Python: Simple Image Recognition System

The goal of the image recognition system explained here is to find parameter values that result in the model’s output being correct as often as possible. This kind of training, in which the correct solution is used together with the input data, is called supervised learning. (Unsupervised learning is out of the scope of this article.)

After the training phase has been done, the model’s parameter values don’t change anymore and the model can be used in the testing phase for classifying images which were not part of its training dataset.

---

Preparation


[1] Python
https://www.python.org/downloads/
Download 3.6.4 (or later) and install it.


[2] TensorFlow
https://www.tensorflow.org/versions/r0.12/get_started/os_setup

Pip installation
Access
http://pip.readthedocs.org/en/stable/installing/

Download
get-pip.py

Then save get-pip.py on your directory.

Run the following scripts on your directory. (e.g., Terminal, Mac OS)
python get-pip.py

https://www.tensorflow.org/versions/r0.12/get_started/os_setup#pip_installation

pip install tensorflow


[3] The CIFAR-10 python version dataset:
https://www.cs.toronto.edu/~kriz/cifar.html
Download "CIFAR-10 python version"

Then extract the downloaded file and save the cifar-10-batches-py holder on your directory.

It consists of 60,000 images (10 different categories * 6,000 images per category). Each image has a size of 32 by 32 pixels. Each pixel is described by three floating point numbers representing the red (R), green (G) and blue (B) values for this pixel. This results in 32 x 32 x 3 = 3,072 values for each image.


[4] Python scripts
Download all the py scripts on
https://github.com/wolfib/image-classification-CIFAR10-tf
then save them on your directory.

---

1.Softmax (not a neutral network)

Run the following scripts on your directory. (e.g., Terminal, Mac OS)
python softmax.py

Step     0: training accuracy 0.08
Step   100: training accuracy 0.3
Step   200: training accuracy 0.26
Step   300: training accuracy 0.24
Step   400: training accuracy 0.34
Step   500: training accuracy 0.28
Step   600: training accuracy 0.35
Step   700: training accuracy 0.27
Step   800: training accuracy 0.37
Step   900: training accuracy 0.37
Test accuracy 0.266
Total time:  3.95s

The accuracy of evaluating the trained model on the test set is about 27% (this might vary on your environment).  So our model is able to pick the correct label for an image it has never seen before around 27% of the time. There are 10 different labels, so random guessing would result in an accuracy of 10%.

---

2. Neutral Network

Let's build a neural network that performs the same task.

two_layer_fc.py
This defines the model

run_fc_model.py
This runs the model (‘fc’ stands for fully connected).

Run the following scripts on your directory. (e.g., Terminal, Mac OS)
python run_fc_model.py


Parameters:
batch_size = 400
hidden1 = 120
learning_rate = 0.001
max_steps = 2000
reg_constant = 0.1
train_dir = tf_logs

2017-12-29 12:50:00.619290: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX AVX2 FMA
Step 0, training accuracy 0.1
Step 100, training accuracy 0.3225
Step 200, training accuracy 0.365
Step 300, training accuracy 0.3925
Step 400, training accuracy 0.4125
Step 500, training accuracy 0.4575
Step 600, training accuracy 0.3925
Step 700, training accuracy 0.4475
Step 800, training accuracy 0.4875
Step 900, training accuracy 0.475
Saved checkpoint
Step 1000, training accuracy 0.5025
Step 1100, training accuracy 0.5125
Step 1200, training accuracy 0.5025
Step 1300, training accuracy 0.45
Step 1400, training accuracy 0.515
Step 1500, training accuracy 0.515
Step 1600, training accuracy 0.5625
Step 1700, training accuracy 0.5575
Step 1800, training accuracy 0.5525
Step 1900, training accuracy 0.5375
Saved checkpoint
Test accuracy 0.4628
Total time: 34.08s


This indicates that our model is not significantly overfitted. The performance of the softmax classifier was around 30%, so 46% is an improvement of about 50%.


Run the following scripts on your directory. (e.g., Terminal, Mac OS)
tensorboard --logdir=tf_logs

Then access the following URL via your browser.
(your hostname).local:6006

---

Source:
http://www.wolfib.com