The Central Idea
A signal-processing or imaging system can often be represented as
\[ \boxed{ \text{source} \longrightarrow \text{measurement system} \longrightarrow \text{sampled data} \longrightarrow \text{processing} \longrightarrow \text{estimate or decision} }. \]The observed data are not usually identical to the underlying signal. Sensors, optics, sampling, noise, and transmission all modify what is measured.
Signals as Functions
A continuous-time signal can be written
\[ \boxed{x=x(t)}. \]A one-dimensional spatial signal might be \(x(s)\), while an image is commonly modeled as
\[ \boxed{f=f(x,y)}. \]Video and volumetric imaging introduce additional spatial and temporal coordinates.
Discrete Signals
Digital systems work with sampled signals. A discrete-time signal is written
\[ \boxed{x[n]}, \]where \(n\) is an integer sample index.
A digital image is represented by pixel values
\[ \boxed{f[m,n]}. \]Sampling
If a continuous signal \(x(t)\) is sampled every \(T_s\) seconds,
\[ \boxed{x[n]=x(nT_s)}. \]The sampling frequency is
\[ \boxed{f_s=\frac{1}{T_s}}. \]Nyquist Sampling Condition
If a signal contains no frequency above \(f_{\max}\), ideal reconstruction requires a sampling rate greater than twice that maximum frequency:
\[ \boxed{f_s>2f_{\max}}. \]This is the classical Nyquist sampling condition under ideal band-limited assumptions.
Aliasing
Aliasing occurs when sampling is too slow to distinguish high-frequency behavior correctly.
Frequencies above the representable range can appear as lower frequencies in the sampled data.
Aliasing is not repaired simply by collecting more samples after the original measurement has already been undersampled.
Anti-Aliasing Filters
Before sampling, an analog low-pass filter is often used to suppress frequency components above the usable sampling band.
The filter and sampling rate should therefore be designed together.
Quantization
Digital measurement also discretizes amplitude. If a sensor uses \(b\) bits, it can represent up to
\[ \boxed{2^b} \]quantization levels.
Finite resolution introduces quantization error even when sampling in time or space is adequate.
Signal Energy
For a continuous signal, an energy measure is
\[ \boxed{ E = \int_{-\infty}^{\infty} |x(t)|^2\,dt }. \]For a discrete signal,
\[ \boxed{ E = \sum_{n=-\infty}^{\infty} |x[n]|^2 }. \]Signal Power
Long-duration or periodic signals are often characterized by average power rather than finite total energy.
For a continuous signal,
\[ \boxed{ P = \lim_{T\to\infty} \frac{1}{2T} \int_{-T}^{T}|x(t)|^2\,dt }. \]Linear Systems
A system \(H\) is linear if
\[ \boxed{ H[a x_1+b x_2] = aH[x_1]+bH[x_2] }. \]Linearity allows superposition and makes many signal-processing systems much easier to analyze.
Time Invariance
A system is time invariant if delaying the input delays the output by the same amount without changing its form.
Linear time-invariant systems are especially important because their behavior is determined completely by their impulse response.
Impulse Response
Let \(h(t)\) be the response of a continuous-time LTI system to an impulse. Then an arbitrary input \(x(t)\) produces
\[ \boxed{ y(t) = (x*h)(t) }. \]Convolution
Continuous convolution is
\[ \boxed{ (x*h)(t) = \int_{-\infty}^{\infty} x(\tau)h(t-\tau)\,d\tau }. \]Discrete convolution is
\[ \boxed{ y[n] = \sum_{k=-\infty}^{\infty} x[k]h[n-k] }. \]Convolution in Imaging
A spatially invariant imaging system can be modeled by
\[ \boxed{ g(x,y) = (f*h)(x,y) +n(x,y) }. \]Here \(f\) is the true image, \(h\) the point-spread function, \(g\) the measured image, and \(n\) measurement noise.
Point-Spread Function
The point-spread function describes how an ideal point source is blurred by the imaging system.
A narrow point-spread function corresponds to better spatial resolution than a broad one, all else equal.
Filtering
A filter modifies signal content to emphasize, suppress, estimate, or separate desired components.
