Regarding the classic computer vision, briefly summarize and list some knowledge points for future study or reference. The reason why it is called classic is mainly because there are some traditional methods in it, compared to the deep learning methods developed in recent years.
1 Introduction
2. Image, sampling, frequency domain processing
2.1 Grayscale, Color, Resolution, Dynamic Range, Storage, etc.
Regarding sampling, we must first understand the interpretation of the signal in the frequency domain, and understand the Fourier transform.
Sampling criteria: In order to reconstruct the original signal from the samples, the sampling frequency must be at least twice the maximum frequency of the sampled signal.
2.2 Discrete Fourier Transform, its properties
Translation, rotation, scale change, frequency domain superposition (linear).
2.3 Transforms other than Fourier
Discrete Cosine Transform
Discrete Hartley Transform
Wavelet Transform
Walsh Transform
2.4 Filtering—Application of Frequency Domain Characteristics
3. Basic image processing operations
3.1 Histogram
3.2 Point operation
Luminance mapping: addition, inversion, scaling, logarithmic, exponential.
Luminance normalization, histogram equalization, thresholding and optimal thresholding.
3.3 Group operations
Template convolution (including frequency domain implementation). Statistical operators: direct averaging, Gaussian filtering, median filtering, mode (truncated median) filtering.
Direct averaging: removes a lot of noise, but blurs feature boundaries;
Gaussian filtering: retains more features, but has little advantage over direct averaging (noise is not Gaussian);
median filtering: retains some noise, but Get clear boundary features;
truncate median filter: remove more noise, but also remove more image details.
Anisotropic diffusion for image smoothing.
force field transformation.
Mathematical morphology: hit or miss transformation, erosion, dilation (including grayscale operators: erosion operator, dilation operator, open operator, closed operator, etc.), Minkowski operator.
4. Low-level feature extraction (including edge detection) (edge detection, corner detection, motion detection)
4.1 Edge Detection
First-order edge detection operators
Difference operation, Roberts cross operator, smoothing, Prewitt, Sobel, Canny and other operators.
Second-order edge detection operator Second-
order difference, Laplacian operator, zero-crossing detection; Marr-Hildreth operator, Laplacian of Gaussian, Gaussian difference, scale space.
Other edge detection operators: Spacek operator, Petrou operator, Susan operator.
The results of all edge operators are achieved with hysteresis thresholding.
For images with a lot of noise, Canny operator and Spacek operator perform better than other operators.
4.2 Phase coherence
The phase consistency method is a feature detection operator with two main advantages:
- Can detect a wide range of features;
- Invariant to local (and smooth) lighting changes.
Frequency domain analysis; detection of a chain of features; photometric invariance, wavelets.
4.3 Location feature extraction (corner extraction)
Plane curvature; corners; Moravec and Harris detectors; scale space.
SIFT (Scale Invariant Feature Transform)
has scale and rotation invariance. There is also a certain invariance to illumination changes.
SURF (Accelerated Robust Features).
Significance operator.
4.4 Describing Image Motion
Differential Detection; Optical Flow, Aperture Problems, Smooth Constraints; Differential Methods; Horn and Schunk Methods; Correlation.
5. Feature Extraction for Shape Matching
5.1 Thresholding and Background Subtraction (Difference)
5.2 Template matching
Direct implementation and Fourier implementation.
5.3 Low-level features
Collection of low-level features for object extraction; frequency-based and component-based methods; detection of distributions of measurements.
Wavelets and Haar wavelets; SIFT and SURF descriptions; and histograms of directional gradients.
5.4 Hough Transform
Feature extraction through matching; detection of quadratic curves and arbitrary shapes using Hough transform; invariant expression.
6. Advanced Feature Extraction: Deformable Shape Analysis
No knowledge of the target shape (i.e. feature) model; the shape is unknown or the fluctuations in the shape cannot be parameterized.
6.1 Deformable shape analysis
Deformable templates; energy maximization; part-based shape analysis;
6.2 Active Contour and Snake Model
Energy Minimization for Curve Evolution; Greedy Algorithm; Kass Snake Model.
6.3 Shape Skeletonization
Distance transforms and shape skeletons; mean axis transforms; discrete symmetry operators; evidence accumulation for symmetrical point distributions.
6.4 Active Shape Model
Shape changes are expressed by statistical methods; shape changes are obtained by feature extraction.
Active shape model; active appearance model; principal component analysis.
7. Objective Description
7.1 Boundary description
Boundaries and areas, how to define boundaries and the areas they delimit, how to form area descriptions and their necessary characteristics?
Basic method: chaincode. Fourier Descriptors: Discrete Approximations; Cumulative Angle Functions and Elliptic Fourier Descriptors.
7.2 Region Descriptor
How to describe the area of the shape?
Basic shape measures: area; perimeter; compactness; dispersion.
Moments: Fundamental Moments; Central Moments; Invariant Moments; Zernike Moments; Properties and Reconstruction.
8. Basics of Texture Description, Segmentation and Classification
8.1 Texture description
What is an image texture and how do I determine several sets of values to identify the texture?
Feature extraction: Fourier transforms, co-occurrence matrices, regions;
modern methods: local binary patterns (LPB) and uniform LBP.
Characterization: energy, entropy, inertia.
8.2 Distance Metrics
Distance Metrics: Manhattan Urban and Euclidean (L1/L2 Distance), Mahalanobis, Bhattacharyya and Cosine; Construction, Visualization and Confusion Matrix.
8.3 Texture Classification
How can I relate the obtained values to known samples?
k-nearest neighbors, support vector machines, and other classification methods (intersection with machine learning).
8.4 Texture segmentation
How to find out the texture area within the image extent?
Convolution calculation, tiling, thresholding.
9. Detection and description of moving objects
9.1 Moving Object Extraction
How to separate moving objects from their background?
Average and median filtering applied to background image estimates. Isolated from background by subtraction. Improved by Gaussian mixture and thresholding.
Problem: Color, Lighting and Shadows.
Use dilation and erosion; open and close operations. Connected component analysis.
9.2 Moving Object Tracking
Time domain consistency is achieved during tracking. Modeling linear system dynamics.
Tracking by local search; Lucas-Kanade method; Kalman filtering; multi-target tracking; comparison of feature points and background subtraction; Camshift and Meanshift methods. Object detection tracking.
9.3 Motion Shape Analysis and Description
Describing motion and extracting motion shapes through evidence collection. Add velocity and displacement to shape description; for identification purposes, describe moving objects.