Edge Detection: Sobel vs. Canny, Corner Detection

• Finding edges and detecting corners can be achieved effectively using image processing in Python

Detecting Edges

• Shape information of an image is mostly enclosed in its edges.
• Representing an image by its edges significantly reduces the amount of data while retaining most of the image information, such as shapes.
• Sobel filtering is a technique for edge detection
• Canny edge detection is a standard method that produces higher accuracy in detecting edges compared to the Sobel algorithm.
• Sobel is a good choice if you need a quick and general edge detection
• Canny is the better option if the task demands more precise and clean edge detection, especially in noisy images.
• The skimage.feature library can be used to apply the Canny filter
• The steps are:
  • Convert the image to grayscale
  • Apply a Canny detector
  • Show the resulting image with edges
• The method of applying the Canny detector includes a sigma value of 0.5, which applies less filltering, this setting allows more edges to be detected.
• Applying the Guassian algorithm to an image removes noise
• Adjust the intensity through the sigma attribute in the Canny function, which by default is 1
• A lower sigma will apply less filltering, allowing for more edges, conversely, a higher sigma value removes more noise, and results in fewer detected edges.

Corner Detection

• Corner detection extracts certain types of features to infer the contents of an image
• Corner detection is used in:
  • Motion detection
  • Image registration
  • Video tracking
  • Panorama stitching
  • 3D Modeling
  • Object recognition
• Features are points of interest in an image that are robust to changes in rotation, translation, brightness, and scale, providing valuable information about the image content.
• Corner detection is also able to identify interest points, but also focuses on corners and edges
• A corner can be defined as the intersection of two edges or a junction of contours.
• Detecting corners as interest points allows us to match objects across images from different perspectives.
• Corners identified in an original image can be matched to those in a downscaled version of the same image which can be used in recognizing and aligning featues across varying scales

Harris Corner Detector

• The Harris Corner Detector is a computer vision operator used to detect corners
• Works by identifying regions in an image where the intensity shows significant changes in multiple directions, indicating the presence of corners
• The skimage.feature library can be used to apply the Harris algotithm
• The Harris corner detector function, "corner_harris", computes the Harris response for each pixel, producing a 2D array (measure_image) where each value represents the "corner strength" at that pixel location.
• Higher values in measure_image indicate areas that are more likely to be corners.
• The corner_peaks algorithm identifies points in the image that have strong corner responses and are at least min_distance pixels apart.
• The output is a 2D array where each row represents the coordinates of one detected corner point.

Face Detection

• Use cases include:
  • Filters
  • Auto focus
  • Recommendations
  • Privacy protection
  • Recognizing emotions
• Scikit-image can be used to detect faces through Machine Learning
• This can be achieved with only a few lines of code using the skimage library!
• Using the classifer:
  • Import the classifer class
  • Load the trained file from the module root
• The trained_file object is a string that holds the file path to the XML file
• The main idea behind the cascade of classifiers is to create classifiers of medium accuracy and ensemble them into one strong classifier
• Advantages of using the cascade classifer is that easy examples can be classified only by evaluating SOME of the classifers which is much faster than the process of evaluating ONE strong classifer!
• A cascade classifer is called this because it uses a series of simple, sequential classifers, arranged in a cascade struture
• Each stage of cascade classifer is a fillter that decides whether a region is likely to containg an object of interest
• To detect faces you can use the detect_multi_scale method from the Cascade class
• This method moves a window across the image to find areas resembling a human face.
• The detect_multi_scale method adjusts the window size to account for images at different scales
• There is also a factor of a step ratio, where a ratio of 1 represents an exhaustive but slow search
• The search window requires min and max sizes in order to detemrine the range of detection

Detected faces

• The detector returns coordinates of the box that contains the face
• Where r represents the row position of the top left corner of the detected window.
• The coordinate position pixel uses "c" representing a column
• The width of the detected window is shown by "width"
• The height is shown by "height"
• A rectangle is drawn around the detected faces

Real World Applications

• Converting to grayscale before detecting edges or corners
• Reducing noise and restoring images
• Blurring Faces
• Approximating object size

Privacy protection

• Implemented by detecting and anonymozing faces
• Code imports cascade of classifiers:
  • The gaussian filter from scikit image of features and gaussian from scikit image of filters allows to detect and blur portions of an image
• To detect each face:
  • Detemrine the image size and step ratio
  • Obtain the face cropped from the detected coordinates
  • Apply gausian to extract the face, using multichannel as true
  • A function is needed to merge the blurred face to the final image, and show it
• The faces will be blurred, and the persons personal information/identity is anonymized

Recap

• Improved contrast
• Restored images
• Applied filters
• Rotated, flipped, and resized!
• Segmented: supervised and unsupervised
• Applied morphological operators
• Created and reduced noise
• Detected edges, corners and faces
• And mixed them up to solve problems!

Whats next?

• Open Source Computer Vision Library
• A comprehensive library aimed at real-time computing
• Scikit is also a powerful tool for image processing
• Open CV Offers functionalities and optimizations beneficial in certain domains
• Features:
  • Real-Time Processing
  • Video Analysis
  • Machien Learning and Deep Learning
  • Camera Calibrations ad 3D Reconstruction
  • Motion Analysis
  • GUI Features
  • Wide range of algorithims for object detection
  • Integration with other technologies
• OpenCV is:
  • Optimized for real-time applications
  • Suitable for comprehensive video processing
  • Has extensive integration for machine learning
  • Capabile of advanced 3D reconstruction
  • Advanced GUI for real time interaction
  • Support for C++, Python, Java
  • Widespread industry support
• SkiKit:
  • Ideal for image processing
  • Suitable for offline analysis
  • Limited video support and built-in machien learning
  • Minimal GUI and is limited for algorithim variety
  • Great support academically

Explore image edge detection techniques, including Sobel and Canny. Understand the advantages of edge representation over pixel data and the role of Gaussian filters. Learn how to apply corner detection for image analysis and feature matching across different viewpoints.