Fashion Product Multilabel Classification

This project focuses on the classification of fashion products based on their visual features using deep learning models. The goal is to build a robust multilabel classification system that can identify various attributes of fashion items, such as category, color, and brand.

You can check out the notebook from files section or from Kaggle.

Dataset

The dataset used in this project is the Fashion Product Images Dataset. It contains images of various fashion products along with their corresponding labels. The dataset is diverse, featuring multiple categories and attributes that are crucial for training and evaluating the classification models.

Finding the Ideal Base Model

Model	Accuracy	Precision & Recall
Base Model: InceptionV3 Epochs: 100 Augmentation: No Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Low
Base Model: EfficientNetB0 Epochs: 100 Augmentation: No Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Low
Base Model: VGG-16 Epochs: 100 Augmentation: No Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Low
Base Model: ResNet-50 Epochs: 100 Augmentation: No Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Low

Image Augmentation

Pipeline

The augment_image function applies the following transformations to enhance image variability:

• Random Scaling: Zooms in and out by up to 35%.
• Random Resized Crop: Crops and resizes the image to the target dimensions with a scale range of 50% to 100%.
• Horizontal Flip: Flips the image horizontally with a 75% probability.
• Rotation: Rotates the image randomly by up to 15 degrees.
• Brightness and Contrast Adjustment: Slightly alters brightness and contrast to simulate different lighting conditions.

Augmentation Examples

Result

Model	Accuracy	Precision & Recall
Base Model: ResNet-50 Epochs: 100 Augmentation: No Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Low
Base Model: ResNet-50 Epochs: 100 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Medium

Regularization

After applying image augmentation, the model improves but overfits, but we can address this issue by incorporating L2 regularization to enhance generalization.

Result

Model	Accuracy	Precision & Recall
Base Model: ResNet-50 Epochs: 100 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: No Input Resolution: Medium
Base Model: ResNet-50 Epochs: 50 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: L2 (0.1) Input Resolution: Medium
Base Model: ResNet-50 Epochs: 50 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: L2 (0.05) Input Resolution: Medium
Base Model: ResNet-50 Epochs: 50 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: L2 (0.02) Input Resolution: Medium
Base Model: ResNet-50 Epochs: 50 Augmentation: Yes Learning Rate: 1e-5 Optimizer: SGD Momentum: 0.9 Batch Size: 32 Regularization: L2 (0.01) Input Resolution: Medium Download Model

Examples

From Train Set

From Test Set