Computer Vision Revolutionizes Carnivore Tooth Mark Identification

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Mar 28 2024

In an article recently published in the journal Nature, researchers explored advancements in taphonomic analysis by employing computer vision (CV) to objectively identify taxon-specific carnivore agency in fossilized tooth marks.

By achieving up to 88% accuracy in discriminating carnivore species, the authors bridged traditional taphonomic-paleontological methods with artificial intelligence-based computer science. This interdisciplinary approach promised to revolutionize the understanding of hominin-carnivore interactions and their impact on human evolution.

Background

The identification of taxon-specific carnivore agency is crucial for understanding hominin-carnivore interactions in archaeological reconstructions. Previous studies have relied on traditional taphonomic methods, which, while useful, suffer from subjectivity and limited ability to discriminate between different carnivore agents. This has led to challenges in accurately determining the roles of specific carnivores in site formation processes and their interactions with hominins. Moreover, the inability to distinguish between different carnivore types has hindered interpretations regarding hominin behavior and evolutionary history.

To address these gaps, this paper presented a novel approach using CV to objectively identify taxon-specific carnivore agency based on tooth marks found on bones. While previous studies have attempted similar methods, they have been limited by small sample sizes or inadequate image standardization, resulting in moderate accuracy rates. This study improved upon previous work by utilizing a larger and more diverse sample of tooth marks from modern carnivores, including lions, leopards, hyenas, and crocodiles.

This research significantly enhanced the accuracy and reliability of CV-based carnivore identification by creating a comprehensive graphic library and developing refined models. These advancements contributed to improving our understanding of hominin-carnivore interactions and paved the way for addressing key questions in human evolutionary history with greater precision and objectivity.

Samples and Methods

The authors aimed to create a robust reference database for identifying taxon-specific carnivore tooth marks on bones using CV methods. A total of 1256 tooth marks from lions, leopards, hyenas, and crocodiles were analyzed. Tooth marks were collected from experimental feeding trials with captive carnivores and naturalistic settings, ensuring a diverse range of marks.

Notably, a new image bank was created using a digital microscope capable of capturing both bidimensional and tridimensional images, improving image quality compared to previous studies. The experimental series involved feeding carcasses of various animals to hyenas, lions, leopards, and crocodiles, with bones collected after exposure to each carnivore.

Additionally, tooth marks from wild hyena dens and leopard feeding trials were included to broaden the dataset's range. Deep learning (DL) methods, specifically convolutional neural networks (CNNs), were employed for tooth mark classification. Four pre-trained architectures—ResNet 50, VGG19, Densenet 201, and EfficientNet B7—were used individually, and their outputs were combined using ensemble learning techniques for improved accuracy.

The DL models were trained and tested on a randomized split of the tooth mark images, with image augmentation techniques applied to enhance training. Exploratory analyses determined the optimal activation function, rectified linear unit (ReLU), and optimizer, stochastic gradient descent (SGD), for each DL model. Regularization techniques such as dropout were employed to prevent overfitting during training.

Overall, the study established a comprehensive database of taxon-specific carnivore tooth marks and demonstrated the efficacy of DL methods in accurately classifying these marks. By leveraging modern technology and experimental data, the research significantly advanced taphonomic analysis, offering a more objective and precise approach to identifying carnivore agency in archaeological and paleontological contexts. These findings paved the way for further interdisciplinary research at the intersection of traditional taphonomy, paleontology, and artificial intelligence.

Results

The study evaluated four DL architectures—Resnet50, EfficientNet B7, VGG19, and Densenet 201—for their ability to classify taxon-specific carnivore tooth marks on bones. Accuracy estimates on the testing dataset exceeded 80%, with Resnet50 achieving the highest accuracy at 88%, followed by Densenet 201 (84.3%), and VGG19 and EfficientNet B7 (both 80%). Ensemble learning approaches, combining the outputs of all models, also demonstrated high accuracy, with an average of 83%.

Dropout regularization effectively prevented overfitting in most models, ensuring robust training. The learning graphs illustrated smooth training processes for Resnet50 and VGG19, with no signs of overfitting. In contrast, EfficientNet B7 and Densenet 201 showed some stagnation after epoch 40 due to increasing overfitting, yet still maintained relatively high accuracy.

Misclassifications were primarily observed in crocodile tooth marks, likely due to the smaller sample size compared to other carnivores. However, F1 score values indicated low classification errors for hyenas, lions, and leopards, with most misclassifications occurring in crocodiles. Removing crocodile tooth marks from the dataset did not significantly impact classification accuracy for the remaining three carnivores. Overall, Resnet50 emerged as the most effective model for classifying taxon-specific carnivore tooth marks, achieving high accuracy across all tested carnivores.

The authors underscored the potential of DL methods, particularly Resnet50, in objectively identifying carnivore agency in archaeological and paleontological contexts. The findings highlighted the importance of robust experimental datasets and demonstrated the feasibility of leveraging modern technology to enhance taphonomic analysis. Future research might explore further optimization of DL models and applying ensemble learning techniques to refine classification accuracy and address imbalanced sample issues, ultimately advancing our understanding of hominin-carnivore interactions and human evolutionary processes.

Discussion

Previous methods for distinguishing carnivore taxa based on tooth marks focused on bidimensional metrics, revealing overlap across carnivore sizes and taxa. Geometric-morphometric (GMM) analyses offer higher discriminatory power but suffer from small sample sizes and potential biases. In contrast, DL models presented by the authors achieved superior accuracy by combining toothpits and scores and utilizing high-quality image data.

These models offered an objective means of differentiating carnivore types in bone surface modifications, enhancing interpretations of agency in archaeological and paleontological contexts. These models could refine paleoecological reconstructions and shed light on early human-carnivore interactions, potentially challenging traditional hypotheses. Moreover, their application extended to modern ecology, providing insights into carnivore impact on ecosystems through prey bone analysis.

Conclusion

In conclusion, the integration of CV with traditional taphonomic analysis represented a significant advancement in understanding hominin-carnivore interactions and their implications for human evolution. By achieving up to 88% accuracy in identifying taxon-specific carnivore agency, this interdisciplinary approach bridged gaps in paleontological and artificial intelligence research. The robust experimental dataset and deep learning models, particularly Resnet50, offered a more objective and precise means of classifying carnivore tooth marks.