Transfer Detection of YOLO to Focus CNN’s Attention on Nude Regions for Adult Content Detection

Abstract

Video pornography and nudity detection aim to detect and classify people in videos into nude or normal for censorship purposes. Recent literature has demonstrated pornography detection utilising the convolutional neural network (CNN) to extract features directly from the whole frames and support vector machine (SVM) to classify the extracted features into two categories. However, existing methods were not able to detect the small-scale content of pornography and nudity in frames with diverse backgrounds. This limitation has led to a high false-negative rate (FNR) and misclassification of nude frames as normal ones. In order to address this matter, this paper explores the limitation of the existing convolutional-only approaches focusing the visual attention of CNN on the expected nude regions inside the frames to reduce the FNR. The You Only Look Once (YOLO) object detector was transferred to the pornography and nudity detection application to detect persons as regions of interest (ROIs), which were applied to CNN and SVM for nude/normal classification. Several experiments were conducted to compare the performance of various CNNs and classifiers using our proposed dataset. It was found that ResNet101 with random forest outperformed other models concerning the F1-score of 90.03% and accuracy of 87.75%. Furthermore, an ablation study was performed to demonstrate the impact of adding the YOLO before the CNN. YOLO–CNN was shown to outperform CNN-only in terms of accuracy, which was increased from 85.5% to 89.5%. Additionally, a new benchmark dataset with challenging content, including various human sizes and backgrounds, was proposed.

An Evaluation of Traditional and CNN-Based Feature Descriptors for Cartoon Pornography Detection

“…In the future, we intend to enhance this work further by considering the problem of small-scale porn regions inside the frames that cannot be addressed directly using CNN-only methods. One of the solutions to overcome small-scale porn regions was to utilize object detection model such as YOLO [49]. The author in previously mentioned work used COCO based pre-trained YOLO to detect human.…”

“…Various pre-trained CNNs such as AlexNet [42], VGG16 [43], GoogleNet [44], Inception3 [45], ResNet [46] have been used for feature extraction in the literature. The performance of previous CNN-only methods may drop sometime due to the complex background of visual images and the small-scale of regions need to be detected [47,48,49]. To address this matter, saliency-aware CNN [47] was found to predict the category and the position of object.…”

“…Additionally, global density fused convolutional network (GDF-Net), that can cascade global features, was proposed to facilitate the detection of a large number of small objects in congested scenes [48]. Furthermore, COCO based pre-trained YOLO was utilized to detect porn regions in any scale in the natural images [49]. This paper evaluates various methods including traditional feature and CNN based feature methods for cartoon pornography detection.…”

Pornographic content classification using deep-learning

“…In recent years, deep-learning algorithms have outperformed most classical feature algorithms that preceded them [7,17,24]. Deep learning algorithms require large sets of examples and ground truth outputs from which they automatically learn to extract a larger spectrum of complex and abstract features [2]. Due to the large number of publicly available examples, Convolutional Neural Network (CNN) have become the state of the art algorithms when dealing with adult pornography detection.…”

“…Whilst being very computationally efficient and robust against generating false positive alarms, hashing algorithms are limited to detecting content identical to that within the LEA's database. Recently, multiple deep-learning frameworks are being developed in an attempt to address this limitation [2,10,22]. However, such data driven solutions tend to suffer from high false positive rates or high negative rates when performing CSA detection since the sensitive nature of CSA material makes it difficult to collect large datasets and would normally be carried out in close cooperation with LEAs's.…”