bird_classification

数据集介绍 一、数据来源 本实验所使用的数据集源自飞桨AI Studio开源数据集平台“275种鸟类数据集”（https://aistudio.baidu.com/datasetdetail/99214）。为满足“50种鸟分类”的课题要求，我们从原始275个鸟类物种中筛选出前50个类别，构建本实验的子数据集。 二、数据规模与划分 经筛选后，数据集总计包含 7,402张 鸟类图像。为保证模型训练的有效性和评估的公平性，我们采用以下划分策略： 数据集 图片数量 占比 用途说明 训练集 5,721张 77.3% 用于模型参数学习 验证集 1,431张 19.3% 用于超参数调优和模型选择 测试集 250张 3.4% 用于最终模型性能评估 具体说明： 训练集：从原始训练数据中按8:2比例划分，80%（5,721张）用于训练，20%（1,431张）作为验证集 验证集：独立于训练过程，用于监控过拟合、调整学习率等超参数 测试集：保留原始数据集的测试划分（每类5张 × 50类 = 250张），仅在最终评估时使用，确保模型未曾见过这些数据 三、图像规格 图像格式：所有图片均为 JPG 格式 图像尺寸：统一为 224 × 224 × 3 的彩色图像 主体占比：鸟类通常占据图像至少50%的像素区域，减少了复杂背景的干扰 四、数据特点与挑战 1. 细粒度分类挑战 不同种类的鸟类在外观（颜色、纹理、体型）上极其相似，仅喙部形状或羽毛纹路存在细微差别，对模型的细节特征提取能力要求较高。 2. 类别平衡性 训练集中每类鸟类约 110-120张 图片 各类别样本数量相对均衡，不存在严重的类别不平衡问题 3. 姿态与背景多样性 数据集包含鸟类的多种姿态（站立、飞翔、觅食）和多种环境背景（树林、水域、天空），有助于模型学习鲁棒性特征。 4. 性别不平衡 数据集中雄性鸟类图像占比约80%，雌性仅占20%。雄性羽毛通常更鲜艳，模型可能偏向学习雄性特征，对雌性个体识别能力相对较弱。Dataset Introduction ### 1. Data Source The dataset used in this experiment is sourced from the "275 Bird Species Dataset" on the PaddlePaddle AI Studio open-source dataset platform (https://aistudio.baidu.com/datasetdetail/99214). To meet the requirements of the "50-Bird Species Classification" project, we selected the top 50 categories from the original 275 bird species to construct the sub-dataset for this experiment. ### 2. Data Scale and Partition After screening, the dataset contains a total of 7,402 bird images. To ensure the effectiveness of model training and the fairness of evaluation, we adopted the following partitioning strategy: | Dataset | Number of Images | Proportion | Purpose Description | |------------------|------------------|------------|--------------------------------------| | Training Set | 5,721 | 77.3% | Used for model parameter learning | | Validation Set | 1,431 | 19.3% | Used for hyperparameter tuning and model selection | | Test Set | 250 | 3.4% | Used for final model performance evaluation | Specific Instructions: 1. Training Set: Divided from the original training data at an 8:2 ratio, with 80% (5,721 images) allocated for training and 20% (1,431 images) as the validation set. 2. Validation Set: Independent of the training process, used for monitoring overfitting, adjusting hyperparameters such as learning rate, and other related tasks. 3. Test Set: Retains the original test split of the dataset (5 images per class × 50 classes = 250 total images), which is only used for final evaluation to ensure that the model has never encountered these data during training. ### 3. Image Specifications - Image Format: All images are in JPG format. - Image Size: Unified as 224 × 224 × 3 color images. - Subject Proportion: Birds typically occupy at least 50% of the pixel area in each image, reducing interference from complex backgrounds. ### 4. Data Characteristics and Challenges 1. Fine-grained Classification Challenge: Different bird species exhibit high similarity in appearance (color, texture, body shape), with only subtle differences in beak shape or feather patterns, which imposes high requirements on the model's ability to extract detailed features. 2. Class Balance: Each bird species in the training set has approximately 110–120 images. The sample size per class is relatively balanced, with no severe class imbalance issue. 3. Diversity of Poses and Backgrounds: The dataset covers multiple poses of birds (standing, flying, foraging) and various environmental backgrounds (forests, waters, skies), which helps the model learn robust features. 4. Gender Imbalance: Approximately 80% of the images in the dataset are of male birds, while only 20% are of female birds. Male birds usually have more vibrant plumage, so the model may tend to learn male-specific features, leading to relatively weaker recognition performance for female individuals.