Deep learning framework to extract anatomy for mosquito image classification

• Published in: مجله مدل سازی در مهندسی Vol. 20; no. 70; pp. 107 - 120
• Main Authors: Marzieh Zare Nazari, Mohsen Sardari Zarchi, Sima Emadi, hadi Pourmohammadi
• Format: Journal Article
• Language: Persian
• Published: Semnan University 01.09.2022
• Subjects: classification; convolutional neural networks; deep learning; segmentation mosquito
• ISSN: 2008-4854; 2783-2538
• DOI: 10.22075/jme.2022.26235.2222

Mosquitoes are the main cause of the spread of dangerous diseases such as malaria, yellow fever, dengue fever, and Zika. The most effective way to control these diseases is to correctly identify the types of mosquito species. In the traditional method of identifying mosquitoes, identification is based on morphological diagnoses by specialized human beings with special skills. The most important classification challenge is to reduce the number of experts and the great diversity of different species of mosquitoes. In order to overcome this challenge, developing an automated method based on deep learning architectures to identify and classify mosquitoes will be a valuable resource for non-specialists.This study proposes a convolutional network model that integrates the ResNet101 architecture and the Mask_RCNN technique to segment and classifies mosquito images. 2354 mosquito images of three species of Anopheles, Aedes, and Culex are compared with each other. In the proposed model, instead of entering the network as a complete image of a mosquito, first, the images are segmented, and then different parts of the abdomen, legs, wings, and head are given to the network as input. The corresponding binary mask of the described parts of the mosquito body is produced by the convolution network to extract the feature for each separate part and then calculate the loss value between the classified values and the image label. The evaluation results showed that the extraction of mosquito anatomy images affects the faster classification of images and the network performed better with 97.84% accuracy than normal.