File(s) not publicly available
Reason: Linked to the publisher's website.
On deploying mobile deep learning to segment COVID-19 PCR test tube images
Effective detection of the COVID-19 pandemic is essential for timely disease treatment and prevention. This work studies compact deep-learning models executed on mobile devices for segmenting COVID-19 RT-PCR test tube images, a crucial image-processing step preceding higher-level tasks. Since the device resource constraints and the need for rapid results necessitate compact and streamlined models with reasonable accuracy, we employ the hyperparameter width multiplierα to the trainable components in the two deep learning models based on the U-Net architecture, including MobileNetV2 and Xception. Our new compact models, calledα-MobileNetV2 andα-Xception, facilitate the progressive simplification of the U-Net model structures, maintaining high accuracy. By varying the width multiplierα, we explore diverse training conditions for the models, analyzing the model size and its performance. The final model achieves a3.2× reduction in size and3× faster inference, with merely a 1.2% loss in accuracy compared to standard MobileNetV2 on segmenting COVID-19 PCR test tube images.