A Deep Learning Approach for Locating Microalgae in Image Captured by Smartphone and Digital Microscope Camera

The utilization of microalgae as a bioindicator for water quality assessment has traditionally relied on the expertise of morphological identification. With the advent of technological advancements, particularly the smartphone camera, there has been increasing interest in using it as a tool for various fields of microbiological and medical diagnostics, including image analysis of microalgae. Despite the growing popularity of smartphone-based image analysis, the image quality of smartphone cameras for detecting microalgae has yet to be thoroughly evaluated. Therefore, the purpose of this study was to investigate the suitability of smartphone-based camera for collecting microalgae images and compare it to that of a digital microscope camera. Additionally, the effect of different microscopic magnifications at 100x and 400x on model performance was thoroughly studied. Four models, YOLOv5, RetinaNet, EfficientDet, and Faster-RCNN, were trained on microalgae image datasets taken by both a smartphone camera (SC) and a digital microscope camera (DMC) under 100x and 400x magnification. A total of 11,608 images were divided into three parts: train (70%), validation (10%), and test (20
^%
), with the test dataset separated into four parts for each device and magnification. The results of the study indicated that Faster-RCNN was the best model for microalgae detection, with the highest average precision
$(\mathbf{AP}^{0.5})$
values of 0.60, 0.84, 0.91, and 0.96 at 100x magnification of SC and DMC and 400x magnification of SC and DMC, respectively. The comparison between the image quality of SC and DMC at each magnification revealed that the SC was less suitable for microalgae identification, particularly at 100x magnification. Conversely, the 400x magnification of SC had the potential to identify microalgae through the use of smartphone-based photography, as well as digital microscope camera.