Introduction

Deep learning, a subfield of machine learning, focuses on developing algorithms that emulate human cognitive processes for problem-solving. Deep learning techniques have gained prominence in various fields, including computer vision, where Convolutional Neural Networks (CNNs) have emerged as a powerful tool. CNNs excel in automatically learning and extracting complex features from images, leading to significant advancements in image classification, object detection, and image segmentation. Image classification is fundamental to organism identification, and CNNs have significantly improved the accuracy and efficiency of this task. In image classification, CNNs are trained to categorize images into predefined classes, such as different species of plants or animals.

CNNs have demonstrated remarkable potential in identifying various organisms, including bacteria, plants, fish, and insects. For instance, researchers have successfully employed CNNs to classify bacterial species with high accuracy. One study reported a CNN model achieving an impressive 99.35% accuracy in identifying 33 different bacterial species [1, 2, 3].

In plant science, CNNs have been applied to automatically identify crop diseases. These applications can serve as valuable tools for assisting experts or for automated screening, contributing to more sustainable agricultural practices and improved food production security [4]. CNNs have proven effective in insect sound recognition. By analyzing spectrograms, which are visual representations of sound frequencies, CNNs can accurately identify different insect species. This capability has applications in ecological monitoring, pest control, and biodiversity studies [5].

While research on CNN-based identification of mollusca is an emerging field, some studies have explored the use of CNNs for identifying cephalopod species based on their beaks. This model utilized pre-trained CNNs (VGG19, InceptionV3, and ResNet50) to extract deep features from beak images. The study found that deep features were more accurate than traditional morphometric features in highlighting differences between cephalopod species. [6]

Despite their successes, CNN models have limitations. One major limitation is the requirement for large amounts of labeled training data. Obtaining and labeling such datasets can be time-consuming and expensive. Another limitation is the potential for overfitting, where the model performs well on the training data but poorly on unseen data. This can occur when the model is too complex or when the training data is not representative of the real-world data. Additionally, CNN models can be sensitive to variations in image quality, lighting conditions, and viewpoint. These variations can affect the model's ability to accurately identify organisms [7].

Mollusks, a diverse and ecologically significant phylum, represent a challenging group for identification due to their vast species richness, morphological plasticity, and the frequent presence of cryptic species. Traditional methods of mollusk identification rely heavily on morphological characteristics, often requiring specialized taxonomic expertise and manual examination of specimens. These methods can be time-consuming, prone to subjectivity, and particularly challenging for closely related species exhibiting subtle morphological differences. This problem is a challenge for all shell collectors.

In this text we describe how a large shell image dataset was collected and processed to make ready for analysis by a CNN. At the time of writing images of more than 50 000 species are collected, with more 1 600 000 images in total.

Methods and Results

Data Collection

Shell images were collected from many online resources, from specialized websites on shell collecting to institutes and universities. One of the largest collections of shell images is available on GBIF. Also online marketplace such as ebay contain a large collection of images. Other large shell image collections are available at , Malacopics, Femorale and Thelsica. A shell dataset created for AI is available [8].

Some online resources have facilities to download images, but most websites require a specialized webscraper. Scrapy , an open source and collaborative framework for extracting the data from websites, is used to create a custom webscraper to extract images and their scientific names. All data was stored in a MySQL database before further processing was performed.

Preprocessing

All names were checked against WoRMS or MolluscaBase for their validity. Names that were not found in WoRMS/MolluscaBase were excluded for further processing. While a large part of this data quality step was automated, a manual verification (time-consuming) step was also included. In addition to text-based quality control, both automated and manual preprocessing steps were applied to the images. Shells were detected in all images and cut out of the original image, having only 1 shell on each image. Other objects on the raw images (labels, measures, hands holding a shell, etc.) were removed. When appropiate the background was changed to a uniform black background.

