Design of deep learning accelerated algorithm for online recognition of industrial products defects

• Published in: Neural computing & applications Vol. 31; no. 9; pp. 4527 - 4540
• Main Authors: Shu, Yufeng, Huang, Yu, Li, Bin
• Format: Journal Article
• Language: English
• Published: London Springer London 01.09.2019; Springer Nature B.V
• Subjects: Algorithms; Artificial Intelligence; Artificial neural networks; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Convolution; Data Mining and Knowledge Discovery; Deep learning; Defects; Feature extraction; Heuristic methods; Image Processing and Computer Vision; Mathematical models; Neural networks; Object recognition; Probability and Statistics in Computer Science; Process parameters; S.I. : Emergence in Human-like Intelligence towards Cyber-Physical Systems; Space storage; Training; Deep learning; Products defects; Accelerated algorithm; Online recognition
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-018-3511-4

With the defects of LED chip as the research object, in LED chip defect recognition, an efficient and scalable parallel algorithm is critical to the deep model using a large data set training. As the implementation of parallel in multi-machine is relatively difficult, the paper puts forward a recognition method based on convolution neural network algorithm, so as to improve the small batch stochastic gradient descent algorithm that is very popular in the industry. This method has overcome the shortcomings of the existing defects recognition algorithms: It requires to manually extract features and needs a heuristic method. Furthermore, this method has made the following improvements to traditional methods: (1) It adds a “copy of model parameters” of critical resources to reduce the waiting time of a GPU in requesting for model parameters in the process of model parameters updating; (2) distribution mechanism of small batch data is designed, which, training set is some small batch data that are randomly distributed at GPU side, and is selected by using the critical resource p variable; (3) a certain amount of memory space storage gradients are created on the GPU side and the parameter server side, and the gradient propagation is passed through a “gradient distribution” thread for scheduling. The experimental data show that the recognition rate of network for short-shot defects has reached 99.4%. In addition, compared with BP neural network, it can be seen from the experiment that the recognition rate of the proposed method is significantly better than that of BP neural network, so the method has a good application prospect.