Common categories include
- low-pass filters;
- high-pass filters;
- band-pass filters;
- band-stop filters;
- smoothing filters;
- edge-enhancing filters.
Moving-Average Filter
A simple \(M\)-point moving average is
\[ \boxed{ y[n] = \frac{1}{M} \sum_{k=0}^{M-1} x[n-k] }. \]It smooths rapid variation but also blurs sharp changes.
Finite Impulse Response Filters
A finite impulse response, or FIR, filter has
\[ \boxed{ y[n] = \sum_{k=0}^{M} b_kx[n-k] }. \]The output depends on a finite set of input samples.
Infinite Impulse Response Filters
An IIR filter may use both current and past inputs and past outputs:
\[ \boxed{ y[n] = \sum_{k=0}^{M}b_kx[n-k] - \sum_{k=1}^{N}a_ky[n-k] }. \]Feedback can create efficient filters but also introduces stability concerns.
Frequency Domain
Signals can be represented as combinations of oscillatory components. The Fourier transform converts a time-domain signal into a frequency-domain representation.
\[ \boxed{ X(\omega) = \int_{-\infty}^{\infty} x(t)e^{-i\omega t}\,dt }. \]Inverse Fourier Transform
Under suitable conditions, the time-domain signal can be reconstructed from its transform:
\[ \boxed{ x(t) = \frac{1}{2\pi} \int_{-\infty}^{\infty} X(\omega)e^{i\omega t}\,d\omega }. \]Why Frequency Representations Help
Oscillations, resonance, periodic interference, bandwidth, and filtering can be easier to understand in frequency space than in the original time domain.
Convolution also becomes multiplication:
\[ \boxed{ \mathcal{F}\{x*h\} = X(\omega)H(\omega) }. \]Discrete Fourier Transform
For \(N\) samples, the discrete Fourier transform is
\[ \boxed{ X[k] = \sum_{n=0}^{N-1} x[n]e^{-i2\pi kn/N} }. \]Inverse Discrete Fourier Transform
\[ \boxed{ x[n] = \frac{1}{N} \sum_{k=0}^{N-1} X[k]e^{i2\pi kn/N} }. \]The DFT converts a finite sample sequence into discrete frequency bins.
Fast Fourier Transform
The FFT is an efficient algorithm for computing the discrete Fourier transform.
It reduces the computational cost from direct \(O(N^2)\) evaluation to approximately
\[ \boxed{O(N\log N)} \]for common implementations.
Frequency Resolution
For \(N\) samples collected at sampling rate \(f_s\), the frequency-bin spacing is
\[ \boxed{ \Delta f = \frac{f_s}{N} }. \]Longer observation windows improve frequency resolution for fixed sampling rate.
Spectral Leakage
A finite observation window generally does not contain an exact integer number of periods of each frequency component.
This causes energy to spread across nearby frequency bins, a phenomenon called spectral leakage.
Windowing
Window functions taper the observed data before transformation to reduce discontinuities at the edges of the sample interval.
Window selection trades spectral leakage against frequency resolution and amplitude characteristics.
Power Spectral Density
A power spectral density describes how signal power is distributed across frequency.
It is useful for vibration, acoustics, communications, biomedical signals, turbulence, and noise analysis.
Time-Frequency Analysis
A single Fourier transform loses information about when frequency components occur.
Time-frequency methods such as the short-time Fourier transform analyze local windows:
\[ \boxed{ X(\tau,\omega) = \int x(t)w(t-\tau)e^{-i\omega t}\,dt }. \]Wavelets
Wavelet methods represent signals using localized functions at multiple scales.
They are useful for transient events, compression, denoising, multiresolution analysis, and signals whose frequency content changes through time.
Noise
A measurement model often takes the form
\[ \boxed{ y=s+n }. \]Here \(s\) is the desired signal and \(n\) noise.
Noise may arise from electronics, photon statistics, environmental interference, digitization, motion, or reconstruction error.
Signal-to-Noise Ratio
A power-based signal-to-noise ratio is
\[ \boxed{ \mathrm{SNR} = \frac{P_s}{P_n} }. \]In decibels,
\[ \boxed{ \mathrm{SNR}_{\text{dB}} = 10\log_{10} \left( \frac{P_s}{P_n} \right) }. \]White Noise
Ideal white noise has constant power spectral density across frequency.
