Parasite Detection in Copro Images with a modified Faster R-CNN

Object detection is an important area in computer vision, which has had a great impact in various areas such as medicine. One of the applications in medicine is the detection of parasites since early detection makes it possible to prevent a possible disease, control a disease, or prevent infections from occurring. In this paper, two models of multiclass detection of parasites are presented: one with the Faster R-CNN network, M1, and another M2 with the Faster R-CNN network modifying its cost functions, termed Faster R-CNNM, due to class imbalance. M1 presented a maximum mean Average Precision, mAP, mAP@0.5:0.95 of 0.8154 at epoch 82, and M2 of 0.8256 at epoch 84 in validation. The two models were analyzed using the precision, P, recall, R, and F1-measure, F, metrics for location and the balanced accuracy metric, BAM, for classification. The analysis consisted of making inferences with an intersection over union, IoU, threshold of 0.5. All images in the data set were used in the analysis. M1 obtained P = 0.814, R = 0.99, F = 0.8935 in localization and BAM = 0.815 in classification. The results of M2 in the localization were P = 0.8417, R = 0.992, F = 0.91, and in the BAM classification = 0.872. These results show that M2 performs better than M1 in location and outperforms M1 in the classification because it is better to deal with the class imbalance situation.