How Multimodal AI Detects Pornographic Videos: Image & Audio Fusion Explained

Background

With the rise of mobile internet, short videos are a primary entertainment medium, but many contain pornographic content that harms youth and threatens social safety.

Technical Challenges

Porn video detection is a multimodal problem involving image and audio recognition. Image detection faces issues such as small pornographic regions and visual similarity between vulgar and pornographic images. Audio‑based detection lacks established theory.

Detection Techniques

Traditional image methods rely on handcrafted features like color histograms, which cannot distinguish vulgar from pornographic images. Deep learning approaches improve but have limited model complexity.

Audio classification typically converts wav to spectrograms and applies 2‑D convolutions; recent work uses log Mel‑spectrograms and a proposed RANet.

Proposed Framework

The framework combines image and audio modalities. It consists of three parts: a porn image recognition model, a porn audio recognition model, and a fusion of their results.

Porn Image Recognition Model

We propose DCNet, which includes classification and detection branches to capture global and local features. The detection branch uses BiFPN for weighted bidirectional feature fusion and an anchor‑free design with FCN‑style dense detection and a center‑point branch to reduce false positives.

Porn Audio Recognition Model

Audio is transformed into log Mel‑spectrograms (one per second) and processed with a TSN‑based architecture to capture temporal information. A frequency‑attention module, consisting of two convolutional layers inserted into a ResNet, extracts key sound features.

Fusion of Image and Audio Results

The outputs of the two models are merged to produce the final porn video classification.

Experimental Results

On a test set of 3,000 videos, the combined model achieved an accuracy of 93.4%.

The content is a translation of a 2021 paper published in the open‑access journal Applied Sciences.