data flowing through the network.Depthwise convolution processes each input channel independently, omitting convolution in the channel domain.Source: 2) SqueezeNet · p. 3
  • MobileNetV2 contains an initial 32-filter fully convolutional layer followed by 19 residual bottleneck layers and uses 2.3 million parameters.Source: 2) SqueezeNet · p. 3

3) MobileNetV2:

The methods describe NASNet's cell-based construction, probability-mean late fusion, and a two-fold cross-validation setup for model evaluation.

3) MobileNetV2:

  • The MobileNetV2 subsection begins by describing the architecture as developed for mobile platforms.Source: 3) MobileNetV2 · p. 3

4) NASNet:

  • NASNet builds a network by optimizing a cell and stacking copies of the best cell; the evaluated NASNet-A uses four repeats, 1056 penultimate-layer filters, and 5.3 million parameters.Source: 4) NASNet · p. 3

C. Late Fusion

  • Late fusion averages the independent classification probabilities from static and motion information to produce the final classification score.Source: C. Late Fusion · p. 3

IV. EXPERIMENTAL SETUP

  • The paper states that dataset, code, and models are available through its project repository.Source: §IV · p. 3
  • Evaluation randomly splits the dataset into two folds, then switches training and testing sets to conduct two analyses per model.Source: §IV · p. 3

A. Elsagate Dataset

The authors introduce a publicly available, cartoon-only Elsagate dataset with separate training/validation and test collections.

A. Elsagate Dataset

  • The dataset contains 285 hours of cartoons: 1,396 Elsagate videos and 1,898 non-sensitive videos.The passage compares its duration with the 140-hour Pornography-2k dataset.Source: A. Elsagate Dataset · p. 3
  • Elsagate videos were gathered from YouTube channels reported in a Reddit thread, while non-sensitive videos came from official channels such as Cartoon Network and Disney Channel.Source: A. Elsagate Dataset · p. 3
  • Collection in February 2018 formed the training/validation set, and a September 2018 collection formed the test set, totaling 1,898 non-sensitive and 1,396 Elsagate videos.The interval was intended to evaluate the model on new videos that could appear; the classes were imbalanced to offer a representative dataset.Source: A. Elsagate Dataset · pp. 3–4
  • The dataset is free to the scientific community, but access requires a formal request and a signed responsibility term because of potential legal liabilities.Source: A. Elsagate Dataset · p. 4

B. Evaluation Metrics

The evaluation reports normalized accuracy and F2-measure, with F2 emphasizing recall because missed sensitive content is treated as especially harmful.

V. RESULTS AND DISCUSSION

B. Evaluation Metrics

  • Normalized accuracy is defined as the average of true positive rate and true negative rate, measuring success independently of class labels.Source: B. Evaluation Metrics · p. 4
  • F2-measure weights recall twice as heavily as precision, reflecting the authors' view that false negatives expose sensitive material and are less acceptable than false positives.Source: B. Evaluation Metrics · p. 4

A. Transfer Learning

Transfer learning favored ImageNet initialization, while mobile-architecture tests showed different leaders before and after finetuning. NASNet achieved the strongest reported finetuned results, whereas SqueezeNet led in transfer learning.

A. Transfer Learning

  • The transfer-learning comparison used GoogLeNet as a feature extractor and trained SVM classifiers on the Elsagate data.Source: A. Transfer Learning · p. 4
  • GoogLeNet pre-trained on ImageNet achieved better accuracy than GoogLeNet pre-trained on pornography videos, so later experiments used ImageNet pre-training.Source: A. Transfer Learning · p. 4, Table I · p. 4

B. Mobile Deep Learning Architectures

  • SqueezeNet outperformed NASNet and MobileNetV2 in transfer learning, but preliminary use of other SqueezeNet feature layers performed poorly.The authors say the architecture modifications produced this unexpected outcome and caution that SqueezeNet is not an adequate feature extractor for Elsagate content.Source: B. Mobile Deep Learning Architectures · p. 4, Figure 4 · p. 5
  • Finetuning initialized the networks with transferred weights and retrained all layers; experiments freezing initial layers did not report good results.Source: B. Mobile Deep Learning Architectures · p. 4
  • 95.3% accuracy and 92.9% F2 were reached by finetuned NASNet, surpassing SqueezeNet and MobileNetV2 and nearing GoogLeNet.The figure compares accuracy and F2 across transfer learning and finetuning.Source: B. Mobile Deep Learning Architectures · p. 4, Figure 4 · p. 5
  • SqueezeNet’s results declined across frames, motions, and late fusion, while MobileNetV2’s frame results also declined after finetuning.The authors hypothesize that models already performing fairly well may hit a “high accuracy barrier” and unlearn the task.Source: B. Mobile Deep Learning Architectures · pp. 4–5

C. Testing

A test set downloaded about seven months after training showed a decline in SqueezeNet performance, leading the authors to select ImageNet-transferred, finetuned NASNet. Their error review found many misclassifications among nursery and counting rhymes, alongside other ambiguous cases.

C. Testing

  • After testing on videos downloaded about seven months later, the authors chose NASNet transferred from ImageNet and finetuned to Elsagate as their final model.The results showed an impressive decline in SqueezeNet performance.Source: C. Testing · p. 5, Table II · p. 5

D. Model Interpretation

  • The authors examined false negatives and false positives because accuracy and F2 alone do not explain model behavior or the reasons for errors.Source: D. Model Interpretation · p. 5
  • Most misclassified videos were nursery or counting rhymes, which appeared in both classes; non-sensitive examples were higher quality and considered less harmful on average.Source: D. Model Interpretation · pp. 5–6, Figure 5 · p. 6
  • Other false positives and negatives included familiar Elsagate characters outside grotesque situations, Elsagate-like color or motion, and non-sensitive videos with controversial scenes.Source: D. Model Interpretation · p. 6, Figure 5 · p. 6

VI. CONCLUSION

The paper presents a first solution for Elsagate detection using deep-learning architectures and static and motion video information. Experiments identified ImageNet-transferred, finetuned NASNet as an effective approach, with 92.6% accuracy.

VI. CONCLUSION

  • 92.6% accuracy was achieved despite challenges including the lack of a formal definition of Elsagate.Source: §VI · p. 6
  • The authors report NASNet as an excellent feature extractor when transferred from ImageNet and then finetuned to Elsagate.Source: §VI · p. 6
  • The work introduces a first Elsagate detection solution, evaluates popular deep-learning architectures, and uses both static and motion video information.Future work is intended to embed the solution in a mobile application and deepen annotation of the phenomenon.Source: §VI · p. 6
Loading 1904.08910v1…