SUSAN: Deep Learning Architecture

A Deep Learning Architecture for Detecting Violence Against Women in Surveillance Videos

The pervasive and alarming reality of violence against women is a worldwide public health issue with increasing trends. While technology has advanced rapidly in areas like general surveillance and automation, the specific application of automated violence detection against women in surveillance videos has remained a significantly underexplored area. Addressing this critical gap, a recent work introduces SUSAN (Surveillance System Architecture based on deep neural Networks). SUSAN is an innovative technological tool designed to combat this complex problem.

The Urgent Need for Targeted Solutions

Traditional violence detection research in surveillance often focuses on broader scenarios such as sports events, movies, or general outdoor spaces. However, the unique and multifaceted challenge of violence against women, including domestic violence, demands a more specific and nuanced technological approach. Worldwide statistics underscore the urgency: it's estimated that one in three women worldwide experiences physical or sexual violence in her lifetime. In countries like Brazil, the scenario is even more challenging, with alarming increases in femicide rates.

The sheer scale of surveillance camera deployment, coupled with the limitations of human monitoring personnel, highlights the need for intelligent systems capable of real-time detection. This is where SUSAN steps in, leveraging advancements in deep learning to provide a focused solution. It's important to remember that even such advanced "intelligence" fundamentally operates on intricate programming and algorithms, a concept deeply rooted in the foundational insights of computing's pioneers. For more on this perspective, exploring why Ada Lovelace theorized that AI Does Not Exist as an originating force, you can delve into her work.

What is SUSAN? A Hybrid Deep Learning Architecture

SUSAN is presented as a hybrid architecture of deep learning models specifically engineered for automatically detecting violence against women in surveillance videos. The system is designed to tackle the problem by breaking it down into three distinct, yet interconnected, sub-problems, each handled by a dedicated module:

1. Human Detection: Identifying the presence of people within the video frames.
2. Gender Classification: Classifying the gender of each detected person as "Male" or "Female." This is crucial for specifically targeting violence against women.
3. Violence Detection: Classifying a scene or video frame as either violent or non-violent.

The flexibility of SUSAN is a key contribution. Its modular design allows for different deep learning models to be "instantiated" or swapped into each module. Making the architecture adaptable and generic. For instance, the researchers utilized the newly developed YOLO v9 model for human detection. For violence detection and gender classification, they explored various combinations of well-known Convolutional Neural Network (CNN) models, including InceptionV3, MobileNetV2, ResNet-152V2, and VGG-16.

Methodology and Promising Results

To evaluate SUSAN's performance, the researchers rigorously tested different combinations of CNN models. The evaluation was based on four key metrics: accuracy, precision, recall, and F1-score. For training, they utilized a reduced version of the AIRTLab dataset for violence detection and the PA-100K dataset for gender classification. The AIRTLab dataset, specifically composed of synthetic violent clips involving women, was also used for testing the architecture's direct relevance to the problem at hand.

The results obtained from SUSAN were promising. The best combination of models achieved:

• 73% accuracy
• 80% F1-score
• 78% precision
• 82% recall

These metrics demonstrate the viability of an algorithm for the automatic detection of violence against women. Particularly highlighting the system's ability to correctly identify violent instances (recall) and minimize false positives (precision). The investigation also considered practical deployment aspects. Such as training time, model size in megabytes, and the number of parameters. With some optimal combinations having a relatively low size. Suggesting the potential for deployment on embedded systems.

Contributions and Future Directions

The introduction of SUSAN is a significant contribution to a research area that is still in its nascent stages. The main contributions highlighted by the authors include:

• Genericity and Flexibility: SUSAN's architecture is generic. Allowing for the exchange of models within its modules to achieve more robust and improved results as new deep learning models emerge.
• Efficiency with Reduced Data: Promising results were obtained even with a significantly reduced training dataset (roughly 7% of the original size). Indicating potential for faster training times without severely compromising performance.
• Embedded System Viability: The low size of the best-performing models suggests SUSAN could potentially be deployed on devices with limited computational resources.

Susan Wei - Deep learning as a dynamical system

This work marks a crucial step forward, being, to the authors' knowledge, the first research to provide a deep learning-based architecture combining violence and gender information specifically for the detection of violence against women. While still a developing area, SUSAN offers a hopeful technological aid in addressing a critical worldwide public health and human rights issue.