Street view number detection is called
natural scene text recognition problem which is quite different from printed
character or handwritten recognition. Research in this field was started in
90’s, but still it is considered as an unsolved issue. As I mentioned earlier
that the difficulties arise due to fonts variation, scales, rotations, low
In earlier years to deal with natural
scene text identification sequentially, first character classification by
sliding window or connected components mainly used. 4 After that word
prediction can be done by predicting character classifier in left to right
manner. Recently segmentation method guided by supervised classifier use where
words can be recognized through a sequential beam search. 4 But none of this
can help to solve the street view recognition problem.
In recent works convolutional neural
networks proves its capabilities more accurately to solve object recognition
task. 4 Some research has done with CNN to tackle scene text recognition
tasks. 4 Studies on CNN shows its huge capability to represent all types of
character variation in the natural scene and till now it is holding this high
variability. Analysis with convolutional neural network stars at early 80’s and
it successfully applied for handwritten digit recognition in 90’s. 4 With the
recent development of computer resources, training sets, advance algorithm and
dropout training deep convolutional neural networks become more efficient to recognize
natural scene digit and characters. 3
Previously CNN used mainly to detecting a
single object from an input image. It was quite difficult to isolate each
character from a single image and identify them. Goodfellow et al., solve this
problem by using deep large CNN directly to model the whole image and with a
simple graphical model as the top inference layer. 4
The rest of the paper is designed in
section III Convolutional neural network architecture, section IV Experiment,
Result, and Discussion and Future Work and Conclusion in section V.
Convolutional Neural Networks
(CNN) is a multilayer network to handle complex and high-dimensional data, its
architecture is same as typical neural networks. 8 Each layer contains some
neuron which carries some weight and biases. Each neuron takes images as
inputs, then move onward for implementation and reduce parameter numbers in the
network. 7 The first layer is a convolutional layer. Here input will be
convoluted by a set of filters to extract the feature from the input. The size
of feature maps depends on three parameters: number of filters, stride size,
padding. After each convolutional layer, a non-linear operation, ReLU use. It
converts all negative value to zero. Next is pooling or sub-sampling layer, it
will reduce the size of feature maps. Pooling can be different types: max,
average, sum. But max pooling is generally used. Down-sampling also controls
overfitting. Pooling layer output is using to create feature extractor. Feature
extractor retrieves selective features from the input images. These layers will
have moved to fully connected layers (FCL) and the output layer. In CNN
previous layer output considers as next layer input. For the different type of
problem, CNN is different.
main objective of this project is detecting and identifying house-number signs
from street view images. The dataset I am considering for this project is
street view house numbers dataset taken from 5 has similarities with MNIST
dataset. The SVHN dataset has more than 600,000 labeled characters and the
images are in .png format. After extract the dataset I resize all images in
32×32 pixels with three color channels. There are 10 classes, 1 for each digit.
Digit ‘1’ is label as 1, ‘9’ is label as 9 and ‘0’ is label as 10. 5 The
dataset is divided into three subgroups: train set, test set, and extra set.
The extra set is the largest subset contains almost 531,131 images.
Correspondingly, train dataset has 73,252 and test data set has 26,032 images.
Figure 3 is an example of the original,
variable-resolution, colored house-number images where each digit is marked by
bounding boxes. Bounding
box information is stored in digitStruct.mat file, instead of drawn directly on
the images in the dataset. digitStruct.mat file contains a struct called
digitStruct with the same length of original images. Each element in
digitStruct has the following fields: “name” which is a string containing the
filename of the corresponding image. “bbox” is a struct array that contains the
position, size, and label of each digit bounding box in the image. As an example,
digitStruct(100). bbox (1). height means
the height of the 1st digit bounding box in the 100th image. 5
This is very clear
from Figure 3 that in SVHN dataset maximum house numbers signs are printed
signs and they are easy to read. 2 Because there is a large variation in
font, size, and colors it makes the detection very difficult. The variation of
resolution is also large here. (Median: 28 pixels. Max: 403 pixels. Min: 9
pixels). 2 The graph below indicates that there is the large variation in
character heights as measured by the height of the bounding box in original
street view dataset. That means the size of all characters in the dataset,
their placement, and character resolution is not evenly distributed across the
dataset. Due to data are not uniformly distributed it is difficult to make
correct house number detection
experiment, I train a multilayer CNN for street view house numbers recognition
and check the accuracy of test data. The coding is done in python using
Tensorflow, a powerful library for implementation and training deep neural
networks. The central unit of data in TensorFlow is the tensor. A tensor
consists of a set of primitive values shaped into an array of any number of
dimensions. A tensor’s rank is its number of dimensions. 9 Along with
TensorFlow used some other library function such as Numpy, Mathplotlib, SciPy
I perform my
analysis only using the train and test dataset due to limited technical resources.
And omit extra dataset which is almost 2.7GB. To make the analysis simpler delete
all those data points which have more than 5 digits. By preprocessing the data
from the original SVHN dataset a pickle file is created which being used in my
experiment. For the implementation, I randomly shuffle valid dataset and then
used the pickle file and train a 7-layer Convoluted Neural Network.
At the very
beginning of the experiment, first convolution layer has 16 feature maps with
5×5 filters, and originate 28x28x16 output. A few ReLU layers are also added
after each convolution layer to add more non-linearity to the decision-making
process. After first sub-sampling the output size decrease in 14x14x10. The
second convolution has 512 feature maps with 5×5 filters and produces 10x10x32
output. By applying sub-sampling second time get the output size 5x5x32.
Finally, the third convolution has 2048 feature maps with same filter size. It
is mentionable that the stride size =1 in my experiment along with zero padding.
During my experiment, I use dropout technique to reduce the overfitting.
Finally, SoftMax regression layer is used to get the final output.
initialized randomly using Xavier initialization which keeps the weights in the
right range. It automatically scales the initialization based on the number of
output and input neurons. After model buildup, start train the network and log
the accuracy, loss and validation accuracy for every 500 steps.Once the process
is done then get the test set accuracy. To minimize the loss, Adagrad Optimizer used.
After reach in a suitable accuracy level stop train the network and save the
hyperparameters in a checkpoint file. When we need to perform the detection, the
program will load the checkpoint file without train the model again.
the model produced an accuracy of 89% with just 3000 steps. It’s a great
starting point and certainly, after a few times of training the accuracy will reach
in 90%. However, I added some additional features to increase accuracy. First, added
a dropout layer between the third convolution layer and fully connected layer. This
allows the network to become more robust and prevents overfitting. Secondly, introduced
exponential decay to calculate learning rate with an initial rate 0.05. It will
decay in each 10,000 steps with a base of 0.95. This helps the network to take
bigger steps at first so that it learns fast but over time as we move closer to
the global minimum, it will take smaller steps. With these changes, the model
is now able to produce an accuracy of 91.9% on the test set. Since there are a
large training set and test set, there is a chance of more improvement if the
model will train for a longer time.