Automated Extraction and Classification of Fossil Whorl Patterns Using Deep Learning: A Novel Approach to Paleontological Image Analysis

With the rapid advancements in information technology, particularly in the domains of big data and computational power, image processing technologies have undergone unprecedented transformations [5]. In the field of AI, and more specifically in image recognition, the application of deep learning—particularly CNNs—has significantly advanced the field by shifting the focus from surface-level features to deeper semantic understanding [6]. These technologies are increasingly demonstrating their unique value within the domain of paleontology, with the recognition of Yangquan Paleontological Fossil Patterns serving as a key empirical case. As crucial evidence of life’s evolutionary history, the intricate and diverse patterns found in paleontological fossils present significant challenges for researchers. Traditional recognition methods were not only time-consuming and subjective but also prone to human error. The integration of modern AI-driven image processing technologies, however, has substantially enhanced both the efficiency and robustness of the recognition process, greatly reducing the potential for human bias and error. As a result, these advancements have allowed for a more accurate and comprehensive understanding of the complex evolutionary narratives embedded within fossil records.

2.1 AI image processing in fossil pattern recognition

The application of AI-driven image processing technologies encompasses various aspects, including image denoising, enhancement, classification, feature extraction, and recognition. Through the integrated application of these techniques, the full potential of image processing is continuously explored and leveraged.

In the context of pattern recognition for fossilized whorls of ancient organisms in the Yangquan region, as illustrated in Figure 1, efficient image preprocessing methods are employed to enhance the quality of input images, mitigate the impact of noise during the training process, and allow the model to focus on more relevant and informative features. Additionally, by utilizing multi-level feature abstraction and the capacity to recognize complex patterns, along with the integration of modern feature matching techniques such as SIFT and multi-scale analysis methods, both the accuracy and efficiency of recognition are significantly improved [7]. Furthermore, the transfer learning capabilities of CNNs are harnessed to rapidly and effectively learn new patterns of paleontological whorl fossils based on limited sample data, thereby facilitating the recognition and classification of fossil patterns in the Yangquan region. The ongoing advancements in image processing technologies have resulted in both visual and technological innovations within the field of paleontology, and the research into fossilized whorl patterns in Yangquan stands as an exemplary case of the fusion of cutting-edge technology with the study of ancient science.

1.png

Figure 1. Process diagram of pattern recognition of Yangquan paleontological fossils

2.2 Application of AI in fossil pattern recognition

The rapid advancements in AI, particularly within the domain of image recognition, have enabled unprecedented speed and accuracy in numerous scientific studies. AI technologies are increasingly being employed to analyze image data at deeper levels, achieving precise and efficient automatic recognition and classification. This is especially true in the field of paleontological research, where AI’s image recognition capabilities offer new methodologies for the identification and classification of paleontological fossils, significantly accelerating progress in related disciplines. In the study of the abundant paleontological fossil resources in the Yangquan region, traditional manual identification methods, limited by their efficiency and accuracy, fail to meet the growing demands of contemporary scientific research. The integration of AI image processing technology has not only addressed these limitations but has also provided a robust technical framework and methodological foundation for ongoing research [8].

In the recognition of paleontological fossil images from the Yangquan area, advanced technologies such as deep neural networks, feature learning, and pattern matching enable AI systems to accurately extract relevant information from complex image data, including lines, patterns, shapes, and colors—critical features for analyzing fossil patterns. By combining classifiers such as CNNs and Support Vector Machines (SVMs), AI technology not only enhances the speed of pattern extraction but also substantially improves the accuracy and reliability of recognition [9].

Through multi-level feature abstraction and synthesis, a highly efficient and intelligent recognition model has been developed. An improved least squares ellipse fitting method, in conjunction with the Faster Regional Convolutional Neural Network (Faster RCNN) algorithm, has enabled the model to achieve precise and robust pattern recognition across diverse and complex backgrounds. When trained and tested on an actual dataset of paleontological whorl fossils, the AI model demonstrated a recall rate exceeding 95% and a recognition accuracy of 99%, significantly outperforming traditional methods. Furthermore, AI image recognition technology exhibits excellent adaptability. By employing strategies such as transfer learning and data augmentation, the system can quickly adapt and accurately recognize new fossil patterns, even when faced with sparse data. This highlights the powerful capabilities of AI in handling diverse and evolving image datasets [10].

