Stanford Dog Dataset

Dataset Overview

Dataset Name

• Stanford Dog Dataset

Dataset Content

• Contains approximately 20,000 images covering 120 different dog breeds.
• Each image includes corresponding annotation information.

Dataset Usage

• Due to the limited data volume, transfer learning techniques are employed, using pretrained VGG16 and VGG16BN models.
• By replacing the top fully‑connected layer and softmax layer, freezing the remaining layers, and employing synthetic image generation to increase image variability.

Dataset Processing

• The dataset is split into training, validation, and test sets, ensuring no class imbalance across subsets.
• Data format is converted to tfRecords to accelerate I/O operations.

Training Results

• After 25 epochs of training the initial CNN, training accuracy reached 94.07% and test accuracy 51.07%.
• Using the VGG16 model for 50 epochs yielded training accuracy 97.8% and test accuracy 40.23%.
• Applying Dropout and Batch Normalization resulted in training accuracy 88.23% and test accuracy 76.53% after 20 epochs.

Prediction and Analysis

• The YOLO algorithm is used for object detection; accuracy is evaluated by comparing predicted boxes with ground‑truth boxes using Intersection over Union (IoU).
• Confusion matrix analysis reveals that “Silky Terrier / Yorkshire Terrier” is the most frequently misclassified dog breed pair.