Image Similarity AI Model

(Image Generated using bing ai)

Introduction

The primary goal of this project is to develop an AI model that can determine the most similar images within a given dataset. This is achieved by first training a classification model on the dataset. After successfully training the model, I use the output from the last hidden layer as a vector representation of each image. By employing cosine similarity, I can then compare these vectors to find the most similar images in the dataset.

Please note that this project is purely a concept, and there is a possibility that it might not achieve the desired results. My intention is to explore the concept and gain valuable insights, regardless of the final outcome.

The insights gained from this attempt, even if the project does not succeed, will be valuable for future iterations and developments in image comparison and similarity analysis. I welcome other researchers and developers to join me in this exploratory journey, as we learn from both successes and failures.

I hope this concept will inspire curiosity and creativity among the AI community. Should you choose to be a part of this exciting endeavor, I welcome your contributions and look forward to working together.

Thank you for your interest in my Image Similarity AI Model project, and let's embark on this challenging yet rewarding exploration together!

Table of Contents

• Introduction
• Data
• Preprocessing
• Architecture
• Installation
• Usage
• Dataset
• Training
• Evaluation
• Results
• Contributing
• License

Data

The images were downloaded using Download All Images

Preprocessing

The Data was First Augmented using Albumentations Pipeline.

The below filters were applied to Images to augment them:

• Horizontal Flip
• Random Brightness Contrast
• Random Rotate 90
• Flip
• RGB Shift
• Hue Saturation Value
• CLAHE
• Random Gamma
• Blur
• To Gray
• Image Compression

The image was splitted into Train-Test-Valid sets with 75:10:15 ratio

Architecture

This model uses a Convolutional Layers of MobileNet

and then adds

• Global Flatten Layer
• Dense Layer with 512 outputs and ReLU activation
• Dense Layer with 64 outputs and ReLU activation
• Dense Layer with 1 output and Sigmoid activation

This Model used Categorical Cross Entropy Loss to learn with an adam optimizer

Installation

You first need to install Tensorflow

To install TensorFlow using pip, run the following command:

pip install tensorflow

To install TensorFlow using conda, run the following command:

conda install tensorflow

Then you have to download the model

Download The model here

Then paste it in your project folder.

Usage

To use the model

import tensorflow as tf
from tensorflow import keras
from keras.models import load_model
from keras.utils import load_img

#load dependencies

model = load_model('model.h5')
# loading the model 

def predict(image:str, model:keras.models.Model):
    img = tf.convert_to_tensor(Image.open(image).resize((256, 256)))
    prediction = model.predict(tf.expand_dims(img, 0), verbose=False)
    return prediction

def predict_multiple(images:list[str], model:keras.models.Model):
    imagesTensor = []
    for img in images:
        imagesTensor.append(tf.convert_to_tensor(Image.open(img).resize((256, 256))))
    imagesTensor = tf.stack(imagesTensor)
    
    return model.predict(imagesTensor, verbose=False)

def find_similar(target:str, dataset:list[str], model:keras.models.Model):
    targetVector = predict(target, model)
    dataMatrix = predict_multiple(dataset, model)

    similarity = cosine_similarity(targetVector, dataMatrix)[0]

    print(similarity)

    dataset = zip(dataset, similarity.tolist())

    return sorted(dataset, key=lambda x: x[1], reverse=True)

find_similar('image/path', ['list', 'of', 'images', 'path', 'to', 'compare', 'similarity'], model)

Training

The Model was trained with only 5 epochs

Evaluation

The model Performed Exceptionally well it training and testing

Results

[Will add the results of my experiments here]

Contributing

You can give your comments in the discussions section or you can email me at ivanrj7j@gmail.com, you can also fork this repo if you want.

License

This project is Licensed under MIT license.

---

Thank you for showing interest in my Image Similarity AI Model project! I hope this concept will pave the way for exciting advancements in image comparison and similarity analysis. Should you have any questions or suggestions, feel free to reach out to me.