Tensorflow를 활용한 유사 이미지 추출하기

Tensorflow를 활용해 학습한 이미지 중 선택한 이미지와 가장 유사한 5개의 이미지를 출력해보자

from tensorflow.keras.applications import vgg16
# resnet, inception ... 등 다양한 모델 적용해볼 예정
from tensorflow.keras.applications.imagenet_utils import preprocess_input
from tensorflow.keras.preprocessing.image import load_img,img_to_array
from tensorflow.keras.models import Model

import os
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity

vgg_model = vgg16.VGG16(weights='imagenet')
#imagenet 가중치 가져오기 
model = Model(inputs=vgg_model.input, outputs=vgg_model.get_layer("fc2").output)
# output 부분 fully-connected를 추가 ! 
# 특징을 추출해서 유사한 이미지를 찾는 것이 목적 . 
# label 예측 XXXXXXXX
model.summary()

img_path = 'style/'

img_width, img_height = 224, 224

most_similar = 5
# 유사한 아이템 몇개를 뽑을것인지 
files = [img_path + x for x in os.listdir(img_path) if 'png' in x]
  
image_1 = load_img(files[0], target_size=(img_width, img_height, 3))
# (img_width, img_height, 3) 3채널이 아닌 4채널로 나오기에 3채널로 변경
plt.imshow(image_1)
plt.show()

img_array = img_to_array(image_1)
# 이미지 array 읽어오기

img_batch = np.expand_dims(img_array, axis=0)
#(batch_size, img_width, img_height, channel)으로 만들기 위함
# batch_size = None이면 모델이 작동하지 않음
print('image_batch_size=',img_batch.shape)
# 4_channel -> 3_channel로 변경
processed_image = preprocess_input(img_batch)
# tf.kears.applications.[model] import preprocess_input
# 불러온 모델에 가장 잘 맞게 이미지를 맞추기위함.
processed_image

img_feature = model.predict(processed_image)
print('number of image features: ',img_feature.size)
img_feature

image_list = []
for f in files:
    file_name = f
    image = load_img(file_name, target_size=(224, 224, 3))
    img_array = img_to_array(image)
    img_batch = np.expand_dims(img_array, axis=0)
    
    image_list.append(img_batch)

images = np.vstack(image_list)
# vstack -> 세로로 합치기 
# hstack -> 가로로 합치기
processed_imgs = preprocess_input(images)

img_feature = model.predict(processed_imgs)

cosine = cosine_similarity(img_feature)
#예측 값으로 유사도 구하기
cos_df = pd.DataFrame(cosine, columns=files, index=files)
#pd.DataFrame(value, columns, index)
#files = img_name
# index, column -> 이미지이름으로 하고 예측했던 값들로 value
cos_df

def similarity_image(img):
    print('-'*30)
    print('selected product:')
    image = load_img(img, target_size=(img_width, img_height, 3))
    plt.imshow(image)
    plt.show()
    
    print('-'*30)
    print('similar products:')
    cosine_img = cos_df[img].sort_values(ascending=False)[1:most_similar+1].index
    
    cosine_img_score = cos_df[img].sort_values(ascending=False)[1:most_similar+1]
    
    for i in range(0, len(cosine_img)):
        image_cosine = load_img(cosine_img[i], target_size=(img_width, img_height, 3))
        plt.imshow(image_cosine)
        plt.show()
  
similarity_image(files[5])

# 코드
from tensorflow.keras.applications import vgg16
from tensorflow.keras.applications.imagenet_utils import preprocess_input
from tensorflow.keras.preprocessing.image import load_img,img_to_array
from tensorflow.keras.models import Model

import os
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from sklearn.metrics.pairwise import cosine_similarity

vgg_model = vgg16.VGG16(weights='imagenet')

model = Model(inputs=vgg_model.input, outputs=vgg_model.get_layer("fc2").output)
model.summary()

img_path = 'style/'
img_width, img_height = 224, 224
most_similar = 5

files = [img_path + x for x in os.listdir(img_path) if 'png' in x]

image_list = []
for f in files:
    file_name = f
    image = load_img(file_name, target_size=(224, 224, 3))
    img_array = img_to_array(image)
    img_batch = np.expand_dims(img_array, axis=0)
    
    image_list.append(img_batch)

images = np.vstack(image_list)
processed_imgs = preprocess_input(images)
img_feature = model.predict(processed_imgs)

cosine = cosine_similarity(img_feature)
cos_df = pd.DataFrame(cosine, columns=files, index=files)

def similarity_image(img):
    print('-'*30)
    print('selected product:')
    image = load_img(img, target_size=(img_width, img_height, 3))
    plt.imshow(image)
    plt.show()
    
    print('-'*30)
    print('similar products:')
    cosine_img = cos_df[img].sort_values(ascending=False)[1:most_similar+1].index
    
    cosine_img_score = cos_df[img].sort_values(ascending=False)[1:most_similar+1]
    
    for i in range(0, len(cosine_img)):
        image_cosine = load_img(cosine_img[i], target_size=(img_width, img_height, 3))
        plt.imshow(image_cosine)
        plt.show()

similarity_image(files[5])

used_Data :::::: https://www.kaggle.com/olgabelitskaya/style-color-images