BIRDS 525 SPECIES 525 Bird Image Dataset

Dataset Overview

The dataset contains 525 bird species, 84,635 training images, 2,625 test images, and 2,625 validation images.

Data cleaning and quality assurance

• De-duplication and denoising: Use analytical tools to clean the dataset and remove duplicate or near-duplicate images, as well as defective and low-information images.
• Dataset isolation: Ensure that there is no information leakage between the training, testing, and validation datasets.

Dataset characteristics

• Image Quality: The images are original and unenhanced, with only one bird in each image, usually occupying at least 50% pixels.
• Expected performance: Models of medium complexity are expected to achieve training and test accuracy of about 90%.

Technical specifications

• Image size: All images are in 224 X 224 X 3 color JPG format.
• Dataset structure: Includes training set, test set and validation set, each set contains 525 sub-directories, each sub-directory corresponds to a bird species.

Recommendations for using the dataset

• Data Generator: It is recommended to use Keras ImageDataGenerator.flow_from_directory to create the data generator.
• Supporting Files: The dataset includes a bird.csv File containing image path, label, scientific name, dataset type, and class index value.

Data Collection and Processing

• Image source: Collected through Internet search, checked and deleted duplicate or near-duplicate images after downloading.
• Image Processing: Crop and resize the image to ensure that the bird image occupies at least 50% pixels.

Dataset limitations

• Image size recommendations: It is recommended to use an image size of 150 X 150 X 3 to reduce training time.
• Document No.: All files are numbered by species, and training images are padded with zeros to maintain order.
• Imbalanced dataset: The number of images of each species in the training set varies, but there are at least 130 images.
• Gender bias: About 80% of the images are male and 20% are female, which may cause the classifier to perform poorly on female images.