Real measurement noise may instead be colored, correlated, non-Gaussian, or dependent on signal intensity.
Gaussian Noise
A common model assumes
\[ \boxed{ n\sim\mathcal{N}(0,\sigma^2) }. \]Gaussian noise is mathematically convenient but should be treated as an assumption to be checked, not a universal property of sensors.
Poisson Noise
Photon-counting and event-counting systems often use a Poisson model:
\[ \boxed{ Y\sim\operatorname{Poisson}(\lambda) }. \]The variance then scales with the mean under the ideal Poisson assumption.
Denoising
Denoising seeks an estimate \(\hat s\) of a desired signal from noisy observations.
Smoothing can reduce noise but may also remove edges, transients, or fine structure. The optimal tradeoff depends on the application.
Matched Filtering
When a known signal shape must be detected in additive noise, a matched filter is designed to correlate strongly with that expected signal.
It is widely used in radar, communications, sonar, and event detection.
Correlation
Cross-correlation measures similarity as one signal is shifted relative to another:
\[ \boxed{ R_{xy}[\ell] = \sum_n x[n]y[n+\ell] }. \]Peaks can estimate delay or alignment.
Autocorrelation
Autocorrelation compares a signal with shifted versions of itself:
\[ \boxed{ R_{xx}[\ell] = \sum_n x[n]x[n+\ell] }. \]Periodic or repeating structure can produce recurring peaks.
Image Formation
An imaging model describes how a scene generates measured pixels or detector readings.
A general linear model is
\[ \boxed{ \mathbf{y} = A\mathbf{x} + \mathbf{n} }. \]The vector \(\mathbf{x}\) represents the unknown image or object, \(A\) the imaging operator, and \(\mathbf{n}\) measurement noise.
Blur
Blur can arise from optics, motion, defocus, scattering, detector size, or reconstruction.
A spatially invariant blur model is
\[ \boxed{ g=f*h+n }. \]Spatial Resolution
Spatial resolution describes the ability to distinguish nearby features.
Pixel spacing alone does not determine resolution. Optics, detector response, blur, noise, and reconstruction all matter.
Contrast
Contrast measures differences in signal intensity between structures or regions.
A high-resolution system can still have poor detectability if contrast is weak relative to noise.
Dynamic Range
Dynamic range describes the span between the smallest and largest signal levels that can be measured usefully.
Saturation at the high end and quantization or noise at the low end limit recoverable information.
Image Gradients
The image gradient is
\[ \boxed{ \nabla f = \begin{bmatrix} \partial f/\partial x\\ \partial f/\partial y \end{bmatrix} }. \]Large gradient magnitude often indicates edges or rapid spatial change.
Edge Detection
Edge detectors approximate spatial derivatives and identify locations of strong change.
Noise can also create large derivatives, so smoothing and derivative estimation must often be designed together.
Image Segmentation
Segmentation divides an image into meaningful regions or objects.
Methods may use thresholding, clustering, graph models, region growth, optimization, or machine learning.
Thresholding
A simple binary threshold is
\[ \boxed{ b(x,y) = \begin{cases} 1,&f(x,y)\ge T,\\ 0,&f(x,y)Geometric Transformations
Image registration may require translations, rotations, scaling, affine transformations, or nonlinear warps.
A two-dimensional affine transform can be written
\[ \boxed{ \mathbf{x}' = A\mathbf{x} + \mathbf{b} }. \]Image Registration
Registration aligns two or more images into a common coordinate system.
The optimization may minimize a mismatch measure:
\[ \boxed{ \theta^* = \arg\min_\theta D\left( I_1, T_\theta(I_2) \right) }. \]Tomographic Imaging
Tomography reconstructs internal structure from measurements acquired along many directions or trajectories.
A simplified linear model is
\[ \boxed{ \mathbf{y} = A\mathbf{x} }. \]Each measurement is influenced by multiple unknown spatial elements.
Inverse Problems
Imaging reconstruction is often an inverse problem: measurements are known and the underlying scene or object must be inferred.
