Effects of Background Color, Flash and Exposure Value on the Accuracy of Smartphone-Based Pill Recognition System Using Deep Convolutional Neural Network (Preprint)
BACKGROUND It is difficult to develop a drug image recognition system due to the difference of the pill color influenced by external environmental factors such as the illumination or presence of flash. OBJECTIVE In this study, we wanted to see how the difference in color between the reference image and the real-world image affects the accuracy in pill recognition under 12 real-world conditions according to the background colors, presence of flash, and exposure values (EV). METHODS We used 19 medications with different features of colors, shapes, and dosages. The average color difference was calculated based on the color distance between the reference image and the real-world image. RESULTS In the case of the black background, as the exposure value lowered, the accuracy of top-1 and top-5 increased independently of the presence of flash. The top-5 accuracy in black background increased from 26.8% to 72.6% with the flash on and from 29.5% to 76.8% with the flash off as EV decreased as well. On the other hand, top-5 accuracy was 62.1% to 78.4% in white background with the flash on. The best top-1 accuracy was 51.1 % in the white background, flash on, and EV+2.0. The best top-5 accuracy was 78.4% in the white background, flash on, and EV0. CONCLUSIONS The accuracy generally increased as the color difference decreased except in the case of black background and EV-2.0. This study reveals that the background colors, presence of flash, and exposure values in real-world conditions are important factors affecting the performance of a pill recognition model.