Deep Active Learning for Pornography Recognition Using ResNet
| Autor: | Sui Lyn Hor, Nouar AlDahoul, Hezerul Abdul Karim, Mohd Haris Lye, Sarina Mansor, Mohammad Faizal Ahmad Fauzi, Abdulaziz Saleh Ba Wazir |
|---|---|
| Jazyk: | angličtina |
| Rok vydání: | 2022 |
| Předmět: | |
| Zdroj: | International Journal of Technology, Vol 13, Iss 6, Pp 1261-1270 (2022) |
| Druh dokumentu: | article |
| ISSN: | 2086-9614 2087-2100 |
| DOI: | 10.14716/ijtech.v13i6.5842 |
| Popis: | The demand for nudity and pornographic content detection is increasing due to the prevalence of media products containing sexually explicit content with Internet being the main source. Recent literature has proved the effectiveness of deep learning techniques for adult image and video detection. However, the requirement for a huge dataset with labeled examples poses a restriction in practical use. Several research has shown that training deep models using an active learning framework could reduce the annotation effort, but this approach has yet to be applied for pornography detection. In this paper, the classification efficiency and annotation requirement of fine-tuned ResNet50V2 model trained using an active learning framework in pornographic image recognition was explored by comparing the method’s performance using three sampling strategies (random sampling, least confidence sampling, and entropy sampling). The baseline for comparison was a fully supervised learning method. The video frames of the public NPDI dataset were utilized to run a 5-fold cross-validation. The results of the experiments demonstrated that similar average test accuracy of five folds could be obtained using the deep active learning method, with only 60% of labeled samples in the training dataset compared to 100% annotated samples in fully supervised learning. |
| Databáze: | Directory of Open Access Journals |
| Externí odkaz: |
|