src/app.py

from colordescriptor import ColorDescriptor
from CLIP import CLIPImageEncoder
from LBP import LBPImageEncoder
from helper import chi2_distance, euclidean_distance, merge_features
import gradio as gr
import os
import cv2
import numpy as np
from datasets import *


dataset = load_dataset("nielsr/CelebA-faces", download_mode='force_redownload')
dataset.cleanup_cache_files()
candidate_subset = dataset["train"].select(range(500)) # This is a small CBIR app! :D

def index_dataset(dataset):
    print(dataset)

    print("LBP Embeddings")
    lbp_model = LBPImageEncoder(8,2)
    dataset_with_embeddings = dataset.map(lambda row: {'lbp_embeddings': lbp_model.describe(row["image"])})

    print("Color Embeddings")
    cd = ColorDescriptor((8, 12, 3))
    dataset_with_embeddings = dataset_with_embeddings.map(lambda row: {'color_embeddings': cd.describe(row["image"])})

    print("CLIP Embeddings")
    clip_model = CLIPImageEncoder()
    dataset_with_embeddings = dataset_with_embeddings.map(clip_model.encode_images, batched=True, batch_size=16)

    print("LBP and Color")
    dataset_with_embeddings = dataset_with_embeddings.map(lambda row: {'lbp_color_embeddings': merge_features(row['lbp_embeddings'], row['color_embeddings'])})


    # Add index
    dataset_with_embeddings.add_faiss_index(column='color_embeddings')
    dataset_with_embeddings.save_faiss_index('color_embeddings', 'color_index.faiss')
    dataset_with_embeddings.add_faiss_index(column='clip_embeddings')
    dataset_with_embeddings.add_faiss_index(column='lbp_embeddings')
    dataset_with_embeddings.save_faiss_index('clip_embeddings', 'clip_index.faiss')


    print(dataset_with_embeddings)
    return dataset_with_embeddings


def check_index(ds):
    index_path = "my_index.faiss"
    if os.path.isfile('color_index.faiss') and os.path.isfile('clip_index.faiss'):
        ds.load_faiss_index('color_embeddings', 'color_index.faiss')
        return ds.load_faiss_index('clip_embeddings', 'clip_index.faiss')

    else:
        return index_dataset(ds)


dataset_with_embeddings = check_index(candidate_subset)

# Main function, to find similar images
# TODO: implement different distance measures

def get_neighbors(query_image, selected_descriptor, selected_distance, top_k=5):
    """Returns the top k nearest examples to the query image.

    Args:
        query_image: A PIL object representing the query image.
        top_k: An integer representing the number of nearest examples to return.

    Returns:
        A list of the top_k most similar images as PIL objects.
    """
    if  "Color Descriptor" == selected_descriptor:
        cd = ColorDescriptor((8, 12, 3))
        qi_embedding = cd.describe(query_image)
        qi_np = np.array(qi_embedding)
        if selected_distance == "FAISS":
            scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
                'color_embeddings', qi_np, k=top_k)
        elif selected_distance == "Chi-squared":
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding,row['color_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        else:
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding,row['color_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        images = retrieved_examples['image'] #retrieved images is a dict, with images and embeddings
        return images
    if "CLIP" == selected_descriptor:
        clip_model = CLIPImageEncoder()
        qi_embedding = clip_model.encode_image(query_image)
        if selected_distance == "FAISS":
            scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
                'clip_embeddings', qi_embedding, k=top_k)
        elif selected_distance == "Chi-squared":
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['clip_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        else:
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['clip_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        images = retrieved_examples['image']
        return images
    if "LBP" == selected_descriptor:
        lbp_model = LBPImageEncoder(8,2)
        qi_embedding = lbp_model.describe(query_image)
        if selected_distance == "FAISS":
            scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
                'lbp_embeddings', qi_embedding, k=top_k)
        elif selected_distance == "Chi-squared":
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['lbp_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        else:
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['lbp_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        images = retrieved_examples['image']
        return images
    if "LBPColor" == selected_descriptor:
        lbp_model = LBPImageEncoder(8,2)
        cd = ColorDescriptor((8, 12, 3))
        qi_embedding = merge_features(lbp_model.describe(query_image), cd.describe(query_image))
        if selected_distance == "FAISS":
            scores, retrieved_examples = dataset_with_embeddings.get_nearest_examples(
                'lbp_color_embeddings', qi_embedding, k=top_k)
        elif selected_distance == "Chi-squared":
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': chi2_distance(qi_embedding, row['lbp_color_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        else:
            tmp_dataset = dataset_with_embeddings.map(lambda row: {'distance': euclidean_distance(qi_embedding, row['lbp_color_embeddings'])})
            retrieved_examples = tmp_dataset.sort("distance")[:5]
        images = retrieved_examples['image']
        return images


    else:
        print("This descriptor is not yet supported :(")
        return []


# Define the Gradio Interface

with gr.Blocks() as demo:
    gr.Markdown("""
    # Welcome to this CBIR app

    This is a CBIR app focused on the retrieval of similar faces.

    ## Find similar images
    Here you can upload an image, that is compared with existing image in our dataset.
    """)
    with gr.Row():
        image_input = gr.Image(type="pil", label="Please upload your image")
        gallery_output = gr.Gallery()

    btn = gr.Button(value="Submit")
    gr.Markdown("""
    ## Settings
    Here you can adjust how the images are found
    """)

    with gr.Row():
        descr_dropdown = gr.Dropdown(["Color Descriptor", "LBP", "CLIP", "LBPColor"], value="LBP", label="Please choose an descriptor")
        dist_dropdown = gr.Dropdown(["FAISS", "Chi-squared", "Euclid"], value="FAISS", label="Please choose a distance measure")

    btn.click(get_neighbors, inputs=[image_input, descr_dropdown, dist_dropdown], outputs=[gallery_output])


if __name__ == "__main__":
    demo.launch()