4.1 Collection of wheel leaf fossils

As early as 2014, lobe fossils were found in waste rock piles of some coal mining mines in Yangquan. In 2016, the staff of the Yangquan Paleontological Fossil Protection Association and Yangquan Planning and Natural Resources Bureau collected them in the open-pit coal mining pit in the Yangquan mining area, and then they were found to be rich in paleontological fossils in several engineering construction areas and collected them.

The fossil images of Verticillata used in this paper mostly come from collecting fossils in the Yangquan area. A total of 200 fossils were collected, and 30 high-resolution images were sourced from China's Geological Museum. An additional 70 images were added from the public Paleobiology Database. All images are standardized to ensure the consistency of resolution and format. The data set consists of 280 images of fossil lobes, covering different species. Among them, complete fossil images account for 60%, and partial fossil images account for 40%. Shooting angles include front, side, and top view angles to ensure the comprehensiveness of features. In addition, the backgrounds of the fossil images in the data set are varied, with rocks, soil, and lab backgrounds, which can help the model work better in more complex settings. Fossils span from the Carboniferous to the Permian, reflecting the characteristics of different geological periods.

4.2 Preprocessing of wheel leaf fossil images

In the data preprocessing stage, we adjust all images to uniform resolution (350×300 pixels) and convert them into BMP format to preserve high-quality details. Through background separation technology, the interference of soil and rock in the image is removed. In addition, we have enhanced the data set, including rotation, scaling, and flipping, and processed the collected original data to enhance the training effect and further expand the data set scale.

When capturing and transmitting images, due to the influence of the environment and equipment, the image information inevitably contains noise, and it is necessary to smooth the image. The median filtering method has a significant effect on eliminating this type of noise and can effectively preserve image edge information. Sort the grayscale values of all pixels within the pixel and its neighboring window, and take the median value as the new grayscale value for that pixel [12]. The principle is shown in Figure 2. Taking a 3×3 window as an example, the original pixel point in the figure is 61, and the grayscale values in adjacent windows are arranged in ascending order, with the sequence being [60, 61, 62, 63, 64, 65, 66, 67, 68]. The median value 64 is replaced with the original value 61 as the new grayscale value, so that the surrounding grayscale values are close to the actual value, thereby eliminating noise points, as shown in Figure 3.

2.png

Figure 2. Principle of binary filtering

3.png

Figure 3. Image denoising

Image segmentation is an important step in image preprocessing, extracting fossil image regions from the image. In practical applications, the use of the image local segmentation ROI style algorithm is widely adopted due to its fast calculation speed and low computational complexity. Assuming the image function is $f(x, y)$, its grayscale value is compared with the set threshold T to generate a new image function through the following method:

$g(x, y)=\left\{\begin{array}{l}1, \text { if } f(x, y)>T \\ 0, \text { if } f(x, y) \leq T\end{array}\right.$

In this formula, the pixel value $f(x, y)$ is compared with a threshold T. If the grayscale value is higher than the threshold, the pixel is marked as an object (with a value of 1); if it is below the threshold, it is marked as background (value 0). In this way, through grayscale threshold segmentation, the image is divided into two parts, usually foreground and background. However, using a fixed grayscale threshold can be affected by factors such as uneven image lighting, resulting in unsatisfactory segmentation results. To solve this problem, the image can be divided into multiple sub images, and the grayscale values of neighboring pixels of a certain pixel point in each sub image can be weighted and averaged to obtain the threshold of the local region [13].

4.png

Figure 4. Local image segmentation

By using the above methods, the impact of uneven lighting can be effectively avoided, thereby achieving effective segmentation of images. Especially for images with different brightness and backgrounds, such as fossil images, using local thresholding can significantly improve segmentation performance. The effect of the segmented fossil image is shown in Figure 4. This method can effectively address the issue of lighting changes in images, thereby improving the accuracy of image segmentation.

