赵金凯, 宁春林, 王为民, 等, xxxx. 基于MobileViT的海洋浮游生物自动分类[J]. 海洋科学进展, x(x): xx-xx. doi: 10.12362/j.issn.1671-6647.20230923001.
引用本文: 赵金凯, 宁春林, 王为民, 等, xxxx. 基于MobileViT的海洋浮游生物自动分类[J]. 海洋科学进展, x(x): xx-xx. doi: 10.12362/j.issn.1671-6647.20230923001.
ZHAO J K, NING C L, WANG W M, et al, xxxx. Automated plankton classification based on mobilevit[J]. Advances in Marine Science, x(x): xx-xx. DOI: 10.12362/j.issn.1671-6647.20230923001
Citation: ZHAO J K, NING C L, WANG W M, et al, xxxx. Automated plankton classification based on mobilevit[J]. Advances in Marine Science, x(x): xx-xx. DOI: 10.12362/j.issn.1671-6647.20230923001

基于MobileViT的海洋浮游生物自动分类

Automated Plankton Classification Based on MobileViT

  • 摘要: 随着海洋生态环境研究的不断深入,对浮游生物群落结构时空变化观测的需求日益增长。然而,目前基于人工智能的浮游生物分类算法,如多层卷积神经网络(Convolutional Neural Network, CNN),在捕捉浮游生物多样性形态方面存在局限,效率和准确率均较低。为解决这一难题,本文提出一种浮游生物分类算法,该算法以轻量化视觉转换模型(Mobile Vision Transformer, MobileViT)为基础架构,使用深度可分离卷积替代标准卷积以减少参数的数量,有效减少模型复杂度;同时,利用局部感知注意力机制降低了注意力计算的复杂度,从而提高了模型的计算效率。本文使用2018年在南太平洋海域采集的剖面浮游生物个体图像,构建了一个包含9个浮游生物类别的数据集。基础分类模型应用于人工分类验证集时,加权平均准确率达到了92.04%。为进一步提升分类性能,本文在模型尾部应用分步概率滤波器消除错误分类,加权平均精确率提高到了96.93%。此外,本文使用同海域另一剖面的浮游生物个体图像作为测试集,对改进后的模型进行了测试,结果显示,该模型在测试集上的top1准确率达到了93.77%,为海洋浮游生物的分类工作提供了一种更加高效、准确的方法。

     

    Abstract: As the research on marine ecological environment continues to deepen, the demand for observing the spatial and temporal changes in the structure of plankton communities is growing. However, current artificial intelligence-based classification algorithms for plankton, such as multi-layer convolutional neural networks (CNN), have limitations in capturing the morphological diversity of plankton, resulting in low efficiency and accuracy. To address this challenge, this paper proposes an innovative algorithm based on the lightweight visual transformation model (Mobile Vision Transformer, MobileViT). It replaces standard convolutions with depthwise separable convolutions to reduce the number of parameters and effectively decrease model complexity. At the same time, it utilizes a local perception attention mechanism to lower the complexity of attention calculations, thereby enhancing the model's computational efficiency. This paper utilizes individual images of plankton collected from the South Pacific Ocean in 2018 to construct a dataset encompassing nine plankton categories. When applied to a manually classified validation set, the base model achieves a weighted average accuracy of 92.04%. To further improve classification performance, this paper innovatively applies a step-by-step probability filter at the tail of the model to eliminate misclassifications, resulting in an increased weighted average precision of 96.9%. Additionally, this paper tests the improved model using individual images of plankton from another profile in the same oceanic region as the test set. The results show that the model achieves a top1 accuracy of 93.77% on the test set, providing a more efficient and accurate method for marine plankton classification.

     

/

返回文章
返回