Author: Vincent Yeo

Image Recognition with Fashion MNIST dataset

Background:

Fashion e-commerce has grown to lucrative business, driven by the accessibility of internet into the global market, connecting wholesalers and consumers via browser. Combined with search engines and mobile devices, e-commerce platform can drives the sales further by allowing consumers to search for products by uploading images of the desired items, taken from their smart phones. Thus, a model to recognise images of clothing would be useful for ecommerce to deliver their service to consumers better to drive sales.

Dataset:

The Fashion MNIST serves as a replacement for the MNIST dataset as benchmark for machine learning algorithms. It contains the similar 28x28 grayscale images associated with 10 classes of objects.

Task:

To build a Convolutional Neural Network (CNN) to classify 10 categories of fashion items from their images

Results & Analysis

Model achieves 80.73% on unseen test data

Based on each precision score, the model's prediction would be realised truely for

• the T-shirt/top, 80% of the time.
• the Trouser, 97% of the time.
• the Pullover, 76% of the time.
• the Dress, 82% of the time.
• the Coat, 58% of the time.
• the Sandal, 93% of the time.
• the Shirt, 55% of the time.
• the Sneaker, 89% of the time.
• the Bag, 94% of the time.
• the Ankle Boot, 90% of the time.

Based on each recall score, the model's prediction for

• the T-shirt/top has about 74% being correct.
• the Trouser has about 95% being correct.
• the Pullover has about 58% being correct.
• the Dress has about 83% being correct.
• the Coat has about 83% being correct.
• the Sandal has about 90% being correct.
• the Shirt has about 50% being correct.
• the Sneaker has about 90% being correct.
• the Bag has about 92% being correct.
• the Ankle Boot has about 93% being correct.

Based on each f1 score, the weighted average of the precision and recall for

• the T-shirt/top is 77%
• the Trouser is 96%
• the Pullover is 66%
• the Dress is 83%
• the Coat is 68%
• the Sandal is 91%
• the Shirt is 53%
• the Sneaker is 90%
• the Bag is 93%
• the Ankle Boot is 91%

Future Works:

• Model can be trained with more specific subclasses of each categorized items for more specific fashion items to be displayed to the consumers.