果蝇是实验教学中最常用的重要生物材料之一。在果蝇实验教学中，每个学生通常需要针对上百只果蝇进行手工辨认，并记录每只果蝇身上的数个不同性状，工作量大且分类标准参差不齐。为了解决这一问题，本文将现代计算机技术融入到遗传学实验教学中，使用深度卷积神经网络来自动统计每只果蝇的性状。采用的是目标检测模型+分类模型的两阶段策略模式。在分类模型的训练设计过程中，创新性利用了关键点辅助分类的方法，有效地提升了模型的可解释性。此外，还针对任务特性改善了RandAugment方法，利用渐进式学习与适应性正则化策略，在有限的计算资源下训练了MobileNetV3架构下的多标签分类任务，并最终在每只果蝇3对性状(红/白眼、长/小翅、雌/雄)的分类任务下分别达到了97.5%、97.5%和98%的准确率。模型经过优化后，可以在手机端10 s内完成600个果蝇性状的分类，该模型具有轻量化的特点，大小不到5 MB，易于在各类安卓系统手机上安装使用。该系统的开发有利于推进以果蝇为研究对象的遗传规律验证等实验的教学，也可用于涉及大量果蝇分类统计分析的科研工作。

Drosophila is a crucial biological experimental teaching material extensively utilized in experimental teaching. In this experimental teaching, each student typically needs to manually identify hundreds of fruit flies and record multiple of each fly. This task involves substantial workload, and the classification standards can be inconsistent. To address this issue, we introduce a deep convolutional neural network that classifies the traits of every fruit fly, using a two-stage consisting of an object detector and a trait classifier. We propose a keypoint-assisted classification model with tailored training session for the trait classification task and significantly enhanced the model interpretability. Additionally, we’ve enhanced the RandAugment method to better fit the features of our task. The model is trained with progressive learning and adaptive regularization under limited computational resources. The final classification model, which utilizes MobileNetV3 as backbone, achieves an accuracy of 97.5%, 97.5% and 98% for the eyes, wings, gender tasks, respectively. After optimization, the model is highly lightweight, classifying 600 fruit fly traits from raw images in 10 seconds and having a size less than 5 MB. It can be easily deployed on any android device. The development of this system is conducive to promoting the experimental teaching, such as verifying genetic laws with Drosophila as the research object. It can also be used for scientific research involving a large number of Drosophila classifications, statistics and analyses.