ANALYSIS: DEEP LEARNING
by using data augmentation, labelling, and ensuring their consistency, accuracy, and completeness. Moreover, this involves protecting and respecting the data and domain knowledge, such as by using encryption, anonymisation, and consent techniques, and ensuring their privacy, security, and integrity.
Trust and transparency enhancement and evaluation One of the possible solutions for overcoming the lack of trust and transparency of deep learning for vision inspection is to provide safe data management and explainable AI techniques. Neurocle, for example, has implemented these solutions in its software, which runs on a local cloud environment that ensures the data sovereignty and security of its customers and partners.
Deep learning can optimise identification and classification of defects in the auto and battery industries
appeal and durability, relies on precise paint inspection, which can be optimised using deep learning. Additionally, the assembly phase, involving components such as headlights, bumpers, mirrors, doors, and wheels, demands the accuracy and completeness that deep learning is capable of delivering. The quality of assembly directly influences the car’s functionality, performance, and, ultimately, customer satisfaction and loyalty.
Examples of deep-learning-based vision inspection in the battery industry As the battery industry is growing rapidly, the quality standards and expectations are also increasing. The customers and regulators demand higher-quality and safer batteries for electric vehicles and other devices. Therefore, the battery industry needs a better inspection technology that can detect defects and anomalies with high accuracy and reliability. Deep learning is a promising technology that can provide the required intelligent and efficient vision inspection solutions. It can improve the inspection of lithium-ion battery components, such as the cathode and anode for defects such as scratches, dents, tears and other imperfections.
Roadblocks and challenges hindering the widespread adoption of deep learning for vision inspection Lack of data and deep learning knowledge Deep learning requires a large amount of data and domain knowledge to train and
www.imveurope.com | @imveurope
optimise the vision inspection models, and to ensure their reliability and robustness. However, data and domain knowledge may not be readily available or accessible, especially for new or niche applications, or for proprietary or confidential information. Moreover, data and domain knowledge may be scattered across different sources and formats and may require preprocessing and annotation to make them suitable for deep learning.
Lack of trust and transparency Deep learning may not be trusted or understood by the management and operators of the vision inspection systems. This is because deep learning may not provide clear and explainable reasons for its inspection decisions and may not be able to handle unexpected or adversarial situations. Moreover, deep learning may not be aligned with the ethical and legal standards and expectations of the stakeholders and users, such as the privacy, security, and safety of the data and the products.
How can these roadblocks and challenges be addressed? Data and deep learning knowledge acquisition and management This involves collecting, organising, and managing the data and domain knowledge that are relevant and useful for deep learning for vision inspection, and ensuring their quality, quantity, diversity, and availability. This also involves preprocessing and annotating the data and domain knowledge, such as
Experience and insights from a deep learning software developer My firm, Neurocle, provides AI deep learning vision software for a diverse range of manufacturing industries, including automotive and battery. We have accumulated extensive experience and insights over several years of development and deployment. Our software, tailored for an extensive range of applications, enhances quality and performance, reduces costs, and boosts flexibility and scalability for our clients and partners. Traditional deep learning vision inspection systems often necessitate skilled engineers to perform numerous adjustments, normally up to 30 attempts to obtain a desirable model performance. Neurocle, however, has simplified this process with its proprietary auto deep learning algorithm, which finds the optimised model performance in a single attempt. This feature optimises hyperparameters, deep learning architecture, and data augmentation with a single click, making high-performance model creation accessible to anyone. Our vision is to democratise high-
performance deep learning model creation, allowing users to navigate the entire process effortlessly – from image input to labelling and training. This is done through user-friendly interfaces and features such as our auto-labelling and smart labelling tools. While recognising the complexity of deep
learning for vision inspection, Neurocle believes in its potential to benefit a wide range of industries. We emphasise the collaborative effort required from users and technology providers alike. Our mission is to provide exceptional AI deep learning vision software and services, contributing to the industry’s progress and adoption of this transformative technology. I
DECEMBER 2023/JANUARY 2024 IMAGING AND MACHINE VISION EUROPE 11
Shutterstock/PP77LSK
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32
