Advancing Surveillance Systems: Leveraging Sparse Auto Encoder for Enhanced Anomaly Detection in Image Data Security

In the realm of surveillance systems, ensuring robust anomaly detection capabilities is crucial for safeguarding against potential security breaches or hazardous incidents. This work uses the Sparse Autoencoder architecture to provide a unique method for improving anomaly identification in surveillance imagine information. The methodology begins with the collection of a comprehensive dataset, termed DCSASS, comprising videos from security cameras capturing a diverse range of abnormal and typical actions across various categories. Each video is meticulously labeled as normal or abnormal based on its content, facilitating the differentiation between routine operations and potential security threats. Individual frames extracted from the videos serve as image samples for subsequent model training and evaluation in anomaly detection tasks. In deep learning-driven anomaly identification networks, the preprocessing stage uses Min-Max Normalization to normalize pixel values, improving model stability and efficacy. Subsequently, the Sparse Autoencoder architecture, consisting of encoder and decoder components, is utilized for anomaly detection. The encoder, designed using convolutional neural network (CNN) architecture, extracts meaningful features from input images while inducing sparsity in learned representations through techniques like L1 regularization. Experimentation with various hyperparameters and architectural choices optimizes performance, with the decoder symmetrically mirroring the encoder's architecture to ensure accurate reconstruction of input images. The proposed approach outperforms existing methods such as CNN, RF and SVM, achieving 99% accuracy, making it 2.3% superior to other methods. Implemented in Python, our methodology demonstrates its efficacy in effectively capturing and reconstructing meaningful representations of input images, thereby enhancing anomaly detection capabilities in surveillance image data for improved security and safety measures.