4.3 Pattern extraction and feature analysis

The multi-level CNN structure used effectively extracts the texture features and color differences of the surface of paleontological whorl fossils. In order to improve the accuracy of the recognition model, an HSV (hue, saturation, brightness) model for the color of paleontological leaf fossils was also added. The RGB image corresponds to the hardware output, while the HSV image is more in line with the intuitive vision of the human eye. Therefore, when processing leaf fossil images, the RGB image is first converted to an HSV image, and the image is processed in the HSV color space. First, the R, G, and B values need to be converted between 0 and 1. Let the largest component of the R, G, and B components of the image be Xmax and the smallest component be Xmin, then the conversion relationship between the RGB model and the HSV model is:

$\begin{aligned} & \mathrm{R}=\mathrm{R} / 255 \\ & \mathrm{G}=\mathrm{G} / 255 \\ & \mathrm{~B}=\mathrm{B} / 255\end{aligned}$     (1)

$\mathrm{H}=\left\{\begin{array}{ll}60 *(G-B) /(X-\min (R, G, B)), & \text { if } X=R \\ 120+60 *(B-R) /(X-\min (R, G, B)), & \text { if } X=G \\ 240+60 *(R-G) /(X-\min (R, G, B)), & \text { if } X=B\end{array}\right\}$     (2)

$S=\left\{\begin{array}{cc}\frac{X-\min (R, G, B)}{X}, \text { if } & X \neq 0 \\ 0, & \text { else }\end{array}\right\}$     (3)

$\mathrm{V}=\max (\mathrm{R}, \mathrm{G}, \mathrm{B})$     (4)

Using the conversion relationship from the RGB model to the HSV model, the image of the fossilized wheel leaf was transformed, and the HSV model image is shown in Figure 5.

5.png

Figure 5. Original image of wheel leaf fossil and HSV model image

In the HSV model, binarization was performed on the dataset of fossilized wheel leaf images [14]. Based on the HSV color reference range in Table 3, the approximate range of feature thresholds for HSV model wheel leaf fossil images is preliminarily determined to be (H=35~77, S=43~255, V=46~255); By observing the histogram of the HSV model (see Figures 6-7) and continuously refining and adjusting it, the binary threshold range of the HSV model was finally determined to be (H=35-60, S=162-214, V=48-125). The features of the wheel leaf fossil image are shown in Figure 8. By statistically analyzing the color characteristics of fossil images of paleontological whorls, the color tone variation patterns of different paleontological whorls were finely distinguished.

Table 3. HSV color reference range

6a.png

(1) Original histogram

6b.png

(2) HSV model histogram

Figure 6. Image histogram

7.png

Figure 7. Image segmentation map of HSV model

8.png

Figure 8. Binary different threshold maps

9.png

Figure 9. Partial wheel leaf fossil image dataset

Further combining texture based gray level co-occurrence matrix, entropy, contrast and other parameters, by analyzing the spatial relationship between pixels and the statistical distribution of gray levels in the image, the co-occurrence relationship between different gray levels in the image can be captured. It can provide information about the grayscale direction, spacing, and magnitude of changes in the image, and calculate feature values through it to describe various texture features in the image, thereby improving the model's ability to recognize the texture of paleontological whorl fossils.

4.4 Experimental model design and methods

In order to effectively extract and identify the patterns of paleontological whorl fossils in the Yangquan area, an image processing technique combining gray level co-occurrence matrix (GLCM) and CNN was adopted. Firstly, over 300 images of paleontological whorl fossils were collected from the fossil repository in the Yangquan area, and the images were uniformly output as 350 pixels by 300 pixels using Matlab, as shown in Figure 9.

Divide the training set and testing set in a 9:1 ratio, with 270 training sets and 30 testing sets. These images need to undergo strict preprocessing before processing to ensure the quality and consistency of the input data. Preprocessing includes steps such as grayscale conversion, noise removal, and contrast enhancement to ensure the accuracy of feature extraction. Subsequently, GLCM was used to extract texture features from the preprocessed image. Based on the texture parameters obtained from the gray level co-occurrence matrix, such as energy, contrast, uniformity, entropy, etc., representative feature vectors were constructed. These complex texture features provide an important basis for identifying patterns of fossilized whorls in ancient organisms. Subsequently, a CNN based deep learning model was designed, which includes multiple convolutional layers, pooling layers, and fully connected layers, as shown in Figure 10. ReLU was used as the activation function to ensure that the model can effectively learn and extract advanced features from images [15].

