wzyn20051216/matlab-agent-skills / skills/matlab-signal-vision-ai
matlab-signal-vision-ai
UnverifiedMATLAB workflow for signal processing, DSP, image processing, computer vision, lidar, medical imaging, deep learning, and AI-assisted reproducible experiments. Use for filters, spectrograms, wavelets, object detection, segmentation, classification, training, inference, generated figures, and for verifying signal or vision results — units, windows, Nyquist, filter response, BER/SNR, masks and boxes, IoU/precision/recall/F1, label alignment, transform invertibility, and split leakage.
MATLAB Signal Vision AI
Use this skill when the core artifact is a processed signal, image, video, point cloud, neural network, or AI experiment.
Skill Root
The installed directory containing this SKILL.md is the resource root. Add its
helper folder for the current MATLAB session only — never through savepath,
startup.m, or a global path edit:
skillRoot = '<absolute path to the installed matlab-signal-vision-ai Skill directory>';
addpath(fullfile(skillRoot, 'scripts'));
Resolve references/, tests/, and helper calls from that root, not from the
caller's working directory. If a shipped helper or reference is missing, report the
exact path and reinstall the Skill; do not scan the disk for a substitute.
Workflow
- 任务识别 — Identify data modality, sampling rate, image size, color space, labels, expected outputs. State the units (Hz vs rad/s), the analysis window, and whether the claim is steady-state or transient.
- 能力选择 —
r = mas_route(prompt); confirm this skill is primary and read the routed tool contract, not the whole toolbox. - 环境检查 — confirm toolbox availability (
mas_precheck) before specialized APIs; a missing toolbox degrades the capability, it does not silently change the method. - MATLAB 执行 — build a tiny reproducible subset first; save intermediate outputs so failures are inspectable.
- 独立验收 — grade with the measurement tools below, never by the figure. Reference values must come from outside the code under test (closed form, hand algebra, a second route).
- 错误恢复 — a
failorUNDETERMINEDis re-run with the specific fix (longer record, stated window, independent truth) and the attempt is recorded; it is not skipped. - 结果交付 — report numbers with units, window, regime and interval. SKIP/BLOCKED is never rewritten to PASS.
Preferred APIs
- Signal:
designfilt,kaiserord,butter/buttord,fir1,filter,freqz(always passfswhen you mean Hz),spectrogram,pwelch,findpeaks, wavelet functions. - Vision:
imread,im2gray,imresize,imbinarize,regionprops,detectSURFFeatures,bboxOverlapRatio. - Deep learning:
dlnetwork,trainnet, datastores,minibatchqueue, pretrained models when licensed. - Lidar/medical imaging: use domain importers and viewers when available; keep metadata.
Measurement tools
c = mas_measure_contract(x, fs, struct('freq',struct('value',50,'unit','Hz'), ...
'rms', 0.707, 'expected', struct('value', 0.5, 'tolRel', 1e-2)));
c.verdict c.checks c.ci c.freqHz c.freqRadPerSec c.powerLinear c.powerDb
f = mas_filter_check(b, a, struct('fs',1000,'type','lowpass','fpass',100,'fstop',250,'astopDb',40));
f.verdict f.achieved.stopbandDb f.achieved.groupDelaySamples f.achieved.maxPoleMag
s = mas_sample_check(x, fs, struct('fMax',200,'truth',struct('f',@(t) sin(2*pi*150*t))));
s.verdict s.measured.fPeak s.measured.aliasFoldHz s.measured.reconRelErr
g = mas_ber_check(588, 262144, struct('scheme','QPSK','ebn0dB',6));
g.verdict g.z g.berMeasured g.berTheory g.ci g.state
Insufficient evidence is not a refutation
Any counted claim (BER, failure rate, error fraction) goes through
mas_stat_verdict(nErr, n, spec), which names one of five states and lets the
state decide the verdict -- never the reverse:
| state | verdict | when |
|---|---|---|
sufficient-support | PASS | the interval excludes worse-than-claimed |
sufficient-refutation | FAIL | the interval excludes the claim |
insufficient-sample | UNDETERMINED | too few expected events to see the rate under test |
interval-straddles-target | UNDETERMINED | powered, and the threshold sits inside the interval |
criterion-missing | UNDETERMINED | nothing was claimed |
0 errors in 128 bits against a theoretical 2.4e-3 is the case that gets judged
wrong most often: the expected count is 0.31, so the run neither certifies nor
refutes anything, and it must not be reported as a refutation. The same zero
count against a target of 3e-2 does certify it (exact bound
1-(1-L)^(1/n) = 2.31e-2), so downgrading that to inconclusive throws away the
only evidence a rare-event test produces. State the claim as
struct('target', 3e-2) or struct('theory', 2.4e-3) and let the tool decide.
