Project 3

1. Data collection and organization

As an initial test, I added two brands: Stussy and Chrome Hearts. I collected the images off of pinterest and added them to their respective folders within a raw dataset.

Since all the images came in different sizes/formats, I created a super short script that automates processing to save me time:

import os
from PIL import Image

# Constants
IMG_SIZE = (224, 224)
IMAGE_FORMAT = 'JPEG'

def preprocess_image(image_path, save_path):
    with Image.open(image_path) as img:
        # Resize the image to (224, 224)
        img = img.resize(IMG_SIZE)
        # Convert the image to RGB format
        img = img.convert('RGB')
        # Save the processed image
        img.save(save_path, format=IMAGE_FORMAT)
        print(f'Processed image saved to {save_path}') # For debugging

def preprocess_directory(input_dir, output_dir):
    # Ensure output directory exists
    os.makedirs(output_dir, exist_ok=True)
    
    for root, _, files in os.walk(input_dir):
        for filename in files:
            if filename.lower().endswith(('.jpg', '.jpeg', '.png')): # Check if the file is an image
                brand = os.path.basename(root)
                brand_output_dir = os.path.join(output_dir, brand)
                os.makedirs(brand_output_dir, exist_ok=True)
                input_path = os.path.join(root, filename)
                output_path = os.path.join(brand_output_dir, filename)
                preprocess_image(input_path, output_path)
            

2. Model Training

Using Tensorflow/Keras, I pre-trained a ResNet50 model, and fine-tuned it with custom layers to adapt to the specific task of ID recognition.

I implemented a script as well which essentially just evaluated test loss and accuracy

First eval:

So... not very good

At it's first run, the model achieved a 50% test accuracy, which is paramount to just random guessing, but this is mainly due to the small sample size of images I started off with. Ideally, having 500-600 images per brand/ID is ideal, and I started off with around 15 each to start. My main struggle here is finding a way to automate gathering the hundreds of images off of pinterest/grailed.

3. Creating the GUI

Next, I used customtkinter to create a GUI where the user can select an image off of their computer to evaluate, which worked perfectly.

Here's a screenshot of the GUI:

The user selects an image, and the model is evaluated, and the result is displayed.

4. Future Work

There is still a lot more work to be done, and I'll gradually update this as I work more on it, but these are the key aspects I plan on implementing:

Automate the image collection process using some API

Create larger sample sizes, ideally 500-600 images per ID

Improve data augmentation by including more transforms (ie. rotation, scaling, noise)

Improve the model by unfreezing specific layers or maybe using different pre-trained models