Fig 1.
Life stages of Dendrocephalus brasiliensis: a) cysts, b) nauplius, c) juvenile, and d) adult.
Fig 2.
Substrate image captured by the XTRAD USB digital microscope model XT-2036 at a 52x magnification.
Each red rectangular bounding box displays a cyst.
Fig 3.
The labeling process using the LabelImg software: a) Sample of substrate image captured by the XTRAD USB digital microscope model XT-2036 at a 52x magnification; b) Using the LabelImg software to label samples of cysts. The green rectangles in the image are the labeled cysts.
Fig 4.
The testing set building process: a) The substrate split into small portions on a white coverslip. b) Folder name (label) that indicates the amount of cyst (142) and the weight of the substrate (0.485g) contained in the folder. c) Sample of images in folder captured by the XTRAD USB digital microscope model XT-2036. The file name indicates the image number and amount of cyst contained in it. For example, in the 1_13.jpg file, 1 indicates the file number and 13 indicates the number of cysts in the image.
Table 1.
The testing set: 10 folders, 135 images, 4.24 grams of substrate, and 1968 cysts.
Fig 5.
Overview of the automatized approach for Dendrocephalus Brasiliensis cysts detection and counting using YOLOv3.
Fig 6.
Overview of the automatized approach for Dendrocephalus Brasiliensis cysts detection and counting using Faster R-CNN architecture.
Table 2.
Cyst detection results at average percentage through 10 test subsets on the DBrasiliensis dataset for YOLOv3 and Faster R-CNN.
Fig 7.
Example of detecting and counting cysts of the Faste R-CNN and YOLOv3 architectures: a) The Faster R-CNN detected and counted only 1 of 20 cysts in the image; b) The YOLOv3 detected and counted 10 of 20 cysts in the image; c) The Faster R-CNN detected and counted 17 of 36 cysts in the image; d) The YOLOv3 detected and counted 35 of 36 cysts in the image.
Fig 8.
Comparison of the number of false positives of the Faster R-CNN and YOLOv3 architectures in different batch sizes.
Table 3.
Detection and counting results for all batch sizes on the 10 testing subsets of the DBrasiliensis dataset for YOLOv3 and Faster R-CNN.
Fig 9.
Example of detecting and counting cysts of the Faster R-CNN (a) and YOLOv3 (b) architectures, both with batch size set at 2. Green boxes are true positives with score detection ≥ 0.3, red boxes are false positives, and blue boxes are false negatives.
Table 4.
RSME, MAE, and R2 through 10 test subsets on the DBrasiliensis dataset for YOLOv3 and Faster R-CNN.