面向社会性昆虫识别的知识迁移DenseNet后训练剪枝轻量化模型研究

Research on Lightweight Model of Post-training Pruning of Knowledgetransferred DenseNet for Social Insects Identification

在资源受限的设备上,如何快速有效地针对蜜蜂、蚂蚁等对生态系统产生重要影响的种群进行图像识别,具有重要的生态保护意义.本文采用DenseNet预训练模型,在蚂蚁蜜蜂小规模数据集上进行知识迁移,并利用非结构化后训练剪枝UPSCALE方法,构建了一个完整的架构.实验证明,该架构可以快速利用小规模数据集,以较高的识别精度实现目标图像识别,且模型参数不到基准方法的1/3,对于部署设备而言,具有更广泛的应用价值.

The rapid and effective identification of populations such as bees and ants,which have significant ecological impacts,on resource-constrained devices holds great ecological conservation significance.In this paper,a DenseNet pre-trained model is employed for knowledge transfer on a small-scale dataset of ants and bees,and an unstructured post-training pruning method known as UPSCALE is utilized to construct a comprehensive framework.Experimental results demonstrate that this framework can rapidly leverage small-scale datasets to achieve target image recognition with high accuracy,while the model parameters are less than one-third of those of the baseline method,thus providing broader application value for deployment on devices.