• Vaishnavi Diliprao Deshmukh
• Dr. Chandrakant P. Navdeti

This research presents an innovative framework for criminal anomaly detection in surveillance footage, leveraging a custom dataset and a hybrid deep learning architecture. The custom dataset, comprising normal and anomalous images extracted from video frames, simulates real-world surveillance scenarios with clear, stable scenes and degraded or suspicious activities (e.g., blur, noise, or unusual behavior). Images are preprocessed to a uniform 224x224 resolution, enabling robust model training. The proposed system integrates a Generative Adversarial Network (GAN) for restoring degraded visuals and a MobileNetV2-based feature extractor combined with a custom convolutional neural network (CNN) for semantic classification. The GAN’s generator reconstructs high-quality features from anomalous inputs, while the discriminator distinguishes between normal and generated features. MobileNetV2, fine-tuned for efficiency, extracts 1280- dimensional semantic features, concatenated with 1024-dimensional CNN outputs to form robust 2304-dimensional vectors for classification. A Single Shot Multi Box Detector (SSD) enhances real-time object detection, enabling dynamic anomaly identification in video streams. The training pipeline employs an 80-20 train-test split, achieving ~96% accuracy with minimal overfitting through regularization. Features are visualized via PCA for clustering, confirming clear separation between normal and anomalous patterns.

"Criminal Anomaly Detection Using GAN with MobileNetV2", International Journal of Emerging Technologies and Innovative Research (www.jetir.org), ISSN:2349-5162, Vol.12, Issue 12, page no.f196-f221, December-2025, Available :http://www.jetir.org/papers/JETIR2512525.pdf

"Criminal Anomaly Detection Using GAN with MobileNetV2", International Journal of Emerging Technologies and Innovative Research (www.jetir.org | UGC and issn Approved), ISSN:2349-5162, Vol.12, Issue 12, page no. ppf196-f221, December-2025, Available at : http://www.jetir.org/papers/JETIR2512525.pdf