jyt109 / pyimage Goto Github PK

#Image Procecessing in Python

If you are interested in image classification problems, e.g. classifying dog from cat, female from male, etc, pyimage is designed to make image transformations and vectorization easy.

1. First you need to download images you are interested to classify. By convention, the directory structure is such that under the parent directory, there are sub-directories containing images of each of the classes you are interested in. Here is a small dataset ^[1] containing images of cats and dogs. The parent directory is cat_dog, and the sub-directories are cat and dog.

   
   

2. Import the ImagePipeline class and instantiate by providing the path of the parent path:

   from pyimage.pipeline import ImagePipeline
   pipe = ImagePipeline('path/to/cat_dog')

   
   

3. Read the images in by using the read function. The read function accepts an optional sub_dirs argument which is a tuple of the sub-directory names. If no argument is provided, all the sub-directories are read. The images are stored as a list of list (of matrices) in the instance variable img_lst2.

   pipe.read(sub_dirs=('cat', 'dog'))

   
   

4. Resize the images by providing the dimensions. The images are by default read in as a 3D tensor, therefore the dimension argument must be a tuple of 3.

   pipe.resize((300, 300, 3))

   
   

5. If you are interested to view some of your images anytime along your resizing/transformation steps, you can use the show function. The first argument is the name of sub-directory and the second argument is the nth image in the sub-directory you would want to view.

   pipe.show('cat', 0)

   
   

6. To transform all the images, use the transform function. The first arugment is the transformation function, usually a function from skimage. The second argument is a dictionary of the additional arguments needed for the transformation function.

   from skimage.color import rgb2gray
   pipe.transform(rgb2gray, {})
   pipe.show('cat', 0)

   
   

   If you want to just transform one image instead of all the images to see what the effects are like, the third and four arguments are the name of the sub-directory and the nth image you want to apply the transformation to. The function will be applied to the one image, but not change the original copy.

   from skimage.feature import canny
   pipe.transform(canny, {}, 'cat', 0)

   
   

7. When you have done all the necessary transformations, call vectorize. The images will be flattened to form a feature matrix where each row are the pixels of an image. The class labels are also generated as a vector. The feature matrix and label vector are accessible in the instance variable features and labels respectively.

   pipe.vectorize()
   pipe.features
   pipe.labels

References:

1. Jeremy Elson, John R. Douceur, Jon Howell, Jared Saul, Asirra: A CAPTCHA that Exploits Interest-Aligned Manual Image Categorization, in Proceedings of 14th ACM Conference on Computer and Communications Security (CCS), Association for Computing Machinery, Inc., Oct. 2007