tSNE.py

from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.resnet50 import preprocess_input
import numpy as np
import os
from sklearn.manifold import TSNE
from sklearn.decomposition import PCA
from PIL import Image
import pandas as pd
import pickle
import matplotlib.pyplot as plt
from matplotlib.offsetbox import OffsetImage, AnnotationBbox
from sklearn.cluster import DBSCAN

# create the base pre-trained model
base_model = ResNet50(weights='imagenet', include_top=False)

# Lists for data collection
all_features = []
files_column = []
files_onlyname_column = []

#   Read each image and extract ResNet50 features
for root, dirs, files in os.walk('./data/images/images/images/'):
    for name in files:
        if name.endswith((".jpg")):
            filename = "{}/{}".format(root, name)

            #   Load image and resize
            img = image.load_img(filename, target_size=(224, 224))
            img_array = image.img_to_array(img)
            img_array = np.expand_dims(img_array, axis=0)

            #   Preprocess input
            preprocessed = preprocess_input(img_array)

            #   Extract features and flatten in array (Shape is (7,7,2048) )
            features = base_model.predict(preprocessed, batch_size=1)
            flatten_features = features.flatten()
            all_features.append(flatten_features)

            files_column.append(filename)
            files_onlyname_column.append(name)

            #   Save thumbnail
            image_resize = Image.open(filename)
            image_resize.thumbnail((64, 64), Image.ANTIALIAS)
            image_resize.save("./data/images/resized_leaflet/{}".format(name), "JPEG")

#   Convert all the features in a Numpy Array
all_features = np.asarray(all_features)
# pickle.dump(all_features,open('features.pkl', 'wb'))
# all_features = pickle.load(open('features.pkl', 'rb'))

#   Apply PCA to reduce features to 2048
embedded_features = PCA(n_components=2048).fit_transform(all_features)

#   Apply tSNE
embedded_features = TSNE(n_components=2, verbose=3, perplexity=25).fit_transform(embedded_features)

#   Find Clusters using DBSCAN
dbscan = DBSCAN(eps=1.5, min_samples=15).fit(embedded_features)

colors = ["black", "brown", "chocolate", "gold", "forestgreen", "teal",
          "dodgerblue", "navy", "darkviolet", "deeppink"]

#   Create data frame and print the results using matplotlib
df = pd.DataFrame()
df["files"] = files_column
df["files_names"] = files_onlyname_column
df["x"] = embedded_features[:, 0]
df["y"] = embedded_features[:, 1]
df["labels"] = dbscan.labels_
df.to_csv("output.csv", index=False)
df = df.sort_values(by=["labels"])

x = df["x"].values * 200
y = df["y"].values * 200

fig, ax = plt.subplots(figsize=(150, 150))
ax.scatter(x, y)


def getImage(path, alpha=1):
    return OffsetImage(plt.imread(path), alpha=alpha)


for x0, y0, path, l in zip(x, y, df["files_names"].values, df["labels"].values):
    if l == -1:
        ab = AnnotationBbox(getImage("{}/{}".format("data/images/resized-leaflet", path), 0.1), (x0, y0),
                            frameon=True, pad=0)
    else:
        ab = AnnotationBbox(getImage("{}/{}".format("data/images/resized-leaflet", path)), (x0, y0),
                            frameon=True, bboxprops=dict(edgecolor=colors[l % len(colors)], lw=10), pad=0)

    ax.add_artist(ab)
plt.savefig('foo.png')

#   Output as JSON
df[df["labels"] != -1].to_json("output.json", orient="records")