Direct inversion can be unstable when measurements are incomplete, noisy, or insensitive to certain image components.
Least-Squares Reconstruction
A basic reconstruction may minimize
\[ \boxed{ \hat{\mathbf{x}} = \arg\min_{\mathbf{x}} \|A\mathbf{x}-\mathbf{y}\|_2^2 }. \]This seeks the image whose predicted measurements best match observed data in the least-squares sense.
Regularized Reconstruction
To stabilize an ill-posed inverse problem, add a regularization term:
\[ \boxed{ \hat{\mathbf{x}} = \arg\min_{\mathbf{x}} \left[ \|A\mathbf{x}-\mathbf{y}\|_2^2 + \lambda R(\mathbf{x}) \right] }. \]The regularizer encodes assumptions or preferences about plausible solutions.
Tikhonov Regularization
A common choice is
\[ \boxed{ R(\mathbf{x}) = \|L\mathbf{x}\|_2^2 }. \]The matrix \(L\) may penalize large values, gradients, or other undesirable variation.
Total Variation
Total-variation regularization penalizes spatial gradient magnitude and can preserve sharp edges better than simple quadratic smoothing.
A schematic form is
\[ \boxed{ R(f) = \int |\nabla f|\,dA }. \]Deconvolution
Deconvolution seeks to undo blur:
\[ g=f*h+n. \]In the frequency domain,
\[ G=FH+N. \]Naively dividing by \(H\) can amplify noise where \(H\) is small.
Wiener Filtering
Wiener filtering balances inverse filtering against noise amplification using statistical information about signal and noise.
It illustrates a recurring principle: the best reconstruction depends on both the forward model and uncertainty.
Compressed Sensing
Under suitable conditions, a signal that is sparse in an appropriate representation can be reconstructed from fewer measurements than traditional sampling intuition might suggest.
A common optimization form is
\[ \boxed{ \min_{\mathbf{x}} \|\mathbf{x}\|_1 \quad \text{subject to} \quad A\mathbf{x}=\mathbf{y} }. \]Sparsity
A sparse signal has relatively few significant coefficients in some basis or transform domain.
Sparsity is a modeling assumption that can enable denoising, compression, and reconstruction.
Image Compression
Compression reduces storage or transmission requirements by representing data more efficiently.
Lossless compression permits exact reconstruction. Lossy compression accepts controlled information loss for greater reduction.
Transform Coding
Many compression methods transform data into coefficients whose importance can be ranked or quantized selectively.
The basic structure is
\[ \boxed{ \text{signal} \rightarrow \text{transform} \rightarrow \text{coefficient reduction} \rightarrow \text{encoding} }. \]Sampling in Images
Digital images sample a continuous optical field on a spatial grid.
Insufficient spatial sampling can create aliasing patterns such as moiré, jagged edges, or false texture.
Interpolation
Resizing, geometric transformation, and registration require values between sampled pixels.
Common interpolation methods include nearest-neighbor, bilinear, bicubic, and higher-order schemes.
Interpolation Does Not Create New Measurement Information
Interpolation estimates intermediate values from existing samples. It can make an image appear smoother or larger but cannot recover information that was never measured unless additional model assumptions are introduced.
Multichannel Signals
Many systems measure several channels simultaneously:
\[ \boxed{ \mathbf{x}[n] = \begin{bmatrix} x_1[n]\\ x_2[n]\\ \vdots\\ x_m[n] \end{bmatrix} }. \]Examples include color images, microphone arrays, multispectral imaging, multilead biomedical signals, and sensor arrays.
Sensor Arrays
Multiple sensors can exploit spatial differences in arrival time, phase, or amplitude to estimate direction or source location.
Array processing appears in radar, sonar, acoustics, wireless communications, and imaging.
Beamforming
A simple weighted array output is
\[ \boxed{ y[n] = \mathbf{w}^H\mathbf{x}[n] }. \]The weights \(\mathbf{w}\) are chosen to enhance signals from desired directions and suppress interference.
Feature Extraction
Signals can be summarized by features such as
- mean and variance;
- spectral peaks;
- band power;
- edge density;
- texture measures;
- shape descriptors;
- wavelet coefficients;
- learned representations.
