An Indian manufacturing client needed to replace manual quality inspectors who were catching only 70% of defects. We designed and deployed a computer vision system using custom-trained CNNs that inspects products on the assembly line in real-time at 30fps, dramatically improving defect detection and reducing costs.
“Inconsistency was our biggest challenge. Two inspectors looking at the same product would give different verdicts. We needed a system that could enforce a single, objective standard around the clock without fatigue or bias.”
Quality Director, Indian Manufacturing ClientWe worked closely with the client's quality team to catalogue all 12 defect types with defined severity levels. The team photographed over 15,000 sample images across all defect categories, creating a comprehensive ground-truth dataset that formed the foundation for model training.
We set up high-resolution camera stations on all 4 production lines with controlled LED lighting for consistent image capture. To address class imbalance in rare defect types, we used synthetic data augmentation techniques including rotation, flipping, colour jittering, and GAN-based generation to balance defect classes and improve model robustness.
We trained a custom ResNet-based architecture optimised for fine-grained defect classification. The model was further optimised for edge deployment using NVIDIA TensorRT, achieving real-time inference at 30 frames per second while maintaining 95% detection accuracy across all 12 defect categories.
We deployed the optimised model on NVIDIA Jetson edge devices at each production line. The system was integrated directly with existing PLC (Programmable Logic Controller) systems for automated rejection of defective products, eliminating any manual intervention in the pass/fail decision loop.
Book a free AI readiness call. We'll discuss your challenges and outline a clear path forward.