Skip to content

Detection Data Management

This component of MAGIST deals with image data manipulation and processing. This class in MAGIST will perform image resizing, integrity verification, and image splitting as depicted by the previous article. This prepares the data so that the final CNN can run properly.

Importing Necessary Libraries

MAGIST will automatically handle all the dependencies as long as the main DataDetectionManager module is imported:

1
from MAGIST.Vision.DetectionDataManager.image_slicer import ImageSlicer

Usage

Now, we must initialize the class and begin to use it:

1
slicer = ImageSlicer("path/to/config.json")

Resize Image

This function allows users to resize all images in a given directory to a standard size.

Danger

It will overwrite existing images. As of now, it does not support functionality to export the images elsewhere. If someone would like to add this functionality, please visit our GitHub Repository to learn how to create a pull-request and request your modifications to be merged.

1
slicer.resizer((length,width), "path/to/directory/with/images")

Important

Please note that the first argument is a tuple with the image dimensions. Also note that the images do not have to be a square for this to work. It will expand or shrink image dimensions on demand.

Coordinate Computation

In order to split each input image into a grid, we must first compute the vertices of the grid itself. This will create an array with all the vertices of the grid boxes that can be used to easily split each image.

1
coordinates = slicer.coordinate_compute((input_length,input_width), (output_length,output_width)) # (1)
  1. Both arguments are tuples in this case!

Important

Although this algorithm has preliminary checks in place to ensure that the computation can succeed, there can be edge-cases that can cause major crashes in the algorithm. To ensure this does not happen, please ensure that the input dimensions divide evenly into the output dimensions. For example, the following input-output pairs will work: (500, 500) --> (50, 50) BUT (500, 500) --> (36, 36) will not.

What does the coordinates variable look like and how does it contain the grid information?

The coordinate variable that you created will look like this:

[[left1, upper1, right1, bottom1], [left1+length_of_output, upper1, right1+length_of_output, bottom1] ...]

The important thing to note here is how the upper1 and bottom1 will not be modified until the rightmost part of the image is reached. It will essentially split the images from left to right and top to bottom. That is why it is essential for the output images to divide perfectly into the input images.

Image Slicing

Now for the best part! We will use the coordinates array that we previously computed to split the images.

1
slicer.crop_segments(coordinates, "path/to/input/images", "path/to/base/output/directory", "image_class")

The first 2 arguments are quite straight forward. The 3rd argument describes where the base data directory is. The the image class (4th argument) will be appended to the "base directory" to make the final output directory.

For example:
Output directory: /home/robot/Project/Data
Image class: Turtles

Path to images: /home/robot/Project/Data/Turtles/

This functionality helps format the data in directory in a way TensorFlow can easily interpret.

Image Integrity Verification

Often times, images directly downloaded from the internet can be corrupt. We can use the integrity verification function to delete corrupt images and ensure that TensorFlow doesn't throw any data errors.

1
slicer.image_integrity_verification("path/to/images", "delete_images_bool")

This function provides the option to simply check if images are corrupt and count them or to also delete these images from the dataset.

Convert Images to PNG (experimental)

This class also allows images to be converted to PNG which allows all images to have an alpha value (RGBA instead of RGB) to further standardize the data. This feature is not extensively tested yet.

1
slicer.convert_img_to_png("path/to/images")

Next Steps

After the images are split and sorted by class into their designated folders, it is time to train the Convolutional Neural Network using our custom model and TensorFlow.