Detection
Detection asks whether a target signal is present.
A test statistic \(T(\mathbf{y})\) can be compared with a threshold:
\[ \boxed{ T(\mathbf{y}) \gtrless_{\mathcal{H}_0}^{\mathcal{H}_1} \eta }. \]The threshold determines the tradeoff between missed detections and false alarms.
Receiver Operating Characteristics
A receiver operating characteristic curve compares detection probability and false-alarm probability across thresholds.
The operating threshold should be selected according to the real costs of errors.
Estimation
Estimation seeks a continuous unknown quantity such as frequency, position, amplitude, phase, delay, or image intensity.
A general estimate is
\[ \boxed{ \hat\theta = g(\mathbf{y}) }. \]Maximum-Likelihood Estimation
If a probability model \(p(\mathbf{y}\mid\theta)\) is available,
\[ \boxed{ \hat\theta_{\text{ML}} = \arg\max_\theta p(\mathbf{y}\mid\theta) }. \]Bayesian Estimation
Bayesian estimation combines a likelihood with prior information:
\[ \boxed{ p(\theta\mid\mathbf{y}) \propto p(\mathbf{y}\mid\theta)p(\theta) }. \]The result is a posterior distribution rather than only one estimate.
Image and Signal Models as Forward Models
A forward model predicts what the sensor would measure from a hypothesized source:
\[ \boxed{ \mathbf{x} \longrightarrow A\mathbf{x} \longrightarrow \mathbf{y} }. \]Reconstruction runs this logic in reverse.
Model-Based and Data-Driven Processing
Traditional signal processing relies heavily on explicit forward models, transforms, probability, and optimization.
Modern systems may instead learn denoisers, detectors, reconstructions, or features from data. Hybrid methods combine both.
Learned Reconstruction
A learned model may estimate
\[ \boxed{ \hat{\mathbf{x}} = f_\theta(\mathbf{y}) }. \]Such models can be powerful but must be validated against distribution shift, artifacts, missing structures, and physically implausible outputs.
Physics-Informed Reconstruction
A hybrid method can combine data fidelity with a learned prior:
\[ \boxed{ \hat{\mathbf{x}} = \arg\min_{\mathbf{x}} \left[ \|A\mathbf{x}-\mathbf{y}\|^2 + \lambda R_\theta(\mathbf{x}) \right] }. \]The forward model constrains consistency with measurements while the learned term captures structure from data.
Calibration
Imaging systems often require calibration of gain, offset, geometric alignment, timing, detector sensitivity, optical distortion, spectral response, or point-spread function.
Calibration parameters are part of the measurement model.
Flat-Field and Gain Correction
If detector elements respond differently to the same illumination, correction may use calibrated gain and offset parameters:
\[ \boxed{ x_{\text{corrected}} = \frac{x-b}{g} }. \]Here \(b\) is an offset and \(g\) a gain factor.
Registration Error
Misalignment between signals or images can create artificial differences, blurred averages, or incorrect fused measurements.
Registration uncertainty should be considered when comparing spatial datasets.
Resolution versus Noise
Higher nominal resolution does not always improve useful information. Smaller samples or pixels may collect less signal and therefore have lower signal-to-noise ratio.
Imaging design often involves a tradeoff among spatial resolution, temporal resolution, noise, dose, exposure time, and computational cost.
Temporal Resolution
Temporal resolution determines how quickly changing events can be distinguished.
Shorter exposure or faster sampling may improve timing but reduce the number of collected signal events per sample.
Uncertainty Propagation
Measurement uncertainty can propagate through filtering and reconstruction. If
\[ \mathbf{y}=A\mathbf{x}+\mathbf{n}, \]uncertainty in \(\mathbf{n}\), \(A\), and calibration parameters all affect the final estimate.
Artifacts
An artifact is structure in processed data that does not correspond to the underlying object or signal of interest.
Artifacts can arise from undersampling, motion, saturation, filtering, reconstruction, calibration errors, model mismatch, or learned processing.
Verification
Verification asks whether the processing algorithm and numerical implementation perform the intended mathematics correctly.
