Machine-based mosquito taxonomy with a lightweight network-fused efficient dual ConvNet with residual learning and Knowledge Distillation

Taxonomy plays a vital role in identifying different mosquito species. Studies show that though not all mosquitoes threaten humanity, specific species exist in less fortunate areas that immensely disrupt people’s lives. As identified, researchers discovered that deficiency in identifying between a vector mosquito that carries a lethal disease apart from non-vectors led to people becoming susceptible. Recently, studies proposed automating these mosquitoes’ classification so that people who lack awareness can soon obtain assistance from an intelligent system. However, most solutions still require expensive computations and specialized resources to operate and even reproduce, making the most vulnerable areas or groups of people unable to benefit from them. Therefore, this work solves this problem with a lightweight model built by compressing, duplicating, and fusing a Deep Convolutional Neural Network model (DCNN), adding a modified residual block, and training it through Knowledge Distillation (KD). Upon assessment, results yielded significant performance improvements, as the proposed model reached 99.22% overall accuracy that only requires 0.33 GFLOPs to operate and consumes only 437 KB of disk space. In addition, results also showed the benefits of KD in saliency. Compared to most studies, previous and current state-of-the-art DCNNs, this work shows promising viability to solve the problem pragmatically. [Display omitted] •The proposed model classified six mosquito species better than most solutions.•It attained improvements via compression, fusion, residual learning and knowledge distillation.•It has better cost-efficiency than the existing solutions with only 40 K parameters.