Dilated YOLOv8 for Blur-Robust Image Classification on the MosquitoLarvae-7400 Dataset

Accurate identification of mosquito larvae is crucial for monitoring and controlling mosquito-borne diseases. This study introducesMosquitoLarvae-7400, a new benchmark dataset comprising about 7,400 high-resolution images of four mosquito species, captured using laboratory microscopes and mobile phones with macro lenses. The dataset is designed to support the development of deep learning models for species-level classification under varied imaging conditions. To demonstrate its utility, we evaluated a wide range of models, including variants of VGG, ResNet, Inception, EfficientNet, and YOLO. We also proposeDilated YOLOv8, a modified YOLOv8 architecture that integrates dilated convolutions to improve robustness to image blur and degradation. Models were trained on both the original dataset and an augmented version six times its size. Our results show that while most models achieved over 99% accuracy on clean images, their performance declined significantly on blurred, distribution-shift images. Most models showed improved robustness when trained with the augmented dataset. Ultimately, EfficientNet-B3 and the proposed Dilated YOLOv8 consistently achieved the highest accuracy; however, our model offers comparable performance at a significantly lower computational cost. The MosquitoLarvae-7400 dataset and source code are publicly available to encourage reproducibility and further research in mosquito-borne-disease surveillance.