Texture Classification: Using Neural Networks to Differentiate a Leopard from its Background

Krishna Natarajan & Vijay Subramanian
Graduate Students, University of Florida
Fl, USA

Application area: Texture classification

Problem Definition
The problem is to distinguish between the leopard and the background, in which it is sitting, in the image shown below. This problem falls under the category called texture classification, which falls under the broader category of pattern classification.

Solution Approach
Sample small 5x5 pixel images from sections of the image that represent the leopard and sections that represents the background. Flatten the sampled 2D images into one-row vectors and use them as training data for a neural network.

Once the neural network has been trained, sample the entire image as 5x5 images, flatten the images and test the neural network on the flattened images.

The neural network’s response to each flattened image can now be used to determine which areas of the image represent the leopard and which areas represent the background.

Solution Steps
1. The image file (leopard.jpg) is read into a matrix.
2. Small 5x5 sub images of the image are sampled out.

Figure 1: Sample small 5x5 pixel sub-images that represent the leopard and the background. Each red square represents a sub-image of the leopard and each yellow square represents a sub-image of the background.

3. Sampled images are flattened into single rows vectors.

% Example %
>> oneimage


oneimage =

	151   150   144   141   144
	154   154   151   149   150
	155   155   154   153   150
	155   156   158   156   150
	158   160   164   162   151

>> flattened_image = nsFlatten(oneimage,1)

flattened_image =

  Columns 1 through 12

   151   150   144   141   144   154   154   151   149   150   155   155

  Columns 13 through 24

   154   153   150   155   156   158   156   150   158   160   164   162

  Column 25

   151

   

4. The one-row vectors are used to train a neural network in NeuroSolutions for MATLAB.

Figure 2: The Learning Curve that indicates how well the problem was learned by the neural network

5. The entire image is sampled as 5x5 sub-images as before and are flatten into one-row vectors.
6. The trained neural network is now tested on all of 5x5 sub images to determine which ones are part of the leopard and which ones are part of the background.

Figure 3: The trained neural network's response indicates which sub-images represent the leopard and which ones represent the background.
The red squares represent the areas that the neural network determined to be the leopard.

Download the complete code for use in NeuroSolutions for MATLAB: tex.zip