Classification of butterfly species with improved models based on transfer learning method of convolutional neural networks

Technical Summary: Classification of Butterfly Species with Improved Models Based on Transfer Learning

Problem Statement
The accurate identification and classification of butterfly species are critical for biodiversity conservation, environmental monitoring, and scientific research. Traditional identification methods rely on manual examination of morphological characteristics (wing features, color, patterns), which is time-consuming, labor-intensive, and prone to human error. While deep learning and Convolutional Neural Networks (CNNs) offer automated solutions, existing literature often suffers from limitations such as reliance on small or ready-made datasets, the direct application of standard transfer learning without architectural innovation, and a lack of robustness testing across diverse datasets with varying numbers of species.

Methodology
The study proposes a two-pronged approach to butterfly species classification using deep learning:

1. Dataset Construction: A novel dataset was curated comprising 5,624 images across 49 butterfly species. Images were sourced from various online platforms and nature photographers, capturing diverse lighting conditions, angles (dorsal and ventral views), and environments. To address class imbalance, three species with insufficient imagery were excluded. The remaining data was split into training (4,643 images), validation (735 images), and test (246 images) sets. Preprocessing included cropping to remove background noise, resizing to 224x224 pixels, normalization, shear transformation, and horizontal flipping.

2. Model Architectures: Eight pre-trained CNN architectures were evaluated: VGG16, VGG19, Xception, InceptionV3, NASNetLarge, InceptionResNetV2, ResNet152V2, and DenseNet201.

3. Approach 1: Standard Transfer Learning (TL): The authors applied standard transfer learning by utilizing pre-trained weights from ImageNet. In this configuration, the base networks were frozen, a flattening layer was added, and a final Softmax output layer was attached. The models were trained for 50 epochs using the Adam optimizer, a learning rate of 0.0001, and a batch size of 32.

4. Approach 2: GDD-TL (Global Average Pooling, Dense, Dropout Transfer Learning): To improve performance, a modified architecture was proposed. This method retained the pre-trained base but inserted three specific layers between the feature extraction base and the final classification layer:

   • Global Average Pooling (GAP): To reduce feature maps to single values, acting as structural regularization.
   • Dense Layer: A fully connected layer with 1,024 units and ReLU activation to learn complex non-linear relationships.
   • Dropout Layer: A 20% dropout rate to prevent overfitting.
   • Output Layer: A Dense layer with 49 units and Softmax activation for multi-class classification.

Key Contributions

• Novel Dataset: Creation of a diverse, high-quality dataset of 49 species (5,624 images) with varied environmental conditions, distinct from many studies relying solely on existing Kaggle datasets.
• Architectural Innovation: The proposal of the GDD-TL method, which integrates Global Average Pooling, a Dense layer, and Dropout into standard transfer learning pipelines to enhance feature learning and generalization.
• Comprehensive Evaluation: A comparative analysis of eight CNN architectures across two distinct methodologies (TL vs. GDD-TL) using multiple metrics (accuracy, precision, recall, F1-score, error rate).
• Robustness Validation: Testing the proposed best-performing model on a second, larger external dataset (13,604 images, 100 species) from the Kaggle platform to verify generalization capabilities.

Experimental Results

• Standard Transfer Learning: Among the eight architectures, DenseNet201 achieved the highest performance, reaching a test accuracy of 91.09% and a validation accuracy of 90.73%.
• GDD-TL Method: The addition of the GAP, Dense, and Dropout layers significantly improved performance for most models. DenseNet201 with GDD-TL achieved a test accuracy of 94.33% (up from 91.09%) and a validation accuracy of 91.28%. The error rate decreased from 8.91% to 5.67%.
• Cross-Dataset Validation: When applied to the larger Kaggle dataset (100 species), the GDD-TL DenseNet201 model achieved a test accuracy of 95.20% (with 97.73% training accuracy and 94.19% validation accuracy). The model demonstrated strong generalization, outperforming several state-of-the-art models previously applied to similar datasets.
• Comparative Analysis: The proposed GDD-TL DenseNet201 model outperformed four selected state-of-the-art studies on the authors' generated dataset, achieving the highest scores in test accuracy, precision, recall, and F1-score.

Significance and Claims
The paper claims that the proposed GDD-TL method is an effective and robust tool for butterfly species classification. The authors assert that while standard transfer learning yields good results, the strategic addition of specific layers (Global Average Pooling, Dense, Dropout) enhances the model's ability to generalize, particularly when dealing with increased numbers of species and images. The study highlights that the DenseNet201 architecture, when combined with the GDD-TL approach, offers superior performance due to its dense connectivity pattern, which facilitates feature reuse and mitigates the vanishing gradient problem. The authors conclude that their model is suitable for real-world applications in nature conservation and photography, noting that the increase in dataset size did not negatively impact the model's generalization ability. Future work involves developing offline mobile and desktop applications using Kivy and PyQt5 to facilitate species identification in areas without internet access.