Anomaly Detection in Fast Moving Consumer Goods using Deep Learning

The project presents a deep learning-based system for anomaly detection in the fastmoving consumer goods (FMCG) sector, where speed and scale make quality assurance highly challenging. Traditional inspection methods often fail to identify subtle or unpredictable defects such as mislabeling, surface cracks, or irregular packaging, leading to product recalls and financial loss. To address these challenges, a Stochastic Irregularity Vision (SIV) system is proposed, which integrates computer vision and advanced deep learning techniques including convolutional neural networks (CNNs), autoencoders, and pretrained feature extractors. The system leverages the MVTec AD dataset and Anomalib library to learn representations of defect-free products, enabling unsupervised anomaly detection without reliance on large labeled defect datasets. During inference, anomalies are highlighted through heatmaps, segmentation masks, and confidence scores, ensuring precise localization and rapid decision-making on production lines. To achieve real-time deployment, the solution is optimized using Intel’s OpenVINO toolkit, which reduces inference latency and facilitates scalability across edge devices. The system’s modular architecture supports adaptability across diverse FMCG product categories, minimizing retraining requirements while maintaining consistent performance. Results demonstrate strong accuracy in detecting both obvious and subtle anomalies, with robustness against variations in texture, labeling, and packaging. Key advantages include reduced manual inspection, cost-effectiveness, and high scalability for Industry 4.0 environments. Future work will focus on active learning integration, cross-domain generalization, and IoT-enabled predictive maintenance to further enhance reliability and efficiency. Overall, the proposed framework provides a practical, intelligent, and scalable solution for real-time automated quality control in FMCG manufacturing.