A Deep Learning Approach for Noisy Image Classification in Automated Driver Drowsiness Detection

Abstract

In recent years, driver drowsiness has been a major cause of road accidents, particularly when the driver has been driving on the highway for an extended period of time. Smart systems can now be used to prevent accidents, and a reliable driver detection system must be applied to alert the driver. In these systems, several external factors have been degrading the performance of these systems, including added noise, interference and low illumination. To overcome these limitations, this paper presents a de-noising approach for noisy images; the results show that the enhanced images improve the overall system performance and classification accuracy. The final validation accuracy is 97.5%, while the testing accuracy for S1 is 96%, S2 is 92%, and S3 is 91%. The test accuracy of S1 decreased to 45% when the Salt and pepper noise is added to the set , , when Gaussian noise is added to S2 the testing accuracy decreased to 85%, and when speckle noise is added to S3 the testing accuracy is reduced to 73%. When the median filter is used the testing accuracy for S1 become 93%, the testing accuracy for S2 increase to 91%, and the testing accuracy for S3 raises to 85%.

Automatic modulation classification based deep learning with mixed feature

“…In the future, we are thinking of getting the same results by forming a network that is smaller in size than the current network. also introduced different augmented and optimization techniques as working techniques in papers [24], [25].…”