Geotag prediction from images using DNN

Introduction

This time, I would like to make geotag prediction from a building image using a trained model. Especially in this article, I started with the purpose of using the output of multiple labels of latitude and longitude from the input image.

Data to use

The data used this time is "European Cities 1M dataset" http://image.ntua.gr/iva/datasets/ec1m/index.html

Use the landmark set image and geotag on this site respectively.

Construction environment

The implementation in this article uses Google Colaboratory. The settings of the environment used are listed below.

• Python3.6.9
• tensorflow 2.3.0
• Keras 2.4.3 --Use GPU

import keras
from keras.utils import np_utils
from keras.models import Sequential, Model, load_model
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.layers.core import Dense, Dropout, Activation, Flatten
import numpy as np
from sklearn.model_selection import train_test_split
import os, zipfile, io, re
from PIL import Image
import glob
from tqdm import tqdm

from sklearn.model_selection import train_test_split
from keras.applications.xception import Xception
from keras.applications.resnet50 import ResNet50
from keras.layers.pooling import GlobalAveragePooling2D
from keras.optimizers import Adam, Nadam

Preprocessing

Here, as pre-processing, processing with latitude and longitude labels is performed. This time, we will learn the latitude and longitude separately as labels, so we will process them so that they can be easily retrieved as a list.

with open("landmark/ec1m_landmarks_geotags.txt") as f:
    label=f.readlines()
    for i in label:
        ans=i.split(' ')
        ans[1]=ans[1].replace('\n','')
        print(ans)

output

['41.4134', '2.153']
['41.3917', '2.16472']
['41.3954', '2.16177']
['41.3954', '2.16177']
['41.3954', '2.16156']
['41.3899', '2.17428']
['41.3953', '2.16184']
['41.3953', '2.16172']
['41.3981', '2.1645']
.....
.....

Data acquisition

Image size is 100 Convert dataset to array Label the image without separating the latitude and longitude

X = []
Y = []
image_size=100

with open("landmark/ec1m_landmarks_geotags.txt") as f:
    label=f.readlines()

    dir = "landmark/ec1m_landmark_images"
    files = glob.glob(dir + "/*.jpg ")
    for index,file in tqdm(enumerate(files)):
        image = Image.open(file)
        image = image.convert("RGB")
        image = image.resize((image_size, image_size))
        data = np.asarray(image)
        X.append(data)
        Y.append(label[index])
 
X = np.array(X)
Y = np.array(Y)

output

927it [00:08, 115.29it/s]

X.shape,Y.shape
((927, 100, 100, 3), (927,))

Next, divide into train, test, valid Latitude and longitude are also divided here.

y0=[]  #Latitude label
y1=[]  #Longitude label

for i in Y:
    ans=i.split(' ')
    ans[1]=ans[1].replace('\n','')
    y0.append(float(ans[0]))
    y1.append(float(ans[1]))

y0=np.array(y0)
y1=np.array(y1)

#x(train,test)Split
X_train, X_test = train_test_split(X, random_state = 0, test_size = 0.2)
print(X_train.shape,  X_test.shape) 

#(741, 100, 100, 3) (186, 100, 100, 3)

#y0,y1(train,test)Split
y_train0,y_test0,y_train1, y_test1 = train_test_split(y0,y1,
                                                      random_state = 0,
                                                      test_size = 0.2)
print(y_train0.shape,  y_test0.shape) 
print(y_train1.shape,  y_test1.shape)

#(741,) (186,)
#(741,) (186,)

#Data type conversion & normalization
X_train = X_train.astype('float32') / 255
X_test = X_test.astype('float32') / 255

#x(train,valid)Split
X_train, X_valid= train_test_split(X_train, random_state = 0, test_size = 0.2)
print(X_train.shape,  X_valid.shape) 

#(592, 100, 100, 3) (149, 100, 100, 3)

#y0,y1(train,valid)Split
y_train0, y_valid0,y_train1, y_valid1= train_test_split(y_train0,y_train1,
                                                        random_state = 0,
                                                        test_size = 0.2)

print(y_train0.shape,  y_valid0.shape) 
print(y_train1.shape,  y_valid1.shape) 

#(592,) (149,)
#(592,) (149,)

Model building

This time, we will use the trained model of Xception by referring to the following article. https://qiita.com/ha9kberry/items/314afb56ee7484c53e6f#データ取得

I also wanted to try other models, so I will try using Resnet as well

#xception model
base_model = Xception(
    include_top = False,
    weights = "imagenet",
    input_shape = None
)

