Transfer Learning Using AlexNet and GoogLeNet

The purpose of the transfer learning is to fine-tune a pretrained convolutional neural network to perform classification on our new collection of images.

According Mathworks, "AlexNet has been trained on over a million images and can classify images into 1000 object categories (such as keyboard, coffee mug, pencil, and many animals)." However, we are looking for a classification accuracy over 99% on a given objects. Transfer learning is commonly used in deep learning applications. We can take a pretrained network and use it as a starting point to learn a new task. Fine-tuning a network with transfer learning is usually much faster and easier than training a network with randomly initialized weights from scratch. We can quickly transfer the learned features onto a new task using a smaller number of training images.

Acquire Images from IP Camera (smartphone)

In order to perform a transfer learning, we have to create an image data store. Taking thousands of pictures with the traditional method will be time consuming and inefficient. Yet, this can easily be performed by using our smartphone as a webcam and the Matlab snapshot function to acquire one image frame at a time from an IP camera using a for loop. Also, imwrite(A,filename) writes image data A to the file specified by filename, inferring the file format from the extension. Once the IP cam is connected, we can take continues pictures automatically by writing a simple looping Matlab script.

• First, obtain the IP address. The URL is made up of the IP address of the camera, followed by a resource designation, such as video.mjpeg or mjpg.cgi, as defined by the camera vendor. For example:

cam = ipcam('http://172.28.17.104/video/mjpg.cgi')

• Connect the smartphone IP camera by creating an object, cam, using the URL of the IP camera as shown above.
• Optionally, Username can be created as a character vector representing the user name for the IP camera, and must be the second argument.
• Then, Password can be created as a character vector representing the password for the IP camera, and must be the third argument.

For example: cam = ipcam('http://172.28.17.193/video.mjpeg', 'UserName', 'PassWord')

• After creating the object you can preview the image and take snapshots from the camera.

Creating Image DataStore:

Using the Matlab script shown below, we have created a Data store of 1000 images for each object and stored the images into designated file folders. We chose to take the pictures at an interval of one second and the pictures will be stored at specific file folder in the specified URL.

Note: our camera (android phone) does not require user authentication and our code does not include the user name and password arguments. If your camera does, see the vendor manual to create the object with user name and password arguments.

%% Automatic Picture taker

clear all

cam = ipcam('http://172.28.17.193/video.mjpeg') % replace the example IP address

partName =  'FK-testImg%d.png'; % base file name

figure

for i = 1:1000  % take a thousand pictures 

    im = snapshot(cam);       % Take a picture

    image(im);                   % Show the picture

    im = imresize(im,[227 227]); % Picture size for Alexnet CNN

    path = sprintf(partName,i);

    realPath = ['C:\xxx\xxx\xxxx\ImageCapturingTest\' path]; % replace the example file path

    imwrite(im, realPath); % writes the image to the folder.

    pause (1); % one picture every second

end

%%

After creating the DataStore of images, we used the following code to perform a successive transfer learning and validation test on our objects. This code is a modified version of the example code from MathWorks website.