基于Point Transformer 方法的鱼类三维点云模型分类

CLASSIFICATION OF 3D POINT CLOUD MODELS OF FISH BASED ON POINT TRANSFORMER APPROACH

  • 摘要: 为实现对不同鱼类的精准分类, 研究共采集110尾真实鱼类的三维模型, 对获取的3D模型进行基于预处理、旋转增强和下采样等操作后, 获取了1650尾实验样本。然后基于Point Transformer网络和2个三维分类的对比网络进行数据集的分类训练和验证。结果表明, 利用本实验的目标方法Point Transformer获得了比2个对比网络更好的分类结果, 整体的分类准确率能够达到91.9%。同时对所使用的三维分类网络进行有效性评估, 3个模型对于5种真实鱼类模型的分类是有意义的, 其中Point Transformer的模型ROC曲线准确率最高, AUC面积最大, 对于三维鱼类数据集的分类最为有效。研究提供了一种可以实现对鱼类三维模型进行精准分类的方法, 为以后的智能化渔业资源监测提供一种新的技术手段。

     

    Abstract: Phenotypic data serve as the foundation for effective monitoring of fish species. Currently, fish classification heavily relies on expertise from relevant professionals, leading to issues such as low efficiency, high errors, potential damage to fish bodies, and susceptibility to subjective factors affecting data quality. In this study, we developed a simplified device for acquiring three-dimensional models of fish, pioneering the creation of a dataset comprising authentic three-dimensional fish models. By leveraging the Point Transformer algorithm, we can rapidly, efficiently, and accurately extract phenotypic features from three-dimensional fish bodies, enabling precise classification of different fish species. A total of 110 authentic fish three-dimensional models were collected in this research, resulting in 1650 experimental samples after preprocessing, rotation enhancement, and downsampling operations on the acquired 3D models. Subsequently, through classification training and validation using the Point Transformer network and two comparative networks for three-dimensional classification, the results indicate that the proposed Point Transformer method outperforms the two comparative networks, achieving an overall classification accuracy of 91.9%. Simultaneously, an effective evaluation of the utilized three-dimensional classification networks was conducted, demonstrating the meaningful classification of the three models for five authentic fish species models. The Point Transformer model exhibited the highest ROC curve accuracy and the largest AUC area, proving its effectiveness in classifying three-dimensional fish datasets. This study presents a method for accurately classifying three-dimensional fish models, offering a new technological approach for intelligent monitoring of fisheries resources in the future.

     

/

返回文章
返回