In the world of product appearance inspection, the human eye can only identify defects larger than 0.5 mm in size with a significant risk of miscalculation. With longer working hours, eye fatigue and aging increase the risk of misjudgment. The traditional machine vision inspection is lacking in image processing and defect positioning, resulting in low accuracy and unstable performance of defect detection, and it is very difficult for the assembly line to achieve intelligent automatic detection. This shows that both traditional machine vision inspection and "naked eye" inspection of product appearance quality is very limited in terms of capability and efficiency, as well as accuracy and scope.

The idea of deep learning originates from "artificial neural network", which draws inspiration from the brain, simulates the human brain to analyze the problem mechanism and build a neural network for analysis and learning. The basic building blocks of neural networks are artificial neurons - neurons that mimic the human brain. Just as the brain's billions of neurons are distributed in several layers of a neural network with tens of thousands of connections between them, deep learning models involve a large number of computational units that interact with each other to learn autonomously about the multi-layered representations of the underlying distribution of the modeled data.