Deep neural networks with attention mechanisms for Spodoptera frugiperda pupae sexing

The Spodoptera frugiperda caterpillar is an object of study of great importance to the national economy of Brazil since it directly attacks the cultivations of a diverse amount of agricultural products. Since the sexing of this insect is an arduous and tiring task due to being done manually, this paper proposes to analyse multiple computer vision techniques allied to neural networks, from older approaches to newer ones that use attention mechanisms, and compare their results to determine which type of neural network can be used to help with the laboratorial processes involving individuals from this species. According to the results obtained in this paper, all four techniques reached promising results, with Faster R-CNN reaching approximately 94.9% accuracy, 90.3% for RetinaNet and 96% for both Side-Aware Boundary Localization and Foveabox, showing that it is indeed viable to use automated methods for the detection of pupae genitalia, and that newer approaches that rely on attention mechanisms also perform better overall.