QH-Bird: A Fine-Grained Bird Visual Dataset and Hierarchical Consistency Denoising Method

To address the lack of bird image datasets in the field of fine-grained image recognition, this paper introduces a new fine-grained bird image dataset and a deep learning-based denoising method. The dataset encompasses 50,000 images across 200 species. Initially, we constructed an initial bird image dataset by sourcing and compiling images from specialized books and online resources. Subsequently, we employed image deduplication methods based on difference hashing (dHash) and structural similarity index (SSIM) to clean duplicate images from the dataset. Following this, we applied a deep learning-based hierarchical consistency learning method to identify noisy web images, using classification loss and semantic consistency prediction. To further enhance the dataset's diversity and balance, we expanded the number of images for each bird species using data augmentation techniques such as rotation, mirroring, random cropping, and adjustments in brightness and contrast. Ultimately, we constructed a high-quality, balanced, and diverse bird image dataset, named QH-Bird, which serves as a valuable resource for research and applications in fine-grained image recognition. We evaluated our dataset using four deep learning models: VGG-16, ResNet-50, EfficientNet, and HCLNet, achieving accuracies of 87.17%, 88.59%, 90.28%, and 91.59%, respectively.