Useful checks include
- known-signal tests;
- impulse-response tests;
- frequency-response checks;
- unit and scaling checks;
- synthetic reconstruction tests;
- convergence studies;
- independent metric calculations.
Validation
Validation asks whether the full measurement and processing model provides reliable information about the real system.
A denoiser may reduce pixel variance while removing real structure. A reconstruction may look sharp while introducing bias. A detector may perform well on laboratory data but fail under different operating conditions.
Phantoms and Reference Targets
Imaging systems are often validated using objects or signals with known properties.
Reference targets can measure spatial resolution, contrast, geometric accuracy, noise response, or reconstruction bias.
Sensitivity Analysis
Important sensitivities may include performance with respect to
- sampling rate;
- noise level;
- blur width;
- filter parameters;
- regularization strength;
- registration error;
- sensor gain;
- model mismatch.
Example: Sampling Rate
A sensor is sampled every
\[ T_s=0.002\ \text{s}. \]Then
\[ f_s = \frac{1}{0.002} \] \[ \boxed{f_s=500\ \text{Hz}}. \]Example: Nyquist Check
Suppose the highest signal frequency is
\[ f_{\max}=180\ \text{Hz}. \]The ideal Nyquist requirement is
\[ f_s>2(180) \] \[ \boxed{f_s>360\ \text{Hz}}. \]A \(500\)-Hz sampling rate clears this ideal threshold, though real systems also require transition margin for anti-aliasing filters.
Example: Moving Average
For samples
\[ 2,\quad5,\quad8, \]a three-point moving average is
\[ \frac{2+5+8}{3} \] \[ \boxed{5}. \]Example: SNR in Decibels
If signal power is \(100\) and noise power is \(1\),
\[ \mathrm{SNR} = \frac{100}{1} =100. \]Therefore,
\[ \mathrm{SNR}_{\text{dB}} = 10\log_{10}(100) \] \[ \boxed{20\ \text{dB}}. \]Example: Frequency Resolution
Suppose
\[ f_s=1000\ \text{Hz} \]and
\[ N=2000. \]Then
\[ \Delta f = \frac{1000}{2000} \] \[ \boxed{\Delta f=0.5\ \text{Hz}}. \]Common Error: Ignoring Aliasing
Sampling a signal below the required rate can create false low-frequency content that appears to be real.
Anti-alias filtering and adequate sampling must be designed before digitization.
Common Error: Equating More Pixels with More Resolution
Pixel count can exceed the resolving capability of the optics or sensor.
Useful resolution depends on the complete imaging chain, not only the digital grid.
Common Error: Smoothing Away the Signal
A filter that reduces noise may also suppress narrow peaks, sharp edges, or short-duration events.
Denoising quality should be evaluated against the task, not merely against visual smoothness.
Common Error: Naive Inverse Filtering
Dividing by a small transfer function can strongly amplify noise.
Deconvolution and inverse problems usually require regularization or statistical modeling.
Common Error: Treating Interpolation as New Information
Enlarging an image does not by itself recover spatial detail beyond what the original measurements contain.
Common Error: Ignoring Calibration
Gain error, geometric distortion, timing offset, or sensor nonuniformity can produce systematic errors that filtering cannot remove correctly unless the measurement model accounts for them.
Common Error: Validating Only on Synthetic Data
Synthetic tests are valuable for verification because the true signal is known. They do not establish that the forward model accurately represents real sensors and real operating conditions.
Physical validation is also required.
A Reliable Modeling Procedure
- Define the signal, image, detection, or reconstruction objective.
- Specify the source, measurement system, sampling process, and noise model.
- Identify relevant time, space, and frequency scales.
- Choose sampling rates and anti-aliasing strategies before digitization.
- Model blur, filtering, geometry, calibration, and other sensor effects that materially alter measurements.
- Select transforms, filters, estimators, or reconstruction methods matched to the task.
- Regularize inverse problems when direct inversion is unstable.
- Verify algorithms using analytical signals and synthetic data with known truth.
- Validate using physical measurements, reference targets, and representative operating conditions.
- Perform sensitivity and uncertainty analysis before relying on processed outputs for consequential decisions.
Practice 1: Sampling Frequency
A signal is sampled every \(0.004\) second. Find the sampling frequency.
