Logo image
Synthsecurenet : an improved deep learning architecture with application to intelligent violence detection
Thesis   Open access

Synthsecurenet : an improved deep learning architecture with application to intelligent violence detection

Ntandoyenkosi Welcome Zungu
M.Eng., University of Johannesburg
2025
Handle:
https://hdl.handle.net/10210/519322

Abstract

Deep learning (Machine learning) Human face recognition (Computer science) Video surveillance Human activity recognition Computer Vision
Violence is a pressing global issue, threatening public safety and demanding effective monitoring methods. Video surveillance, particularly remote systems, plays an important role in detecting and responding to violent incidents but faces challenges in terms of accuracy, speed, and resource demands. Recent advancements in deep learning have shown promise in enhancing surveillance capabilities, providing faster and more accurate detection. In this dissertation, a new deep learning architecture is presented, named SynthSecureNet, which addresses these limitations by hybridizing two popular architectures: MobileNetV2 and ResNet50V2. This study aims to harness the combined strengths of these models to improve violence detection accuracy while maintaining computational efficiency. SynthSecureNet leverages the pre-trained weights of MobileNetV2 and ResNet50V2, employing transfer learning to initialize the network. The architecture is then fine-tune using the Real-Life Violence Situations Dataset, consisting of 2000 labelled surveillance videos. The model’s performance is evaluated using metrics including accuracy, precision, recall, and F1 score, with particular focus on optimizing the fusion process between the two base architectures. Experimental results demonstrate significant improvements in violence detection performance. While MobileNetV2 achieves 90% accuracy and ResNet50V2 reaches 94%, SynthSecureNet attains a remarkable 99.22% accuracy. This substantial enhancement in performance is consistent across various dataset sizes, ranging from 400 to 2000 videos. The SynthSecureNet architecture offers a comprehensive solution that addresses the limitations of existing single-model approaches. By combining the efficiency of MobileNetV2 with the depth of ResNet50V2, this model provides a more robust and accurate tool for violence detection in surveillance footage. These findings pave the way for more effective surveillance and crime prevention strategies, potentially improving response times to violent incidents and enhancing public safety.
pdf
Zungu_NW_Masters_Dissertation1.70 MBDownloadView
Open Access

Metrics

2 File views/ downloads
1 Record Views

Details

Logo image