Comparative Study of Neural Networks Used in Halyomorpha Halys Detection

The paper’s purpose was to investigate some methods based on neural networks for the detection and classification of harmful insects for agriculture as the Halyomorpha Halys. The implementation of different object detection networks for image categorization was analyzed. Images from the Maryland Biodiversity database were used for neural network training and testing. Rotation, scaling, blurring, mirroring, and other techniques were employed for data augmentation. For the detection and classification of Halyomorpha Halys, some neural networks that include multiple smaller networks were implemented and investigated. The networks used are the following: YOLOv5s, SSD with different backbones such as MobileNet V1, MobileNet V2, and ResNet-50, Faster R-CNN with ResNet-50 backbone, and EfficientDet-D0. Moreover, neural networks were evaluated and compared based on performance metrics such as accuracy and time. Performances like accuracy between 0.49 – 0.86 and time between 36 ms – 55 ms were obtained. The best results were obtained for YOLOv5s, in terms of accuracy, and EfficientDet-D0, in terms of time.