Dataset Composition

A total of 1 640 000 images were collected (status dec. 2024), representing 58 233 species. After elimination of all images where no valid scientific species name was found, 1 460 526 images are left for further processing. The dataset is unbalanced as shown in the chart below. Note that a logarithmic scale is used in the chart.

Because CNN models need a sufficient amount of images to provide a reliable result, species with less than 25 images were excluded. A total of 11,060 species have more than 25 images, representing many of the most commonly observed species. The dataset is still unbalanced, where the majority of the species have 25-50 images and a few species have 1000 or more images. This distribution reflects the occurrence of the species as well as their popularity among shell collectors. The table shows the number of images by family, and also the number of species among those images. The last column shows the total number of species for which at least 1 image was found. It clearly shows the families which are the most popular (high ratio between number of species with more than 25 images divided by the total number of images with at least 1 image).

Model selection

Recent research has shown that EfficientNetV2 models are a family of convolutional networks that achieve faster training speed and better parameter efficiency than previous models. This is achieved through a combination of training-aware neural architecture search and scaling, which jointly optimize training speed and parameter efficiency . This research also showed that EfficientNetV2 models can be trained efficiently at a large scale. They outperform Vision Transformers in terms of both accuracy and training efficiency. For example, by pretraining on the same ImageNet21k, EfficientNetV2 achieves 87.3% top-1 accuracy on ImageNet ILSVRC2012, outperforming the recent ViT by 2.0% accuracy while training 5x-11x faster using the same computing resources [9]. We are using Keras which has a large selection of CNN applications, including EfficientNetV2 models.

The results in the above table shows that the results between models are not significant different. A few EfficientNet models perform slightly worse, but the results do not help in the decision to select the optimal model to continue further training. Because performance metrics are similar, other factors like training time, computational resource requirements, and ease of deployment become crucial in deciding the best model. EfficientNet V2 models, for example, are designed to be more efficient and faster to train than their predecessors, which might be a deciding factor if performance differences are minimal. Among EfficientNet V2 models, a decision is made based on a balance between performance, training speed and model complexity. Considering these factors, EfficientNetV2 B2 was selected for further training and experimentation.

Class unbalance

Ideally, the dataset should contain an equal number of images for each class. However, in many seashell image datasets, this is not the case. Some classes may have significantly more images than others, leading to an unbalanced dataset. This imbalance can significantly impact the performance of the CNN model, causing it to be biased towards the majority class and perform poorly on the minority class.

Several techniques can be employed to address the challenges posed by unbalanced classes in image datasets including resampling, data augmentation, class weighting, cost-senitive learning, transfer learning, ensemble methods and threshold moving. The keras implementation of EfficientNet includes transfer learning with Imagenet. A model pre-trained on ImageNet can be fine-tuned on a smaller, imbalanced image dataset to improve its performance on the minority class.
Class weighting is also used when training the EfficienNet models which involves assigning different weights to different classes. Higher weights are assigned to the minority class to give it more importance during the learning process. This can help to counter the bias towards the majority class. The class weights are provided as a parameter when training the image dataset:

history = model.fit(train_data, epochs=epochs, validation_data=val_data, class_weight=classWeights, callbacks=[early_stopping, reduce_lr])

When dealing with unbalanced classes, also a good choice of the hyperparameters is important. The batch size used during training can significantly impact the model's ability to learn from the minority class. Ideally, each batch should contain multiple minority class examples. The batch size should be several times greater than the imbalance ratio. For example, if the imbalance ratio is 10:1, then the batch size should be at least 50. A batch size of 64 was used in all training described in this article.

Discussion

The results indicate that CNNs can be used to identify shells with high accuracy. However, the accuracy of the models is affected by the quality of the images and the number of images available for each species. Future work will focus on improving the accuracy of the models by using more data and by developing more robust methods for handling image quality variations.

One of the primary obstacles encountered in this study was the inherent imbalance in the dataset. The uneven distribution of images among species reflects both ecological realities and biases in data availability, with common or visually distinctive species being overrepresented. While species with fewer than 25 images were excluded to ensure model reliability, this exclusion limits the applicability of the model to rarer species. Addressing this imbalance through targeted data collection or synthetic augmentation methods, such as generating images using generative adversarial networks (GANs), could enhance the comprehensiveness of the dataset.

