|
CSME 2023/08
Volume 44 No.4
:
379-388
|
|
Processing-in-Memory (PIM) Based Defect Prediction of Metal Surfaces Using Spiking Neural Network
Mohammed Siyad Ba and R Mohan a
aDept. of Computer Science and Engineering, National Institute of Technology, Tiruchirappalli
|
Abstract:
Metal industries have long used the benefits of computer vision to automate various applications in product development, testing, and mistake repair. Most computer vision models for defect prediction use neural networks for accurate and exact defect classification. Reloading weights, saving, and retrieving activations are all impacted by high data processing layers, which results in restricted network bandwidth and latency. To solve this issue, a novel Processing-in-Memory (PIM) based defect prediction using Spiking Neural Network (SNN) has been proposed, in which the weight values in metal surface image processing are given to SNN that carries data only when a specific threshold is reached by LIF neurons, thereby decreasing the processing latency. The emerging non-volatile memory technologies like memristors have been shown to follow biological neurons and synapses which is irreplaceable to the processing of memory concepts in neuromorphic computing. To eliminate the sneak path problem, a novel approach has been implemented which utilizes two memristors and one transistor and reduces the on-chip memory overhead by changing the modes of transistors. Experimental results show an accuracy of 98% and an F-score of 96.5 which outperforms most of the computer vision methods taken for comparison.
|
|
Keywords: Defect prediction, Spiking Neural Network, Processing in Memory, Memristor crossbar array, Neuromorphic Computing.
|
Download PDF
|
|
*Corresponding author; e-mail: 406320001@nitt.edu
|
©
2023
CSME , ISSN 0257-9731
|