The Effect of LAB Color Space with NASNetMobile Fine-tuning on Model Performance for Crowd Detection

Abstract

In the COVID-19 pandemic, computer vision plays a crucial role in crowd detection, supporting crowd restriction policies to mitigate virus spread. This research focuses on analyzing the impact of using the RGB LAB color space in advance to perform NASNetMobile better. The fine-tuning process, involving freezing layers in various NASNetMobile base model variations, is considered. Results reveal that the model with LAB color space outperforms the model with RGB color space, with an average accuracy of 94.68% compared to 94.15%. From all the test iterations, it was found that the highest performance for the NASNetMobile model occurred when freezing 10% of the layers from the back for both model LAB and RGB color spaces, with the LAB color space achieving an accuracy of 95.4% and the RGB color space achieving an accuracy of 95.1%.