The evaluation of EfficientNetV2 models illustrates trade-offs between accuracy, computational efficiency, and scalability. While the differences in performance metrics among the tested models were not substantial, the selection of EfficientNetV2 B2 for further experimentation was guided by practical considerations, such as training speed and resource demands. This decision underscores the importance of balancing technical and logistical factors in model deployment, particularly when working with large-scale datasets.

The relatively high training and validation accuracies achieved indicate that the model effectively learned to distinguish among shell species. However, the potential for overfitting remains a concern, especially given the controlled preprocessing steps that standardized image backgrounds and removed extraneous elements. Future work could explore the model’s robustness to variations in image quality, lighting, and orientation to better simulate real-world conditions.

Despite its successes, this study highlights several areas for improvement. The reliance on large, labeled datasets remains a bottleneck, as does the exclusion of species with limited image availability. Additionally, the focus on image-based identification does not account for other diagnostic features, such as location or ecological data, which could provide complementary insights. Future research could explore multimodal approaches combining image data with other types of information.

Introduction

The performance of a CNN model is intricately linked to the number of images used in the training process. Generally, a larger and more diverse dataset leads to better performance, as the model can learn more robust and generalized features [1]. This is because CNNs learn by identifying patterns and features in the training data, and a larger dataset provides a more comprehensive representation of the underlying data distribution, allowing the model to learn more diverse and nuanced representations.

However, the relationship between image quantity and CNN performance is not always linear. While increasing the dataset size generally improves accuracy, there is a point of diminishing returns where adding more data provides only marginal gains. The optimal number of images required for training a CNN depends on several factors, including the complexity of the task, the variability within the dataset, the architecture of the CNN itself, and the number of classes [2].

When insufficient training data is available, CNN models are prone to overfitting. Overfitting occurs when the model learns the training data too well, capturing noise and specific details that do not generalize to unseen examples. This results in poor performance on new data, as the model fails to recognize the underlying patterns and features.
Limited data can also lead to class imbalance, where some classes have significantly fewer images than others. This imbalance can bias the model towards the majority classes, resulting in poor performance on the minority classes. The impact of class imbalance can be particularly severe in real-world scenarios where certain events or objects are inherently rare. Moreover, the influence of class imbalance on classification performance increases with the scale and complexity of the task [3, 4].

Methods

We used the Harpa genus dataset for these tests. For species Harpa major, 1228 images are available. When all images are used to create a model, a recall of 0.92 was obtained. We have created a model with a varying number of images of Harpa major, while the number of images for the other species of the genus Harpa stays the same. The images included in the dataset were selected at random, as well as 100 images for a test dataset. For the training and validation dataset a 80-20 split was used.

Results

The figure below illustrates the relationship between the number of images of Harpa major and the recall metric. Between 5 and 1000 images were included in the dataset to create a model to identify the species of the Harpa genus. The dataset included 12 species, each represented by 38 to 833 images. The amount of images in all classes stays the same for each run, only for Harpa major the number of images varies between runs.

When only 5 images of Harpa major were included, the recall was 0.06, an unacceptably low value given the dataset's 12 species. Recall values consistently exceeded 0.5 when at least 25 images were included. To achieve maximum recall, at least 300 images were needed, with additional images providing only marginal improvements beyond this point.

Discussion

The tradeoff between including a species in the dataset with a limited number of images and ensuring adequate model performance is critical, because a CNN model will always identify a species, but often with a low probability for unknown species. Therefore a threshold of 25 images per species is decided with is a good balance between model performance and species completeness.
It is important to note that these results may vary for other genera, as they depend on factors such as the total number of images, the type of seashells, and image quality. Future studies should evaluate these variables to develop more generalized guidelines for dataset preparation.