Image Processing in Python

• Images can be brought to life using scikit-image and image processing in Python.

What is Image Processing?

• Image processing involves operations on images and videos.
• These operations enhance an image.
• Image processing extracts useful information.
• It analyzes data to make informed decisions.

Applications of Image Processing

• Medical image analysis is one application.
• It is used in artificial intelligence.
• Image restoration and enhancement is another application.
• Geospatial computing uses image processing.
• It enables surveillance.
• Image processing is used in robotic vision.
• It contributes to automotive safety.
• There are many more applications.

Image Processing in Agriculture

• Image processing is used in Smart Agriculture Modelling through image processing and remote sensing technology.

Purposes of Image Processing

• Visualization is used to reveal objects not visible to the naked eye.
• It sharpens and restores images to improve quality.
• Image retrieval is used to search for specific images of interest.
• It measures patterns to analyze various objects.
• It recognizes images to distinguish objects.

Intro to scikit-image

• Scikit-image is easy to use.
• Scikit-image leverages machine learning.
• It provides out-of-the-box complex algorithms.

What is an Image?

• An image is a matrix of pixels, referred to as picture elements.
• An image can be just a matrix of numbers [0, 255].
• An example is a 1080 x 1080 x 3 matrix for an RGB image.

Colormaps

• Colormaps are mappings, or lookup tables, that translate scalar values like grayscale intensity, or other single-channel values, into colors.

Images in scikit-image

• To import data, use the following code: from skimage import data.

RGB Channels

• RGB channels create Red channel, Green channel and Blue channel images.

Grayscaled Images

• In grayscaled images, 0 means black with no light.
• 255 indicates white, or full light.

RGB vs Grayscale

• Converting an image to grayscale involves a call to color.rgb2gray(original) .
• To visualize the grayscale image, apply show_image(grayscale, "Grayscale").
• When converting back to RGB using color.gray2rgb(grayscale), each pixel in the grayscale image gets replicated across all three color channels to form an equivalent RGB image.
• The resulting image lacks the original colors and has identical values across all three channels.
• In RGB, a pixel where R=G=B is on the grayscale spectrum, meaning it does not exhibit a color bias towards red, green, or blue.

NumPy for Images

• NumPy is used for fundamentals of image processing techniques.
• These techniques include flipping images.
• It involves extracting and analyzing features.

Images as NdArrays

• Images can be loaded as NdArrays using Matplotlib using the code: madrid_image = plt.imread('/madrid.jpeg').

Colors with NumPy

• Colors can be manipulated using NumPy.
• To obtain the red values of the image, use: red = image[:, :, 0].
• To obtain the green values of the image, use: green = image[:, :, 1].
• To obtain the blue values of the image, use: blue = image[:, :, 2].

Image Properties

• Image shapes can be accessed using madrid_image.shape.
• Image size can be accessed using madrid_image.size.

Flipping Images

• Flipping an image vertically is done using np.flipud(madrid_image).
• Flipping an image horizontally is done using np.fliplr(madrid_image).

What is a Histogram?

• Histograms are used in image processing.

Applications of Histograms

• Histograms perform analysis.
• They enable thresholding.
• Histograms are used to control image brightness and contrast.
• They equalize images.

Histograms in Matplotlib

• To obtain the red histogram of an image, use: plt.hist(red.ravel(), bins=256).

Visualizing Histograms with Matplotlib

• To visualize a blue histogram, use: blue = image[:, :, 2].
• Then, to display, use plt.hist(blue.ravel(), bins=256).

Thresholding

• This technique partitions an image into foreground and background by converting it to black and white.
• 255 (white) if pixel > thresh value
• 0 (black) if pixel < thresh value
• It is the simplest method of image segmentation.
• It is used to isolate objects for object and face detection.

Implementation of Thresholding

• Thresholding is only from grayscale images.

Applying Thresholding

• To obtain the optimal threshold value, set thresh = 127.
• Apply thresholding to the image with binary = image > thresh.
• Display the image using show_image(binary, 'Thresholded').

Inverted Thresholding

• To apply thresholding to the image in an inverted manner, use: inverted_binary = image <= thresh.
• To show the altered image, apply the thresholded image with show_image(inverted_binary, 'Inverted thresholded').

Thresholding Algorithms

• To examine other thresholding algorithms, from skimage.filters use the function try_all_threshold .
• Obtain all the resulting images with the code fig, ax = try_all_threshold(image, verbose=False).
• Show resulting plots with show_plot(fig, ax).

Optimal Thresh Value (Global)

• Uniform backgrounds work best
• Call the otsu threshold function from skimage.filters with import threshold_otsu.
• Apply thresh = threshold_otsu(image) to obtain the optimal threshold value.
• Apply this to the image with binary_global = image > thresh.
• The thresh parameter here is a single value, or is usually an int.
• The original and binarized image is displayed by show_image(binary_global, 'Global thresholding').

Thresholding Categories

• Global type is good for uniform backgrounds.
• Local or adaptive type is for when you have uneven background illumination.

Optimal Thresh Value (Local)

• This involves using local import threshold filters, specifically threshold_local.
• To set the block size to 35: block_size = 35.
• Obtain the optimal local thresholding with the function local_thresh = threshold_local(text_image, block_size, offset=10).
• Apply local thresholding and get the binary image with binary_local = text_image > local_thresh.
• The parameter local_thresh parameter is a matrix in which the elements represent the calculated local threshold values.

Block Size

• Block size refers to the size of the neighborhood around each pixel used to calculate local threshold.
• It is particularly useful for images with varying lighting.
• These images do not have a consistent intensity either.

Block Size and Its Effects

• A small block size closely adapts to intensity variations in high-resolution images, enhancing fine details but increasing susceptibility to noise, exemplified by a 3x3 size.
• A larger block size smooths out local variations and noise in images with gradual illumination changes, offering stability at the expense of finer detail, exemplified by 35x35 size.
• Selecting the optimal block size for local thresholding requires experimentation.

Offset Value

• The offset value in local thresholding is a numerical adjustment applied to the calculated threshold level for each block (or region) of an image.
• It fine-tunes the thresholding process, compensating for uneven illumination and enhancing image segmentation.

Mechanics of Setting the Offset

• Without Offset: Initially, the local threshold for a block is calculated as the mean – or median – pixel intensity within that block.
• With Offset: The offset value is added from this local threshold, and calculates a new threshold.
• For example, if the calculated local threshold is 150 and the offset is 10, the effective threshold used for segmentation in that block is now 160.

Offset Value Effects

• Positive Offset: Makes the algorithm more conservative by raising the threshold, prevents over-segmentation, ensures that very bright pixels are considered foreground.
• Negative Offset: Lowers the threshold, helping ensure more pixels are classified as foreground.