Lightweight High Precision Sensitive Image Recognition Method Based on Improved Runge Kutta and Feature Guidance Fusion

Abstract

The dissemination of sensitive content such as pornography and violence on social networks has adverse effects on the physical and mental well-being of young individuals, as well as on societal order. Accurately identifying and categorizing these sensitive images is crucial for upholding public safety. However, challenges such as varying lighting conditions, significant changes in sensitive areas, and background interference often impede the effectiveness of image recognition. Moreover, existing methods often neglect considerations of computational efficiency.To address these issues, this study proposes a novel approach named Improved Runge Kutta and Feature Guidance Fusion (IRFGN-Net) for sensitive image recognition. The IRFGN-Net aims to strike a balance between computational efficiency and classification accuracy. It introduces an enhanced Runge Kutta method to extract more detailed features, thereby improving the discrimination of key characteristics while minimizing the impact of noise and irrelevant background. Additionally, a novel feature guidance fusion module is introduced to integrate texture and depth features of sensitive images, effectively reducing the influence of illumination and noise on recognition accuracy.A comprehensive series of experiments was conducted to validate the effectiveness of the proposed method. Initially, ablation experiments were performed to demonstrate the efficacy of the improved Runge Kutta algorithm and feature guidance fusion module. Subsequently, IRFGN-Net was compared extensively with several widely used convolutional neural network models. The experimental results unequivocally demonstrate the superior performance of IRFGN-Net in sensitive image recognition tasks. Not only does it exhibit competitive inference speed, but it also achieves a significant enhancement in classification accuracy.

DOI: 10.61416/ceai.v27i1.9146