Practice 2: Nyquist Requirement
A signal contains frequencies up to \(75\) Hz. What ideal minimum sampling rate must be exceeded?
Practice 3: Moving Average
Find the four-point moving average of
\[ 3,\quad7,\quad9,\quad5. \]Practice 4: Signal-to-Noise Ratio
Signal power is \(40\), and noise power is \(2\). Find the linear SNR.
Practice 5: SNR in Decibels
Convert the SNR from Practice 4 to decibels.
Practice 6: Frequency Resolution
A record contains \(N=4096\) samples at \(f_s=2048\) Hz. Find the DFT frequency spacing.
Practice 7: Quantization Levels
How many amplitude levels can an ideal \(12\)-bit converter represent?
Practice 8: Convolution Length
A finite signal has \(N=10\) samples and an FIR filter has \(M=4\) coefficients. What is the length of their full linear convolution?
Practice 9: Detection Threshold
A detector lowers its decision threshold. Qualitatively, what usually happens to both detection probability and false-alarm probability?
Practice 10: Choose the Better Model
A blurred low-light image is restored by dividing its Fourier transform by the measured optical transfer function. Frequencies where the transfer function is nearly zero become dominated by amplified noise. Explain why direct inversion is inadequate and identify a better modeling direction.
Practice Answers
Practice 1
\[ f_s = \frac{1}{0.004} \] \[ \boxed{f_s=250\ \text{Hz}}. \]Practice 2
\[ f_s>2f_{\max} \] \[ f_s>2(75) \] \[ \boxed{f_s>150\ \text{Hz}}. \]Practice 3
\[ \frac{3+7+9+5}{4} = \frac{24}{4} \] \[ \boxed{6}. \]Practice 4
\[ \mathrm{SNR} = \frac{40}{2} \] \[ \boxed{\mathrm{SNR}=20}. \]Practice 5
\[ \mathrm{SNR}_{\text{dB}} = 10\log_{10}(20) \] \[ \boxed{\mathrm{SNR}_{\text{dB}}\approx13.0\ \text{dB}}. \]Practice 6
\[ \Delta f = \frac{2048}{4096} \] \[ \boxed{\Delta f=0.5\ \text{Hz}}. \]Practice 7
\[ 2^{12}=4096. \] \[ \boxed{4096\text{ levels}}. \]Practice 8
Full linear convolution of sequences of lengths \(N\) and \(M\) has length
\[ N+M-1. \]Therefore,
\[ 10+4-1 = \boxed{13\text{ samples}}. \]Practice 9
Lowering the threshold generally makes the detector more willing to declare a target present. Therefore, both detection probability and false-alarm probability usually increase.
Practice 10
Direct inversion is unstable because dividing by a transfer function near zero magnifies measurement noise enormously. A better approach uses regularized deconvolution, Wiener filtering, Bayesian reconstruction, or another inverse method that balances measurement fit against noise and prior information.
What to Remember
- signals and images are mathematical functions or arrays measured through imperfect instruments;
- sampling rate and anti-alias filtering determine what frequency content can be represented reliably;
- convolution describes the action of linear shift-invariant systems and image blur;
- Fourier methods translate signals between time or space and frequency representations;
- filtering changes signal content and always involves tradeoffs among noise, resolution, delay, and distortion;
- noise models should reflect the actual measurement process rather than being assumed automatically;
- imaging reconstruction is an inverse problem and often requires regularization;
- resolution, contrast, dynamic range, sampling, and noise are distinct properties of an imaging system;
- model-based and learned processing can be combined when measurement physics and data-driven structure are both useful;
- verification, physical validation, calibration, sensitivity analysis, and uncertainty quantification are essential before processed signals support consequential conclusions.
Where We Go Next
Imaging and signal-processing models show how mathematical representations can transform raw measurements into estimates, features, and decisions. Similar modeling questions arise in social systems, where observations are noisy, interactions occur through networks and institutions, and policy changes can alter the behavior being modeled.
The next lesson, Social and Policy Systems, develops models of populations, institutions, incentives, networks, diffusion, policy interventions, causal effects, uncertainty, and decision tradeoffs in social systems.