#resnet model
base_model = ResNet50(
    include_top = False,
    weights = "imagenet",
    input_shape = None
)

Enter one predicted value at the end of the regression problem Prepare label output for latitude and longitude

x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1024, activation='relu')(x)
predictions1 = Dense(1,name='latitude')(x)
predictions2 = Dense(1,name='longitude')(x)

Put predictions1 and predictions2 in output. Others are as in the reference article This time, we will study using Adam and Nadam as optimizers.

model = Model(inputs = base_model.input, outputs = [predictions1,predictions2])

#Freeze up to 108 layers
for layer in model.layers[:108]:
    layer.trainable = False

    #Unfreeze Batch Normalization
    if layer.name.startswith('batch_normalization'):
        layer.trainable = True
    if layer.name.endswith('bn'):
        layer.trainable = True

#Learn after 109 layers
for layer in model.layers[108:]:
    layer.trainable = True

# layer.compile after setting trainable
model.compile(
    optimizer = Adam(),
    #optimizer=Nadam(),
    loss = {'latitude':root_mean_squared_error,
            'longitude':root_mean_squared_error
        }
)

Learning

history = model.fit( X_train,   #decode_train
                    {'latitude': y_train0,
                     'longitude':y_train1},
                    batch_size=64,
                    epochs=50,
                    validation_data=(X_valid,    decode_valid
                                     {'latitude' :y_valid0,
                                      'longitude':y_valid1}),
                    )

output

Epoch 1/50
10/10 [==============================] - 4s 409ms/step - loss: 0.6146 - latitude_loss: 0.4365 - longitude_loss: 0.1782 - val_loss: 1.6756 - val_latitude_loss: 1.3430 - val_longitude_loss: 0.3326
Epoch 2/50
10/10 [==============================] - 4s 404ms/step - loss: 0.5976 - latitude_loss: 0.4415 - longitude_loss: 0.1562 - val_loss: 0.7195 - val_latitude_loss: 0.5987 - val_longitude_loss: 0.1208

...
...

Plot the results

import matplotlib.pyplot as plt


plt.figure(figsize=(18,6))

# loss
plt.subplot(1, 2, 1)
plt.plot(history.history["latitude_loss"], label="latitude_loss", marker="o")
plt.plot(history.history["longitude_loss"], label="longitude_loss", marker="o")
#plt.yticks(np.arange())
#plt.xticks(np.arange())
plt.ylabel("loss")
plt.xlabel("epoch")
plt.title("")
plt.legend(loc="best")
plt.grid(color='gray', alpha=0.2)

plt.show()

result

Evaluation

#batch size 64  Adam
scores = model.evaluate(X_test,{'latitude' :y_test0,
                                'longitude':y_test1}, 
                             verbose=1)

print("total loss:\t{0}".format(scores[0]))
print("latitude loss:\t{0}".format(scores[1]))
print("longtitude loss:{0}".format(scores[2]))

output

total loss:	0.7182420492172241
latitude loss:	0.6623533964157104
longtitude loss:0.05588864907622337

Forecast

# show image, prediction and actual label
for i in range(10,12):
    plt.figure(figsize=(10,10))
    print('latitude:{} \tlongititude{}'.format(
        prediction[0][i],
        prediction[1][i],
        ))

    plt.imshow(X_test[i].reshape(100, 100, 3))
    plt.show()

latitude:[39.69221] longititude[2.2188098]

latitude:[39.728386] longititude[2.224149]

There are numbers like that, but when expressed on the map (Google Map), it is in the middle of the sea as shown below, which is not enough to use.

Comparison with other parameters

finally

In this trial, it was found that the combination of Xception and Nadam has the highest accuracy. In the future I will use another model or create a model from scratch

Reference, citation

data set

Publications Conferences Y. Avrithis, Y. Kalantidis, G. Tolias, E. Spyrou. Retrieving Landmark and Non-Landmark Images from Community Photo Collections. In Proceedings of ACM Multimedia (MM 2010), Firenze, Italy, October 2010.

Journals Y. Kalantidis, G. Tolias, Y. Avrithis, M. Phinikettos, E. Spyrou, P. Mylonas, S. Kollias. VIRaL: Visual Image Retrieval and Localization. In Multimedia Tools and Applications (to appear), 2011.

article

--https://qiita.com/ha9kberry/items/314afb56ee7484c53e6f # Overview

• https://qiita.com/cvusk/items/1439c1c6dde160c48d13