基于改进原型网络的小样本古生物化石识别研究

Research on few-shot paleontological fossil identification based on improved prototype network

传统古生物化石鉴定方法多依赖于古生物学家的经验知识,现有的人工智能识别方法需要大量的化石训练样本才能达到较高的准确率。为解决上述问题,在少量化石图像样本情况下准确识别化石,笔者等尝试使用残差网络和注意力模块相结合的方法,并将其运用于小样本的化石鉴定。首先以残差网络作为模型的特征提取模块,并在残差网络的残差块中嵌入CBAM卷积注意力模块,提高模型对于化石纹理特征的关注,以提取更为全面的深层次化石图像特征,再使用小样本度量学习中的原型网络对提取特征进行原型计算,最后通过多轮次迭代训练得出最佳的化石判别模型。使用本文方法与5种经典的小样本学习方法进行对比实验,实验结果表明本文方法的识别准确率最高,在样本数量为1和5的情况下,准确率达到了86.32%和94.21%,对稀缺样本下的化石识别具有更显著的优势。

Objectives:Traditional methods for identifying fossil organisms often rely on a paleontologist's knowledge and experience,while existing artificial intelligence recognition methods require large amounts of fossil training samples to achieve high accuracy.To address this issue,this paper attempts to use a combination of residual network and attention module and apply it to the identification of fossils in small-sample images.Methods:First,a residual network is used as the model's feature extraction module,and CBAM convolutional attention modules are embedded in the residual blocks of the residual network to improve the model's focus on fossil texture features,extract more comprehensive deep-level fossil image features,then use a prototype network in few-shot metric learning to calculate the extracted features,and finally train the best fossil discrimination model through multiple iterations.Results:This paper compares this method with five classical few-shot learning methods,and experimental results show that this method has the highest recognition accuracy.In the case of 1 and 5 samples,the accuracy is 86.32%and 94.21%,respectively,showing significant advantages in recognizing rare fossil samples.Conclusions:The proposed method in this paper adopts the prototype network,which is a common framework used in few-shot learning,as the backbone.The CBAM convolutional attention module is embedded into the residual block of the ResNet12 residual network to enhance the feature extraction ability of the network.This approach achieves high recognition accuracy with only a small amount of training data for fossil images,which solves the problem of traditional convolutional neural networks requiring a large number of fossil image samples for high accuracy despite the limited availability of such data.However,the current work in this paper only involves training few-shot models on seven species-level categories of fossil images collected,and the fossil dataset still needs further expansion to include more diverse categories.The future research direction of this paper is how to achieve high recognition accuracy using small-sample learning models in situations where multiple categories with significant intra-class differences exist.