In order to speed up the model training and improve the recognition accuracy, the transfer learning strategy is adopted, and the pre-trained network weights are used as the starting point to train the model, which shortens the training time and improves the recognition rate [16]. In the experiment, firstly, the SIFT algorithm is used to extract local feature points and descriptors of images, and preliminary feature matching is carried out to screen out possible similar image subsets. Then, using CNN's powerful feature learning ability, these candidate images are more finely extracted and classified to judge the similarity of images and improve the accuracy and efficiency of retrieval. Aiming at the image data set of paleontological lobe fossils, the experiment adopts cross-validation to ensure the generalization ability of model training. All experiments are carried out on the same training set and test set, which ensures the objectivity and accuracy of the evaluation results. Figure 11 shows the network architecture diagram of this paper. The features of the input image are extracted through multiple convolution layers and pooling layers, and finally the classification results are output through the fully connected layer and SoftMax layer.

By comparing the baseline model with the model using advanced feature extraction techniques, the results show that the latter can significantly improve recognition accuracy, reaching over 95%, which is significantly better than manual pattern matching and traditional machine learning methods. The AI image processing technology presented in this article provides a powerful tool for extracting and applying fossil patterns of ancient organisms in the Yangquan area, opening up new paths for paleontological research and cultural heritage protection [17].

In order to verify the effectiveness of the network architecture, this paper uses ResNet-50 as the basic model. We compare the effects of different convolution kernel sizes of 3×3 and the activation function ReLU, and fine-tune them to adapt to the recognition task of paleontological lobe fossil patterns. The experimental results show that the classification accuracy of the network using a 3×3 convolution kernel and ReLU activation function is above 95% on the test set, which is obviously superior to manual pattern matching and traditional machine learning methods. The AI image processing technology in this paper provides a powerful tool for the extraction and application of paleontological fossil patterns in the Yangquan area, and opens up a new path for paleontological research and cultural heritage protection.

10.png

Figure 10. Model architecture

11.png

Figure 11. Model training diagram

4.5 Analysis of experimental results

In this paper, the pattern recognition and feature extraction of paleontological lobe fossils in the Yangquan area are discussed in depth. Through the large-scale data training of more than 300 fossil pictures of paleontology, the experiment uses advanced AI image processing technology and CNNs, which significantly improve the accuracy and efficiency of pattern recognition. In the specific experimental operation, the CNN model used in this paper is optimized by a specific algorithm, and in the process of multi-layer feature extraction and abstraction, supplemented by appropriate image enhancement and regularization skills, which ensures the network's deep learning and recognition ability for the complex texture of paleontological lobe fossil patterns. The experiments on the verification set show that the recognition accuracy of the optimized model for the fossil patterns of Yangquan paleontology is over 95%, which is much higher than the traditional manual recognition efficiency, and it shows excellent transfer learning ability and generalization performance in the new pattern recognition task. Further analyzing the experimental results, it is found that the model is particularly obvious in learning the unique texture features of paleontological lobe fossils, such as patterns with clear texture and strong color contrast, and the visual results show the high sensitivity of the model in feature recognition. Through SIFT feature matching technology, the model realizes more accurate and stable extraction of key feature points, which greatly improves the speed and accuracy of feature matching. In the face of pattern change and noise disturbance, the model can still maintain high robustness and recognition rate.

In addition, AI technology has important practical application value in pattern extraction and classification recognition of paleontological fossils. Traditional methods rely on manual observation and measurement, which is inefficient and subjective, while AI technology can automatically and efficiently process a large number of fossil image data, which significantly improves the research efficiency. AI models can complete the classification and feature extraction of a fossil image in a few seconds, while traditional methods may take hours or even days. Fossil features extracted by the deep learning model are more accurate and can identify subtle differences that are difficult for human eyes to detect. As can be seen from Figure 12, the F1-score value of the SIFT and CNN algorithms is 0.099 higher than that of the SIFT algorithm, but the performance of the traditional image processing method is lower than that of the deep learning method. Compared with the image contour extraction algorithm based on the ResNet-50 model, the improved image feature point extraction algorithm based on the SIFT algorithm improves the F1-score value by 0.028 and has better performance. The automated process of AI technology ensures the repeatability of research results and reduces human error. AI technology can process massive fossil image data, which provides unprecedented data support for paleontology research.

12.png

Figure 12. Graph of artificial and AI technology algorithm