Partha Deka, is a seasoned Data Science Leader with over 15 years of experience in the semiconductor supply chain and manufacturing sectors. As a Senior Staff Engineer at Intel, he has led teams in developing AI and machine learning solutions that drive cost savings and process optimizations. Notably, he spearheaded the creation of a computer vision system that improved logistics efficiency, earning recognition as a CSCMP Innovation Award finalist.
A Senior IEEE Member, published author, and frequent speaker at events like NeurIPS, Partha also reviews submissions for NeurIPS, contributing to advancements in AI and machine learning, particularly in semiconductor manufacturing.
In this article Partha and his co-author Rohit Mittal, writes about Quality Inspection in Manufacturing Using Deep Learning-Based Computer Vision.
Automation in industrial manufacturing is on the rise, and material quality inspection is increasingly being automated with minimal human intervention. The goal is to achieve human-level or greater accuracy in quality inspection while maintaining production efficiency. This article explores the use of deep learning to automate quality inspection, optimizing both the algorithms and the hardware to achieve optimal performance.
Deep Learning for Quality Inspection TraditionalTraditional quality inspections in manufacturing involve manual inspection after production, which is time-consuming and inefficient. Deep learning-based computer vision offers the opportunity to automate quality checks during production, reducing human intervention and improving overall factory efficiency. Deep learning has proven to be highly effective in areas like object detection and medical imaging, achieving human-level accuracy.
Deep learning involves training neural networks with multiple layers (input, hidden, and output) to learn patterns from vast amounts of data. Convolutional Neural Networks (CNNs) are especially useful for image recognition tasks, as they can identify patterns like edges and textures in an image and then use these to recognize complex objects.
In manufacturing, detecting defective materials (e.g., scratches or cracks) manually is prone to errors and consumes valuable time. To automate defect detection, we developed a deep learning system that identifies defects within milliseconds and provides human-level accuracy. This system uses CNNs to identify defects and highlights the defect areas with heat maps for interpretability.
We initially adopted a hybrid approach using traditional computer vision methods to extract regions of interest (ROIs) before applying deep learning to detect defects. By focusing the deep learning models only on the areas of interest, we improved accuracy.
To prevent overfitting, we augmented our dataset by creating multiple variations of each image through random crops and rotations. We also used pre-trained CNN models like ResNet50 and InceptionNet, and applied optimizations like ImageNet initialization and cross-entropy loss to improve model performance.
In a second approach, we replaced traditional computer vision methods with a deep learning-based ROI prediction system. By labeling images with bounding boxes, we trained a model to predict the ROI, reducing the need for code changes when product types or camera setups changed.
Our models achieved high accuracy, with improved metrics such as an Intersection over Union (IOU) score of 0.8577 and a significant reduction in false negatives. We optimized our models for inference using Intel’s OpenVINO toolkit, which accelerated performance across CPUs, GPUs, and other hardware like Intel Movidius and FPGAs.
To increase the interpretability of our CNN models, we applied heat map visualizations to highlight which areas of an image the model focused on when detecting defects. This technique helps users understand why certain areas are classified as defective, enhancing trust in the system.
By applying deep learning-based computer vision, we were able to automate material quality inspection with human-level accuracy, reduce human intervention, and optimize manufacturing processes. Our solution not only detects defects but also interprets the defect areas, providing a comprehensive view of the quality inspection process.
https://www.coursera.org/learn/convolutional-neural-networks
https://arxiv.org/abs/1512.03385
http://cs231n.github.io/convolutional-networks/
https://arxiv.org/pdf/1512.04150.pdf
.svg)
.svg){kind=small}
.svg)
