Outline • Linear Shift-invariant system • Linear filters • Fourier transformation – Time and frequency representation • Filter Design Linear System Theory • What is a system? – A system is anything that accepts an input and produces an output in response y[n] = T{x[n]} where x[n] is the input sequence and y[n] is the output sequence in responses to x[n] – How to represent a sequence? x[n] x[k ] [n k ] k 5/29/2016 Visual Perception Modeling 2 Linear System • Linearity – – – – y1[n] = T{x1[n]} y2[n] = T{x2[n]} Then y1[n]+y2[n] = T{x1[n]+x2[n]} 5/29/2016 Visual Perception Modeling 3 Shift-Invariant System • Shift invariance – y[n] = T{x[n]} – y[n-T] = T{x[n-T]} • LSI system – A LSI system is completely characterized by its impulse response h[n] – For any other input, we can obtain the response through convolution 5/29/2016 Visual Perception Modeling 4 Filtering • Closely related to convolution • Filter examples – Smoothing by averaging – Smoothing by Gaussian 2 2 1 (x y ) G ( x, y) exp 2 2 2 2 5/29/2016 Visual Perception Modeling 5 Multi-scale Representation • Scale in the Gaussian function – is the standard deviation of the Gaussian distribution – When is small, no smoothing or very little – When is large, the noise will be largely disappear. However, the image detail will disappear along with the noise 5/29/2016 Visual Perception Modeling 6 Gaussian Pyramid 5/29/2016 Visual Perception Modeling 7 Why Gaussian Smoothing? • Scale space – If we convolve a Gaussian with a Gaussian, it will also be a Gaussian G 1 * G 2 G – Efficiency 12 22 • A small kernel is generally enough • Separable – Central limit theorem 5/29/2016 Visual Perception Modeling 8 Spatial Frequency Analysis • Filter response analysis – For example, why does smoothing reduce noise? – What is the difference between the discrete image representation and a continuous surface representation? – Is there any way we can design the best filter for a certain task? • For smoothing, how can we have the best smoothing kernel? 5/29/2016 Visual Perception Modeling 9 Fourier Transforms • Fourier transform F ( g ( x, y ))(u , v) g ( x, y )e j 2 ( ux vy ) dxdy – The transformation takes a complex valued function x, y and returns a complex valued function of u, v – U and v determine the spatial frequency and orientation of the sinusoidal component 5/29/2016 Visual Perception Modeling 10 Inverse Fourier Transform • Inverse Fourier transform g ( x, y ) F ( g ( x, y))(u, v)e j 2 ( ux vy ) dudv – It recovers a signal from its Fourier transform 5/29/2016 Visual Perception Modeling 11 Some Fourier Transform Pairs • Step function • Window function • sinc function x sinc ( x) sin( x) • Gaussian function 5/29/2016 Visual Perception Modeling 12 Filter Design • Design filters to accomplish particular goals • Lowpass filters – Reduce the amplitude of high-frequency components – Can reduce the visible effects of noise – Box filter – Triangle filter – High-frequency cutoff – Gaussian lowpass filter 5/29/2016 Visual Perception Modeling 13 Filter Design – cont. • Bandpass and bandstop filters • Highpass filters • Optimal filter design – In some sense, optimal of doing a particular job – Establish a criterion of performance and then maximize the criterion by proper selection of the impulse response – Wiener estimator – Wiener deconvolution 5/29/2016 Visual Perception Modeling 14 Other Transformations • Fourier transform is one of a number of linear transformations that are useful in image processing • Basis functions – How to represent an image by weighted sum of some functions of our choice? 5/29/2016 Visual Perception Modeling 15 Principal Component Analysis • Optimal representation with fewer basis functions – We want to design a set of basis functions such that we can reconstruct the original image with smallest possible error with a given number of basis functions 5/29/2016 Visual Perception Modeling 16 PCA for Face Recognition 5/29/2016 Visual Perception Modeling 17 PCA for Face Recognition – cont. First 20 principal components 5/29/2016 Visual Perception Modeling 18 PCA for Face Recognition – cont. Components with low eigenvalues 5/29/2016 Visual Perception Modeling 19 PCA for Face Recognition – cont. 5/29/2016 Visual Perception Modeling 20