Obscene image detection using transfer learning and feature fusion

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning excels at detecting explicit content but struggles with limited, labeled datasets, leading to overfitting and unstable classification.
• Existing methods face challenges due to data scarcity and the computational demands of training large models.

Purpose of the Study:

• To propose an automatic pornographic image detection method using transfer learning (TL) and feature fusion (FFP).
• To address overfitting, improve model performance, and reduce computational burden in explicit content detection.

Main Methods:

• Developed a TL-based feature fusion process (FFP) integrating low-level and mid-level features from pre-trained models.
• Generated a labeled dataset (GGOI) using Pix-2-Pix GAN and enhanced model architectures with batch normalization and mixed pooling.
• Selected high-performing models (MobileNet V2 + DenseNet169) for FFP and retrained the final layer for TL-based detection.

Main Results:

• The proposed TL model with fused MobileNet V2 + DenseNet169 achieved state-of-the-art performance.
• Achieved average classification accuracy of 98.50%, sensitivity of 98.46%, and F1 score of 98.49% on benchmark and generated datasets.

Conclusions:

• The TL-based feature fusion method effectively enhances pornographic image detection accuracy and stability.
• This approach offers a computationally efficient solution for explicit content detection, overcoming limitations of traditional deep learning models.