Every tool returns verdict (PASS | FAIL | UNDETERMINED) and per-check status
(pass | fail | na | undetermined). na = the contract does not apply; it does
not drag the verdict. undetermined = missing evidence or missing criterion and
forces UNDETERMINED: no acceptance criterion, no PASS.
Boundary rules these tools encode (each one a paid-for defect):
rms <= peakfor ANY signal; ratio ~sqrt(2) means sinusoid, ratio ~1 does not.- dB power is
10*log10; an applied20*log10is a factor-2-in-dB error, and -3.01 dB IS 0.5. - The windowed periodogram power identity
sum(x.^2.*w.^2)/sum(w.^2)is notmean(x.^2); the gap is taper bias, not a scaling bug. - Stopband = WORST over the whole band, not the value at
fstop: a comb null reads 61.8 dB at its token point and 11.3 dB where it actually matters. - Linear phase is graded only for symmetric/antisymmetric FIR; an IIR is
na, notfail, when the spec never asked. - Nyquist: the fold location is printed; a spectrum can confirm content below the claim but can never rule out content above
fs/2. - BER is graded by z-score against counting noise; zero errors certify an upper bound only; tolerances follow the error count, not taste.
- A frequency confidence interval is quoted only when blocks hold >= 10 cycles of the tone; below that the tool refuses a tight-and-wrong interval.
Full contract, worked examples, negative controls: references/signal-measurement-contract.md.
Vision verification tools
s = mas_seg_check(predMask, trueMask, struct('iouThreshold',0.5, ...
'thresholdSource','validation', ...
'metrics', struct('name','f1','value',0.82,'tolRel',1e-3)));
s.verdict s.metrics.f1 s.perClass s.objects s.objectMetrics s.thresholdSweep
d = mas_detect_check(gtBox, gtLab, predBox, predLab, scores, struct( ...
'iouThreshold',0.5,'imageSize',[480 640], ...
'scoreThreshold',0.3,'scoreThresholdSource','validation', ...
'metrics', struct('name','f1','value',0.75,'tolRel',1e-3)));
d.verdict d.metrics d.perClass d.matches d.scoreSweep
v = mas_image_contract(I, struct('expectedSize',[256 256 3], ...
'expectedClass','uint8','expectedRange',[0 255],'labels',L, ...
'normalization', struct('original',I0,'inverse',@inv), ...
'transform',T,'points',P,'pointsTransformed',Q, ...
'edgeMap',E,'edgeTruth',E0,'edgeF1Min',0.9, ...
'split', struct('trainIdx',tr,'valIdx',va,'sampleHash',h)));
v.verdict v.image v.labels v.normalization v.transform v.edge v.split
- Pixel accuracy is not segmentation quality: 95 % background scores 95 % accuracy while finding nothing. Per-class values are always reported, and an accuracy-only claim on imbalanced data is UNDETERMINED.
- Pixel IoU is not object IoU. Objects are matched one-to-one at an EXPLICIT IoU threshold; a second prediction on one object is a false positive, and an object-metric claim without a stated threshold is UNDETERMINED.
- Empty conventions are stated, never 0/0-as-good: empty prediction vs empty truth is a perfect match by convention; empty truth with predictions is IoU 0.
- A score/binarization threshold tuned on the set being reported is a leak (
testfails, no stated source is UNDETERMINED). A metric without a tolerance is not a criterion. - Detection AP is all-point interpolated per class; mAP averages only classes with GT. A predicted label missing from the GT set raises a label-alignment warning, because a name-mapping bug looks exactly like an invented class.
- Image checks are numeric: size/dtype/range, label alignment, normalization round-trip, invertible affine geometry, edge/feature agreement with an independent truth within a pixel tolerance, and split leakage by index AND content hash AND group.
- "The picture looks right" is never a criterion; a missing truth or criterion is UNDETERMINED.
Full contract: references/vision-verification-contract.md.
Legacy metrics tools
r = mas_spectrum_check(x, fs, struct('window','hann','nfft',[]));
r.leakage r.cycles r.concentration r.parsevalErr r.taperBias r.warnings
o = mas_ofdm_ber([0 2 4 6]); o.berSim o.berTheory o.relErr o.verdict
m = mas_image_metrics(predMask, trueMask);
m.iou m.dice m.precision m.recall m.perObject m.objectIouWorst
s = mas_split_seed(labels, struct('trainFrac',0.7,'seed',0,'stratify',true));
s.trainIdx s.valIdx s.noOverlap s.noGap s.reproducible s.classBalance
What they pin:
mas_spectrum_check: a half-bin offset drops concentration 1.000 -> 0.408 (measured); Parseval is tested against the window-weighted identity, taper bias reported separately.mas_ofdm_ber: two-sided against0.5*erfc(sqrt(Eb/N0))— a BER below theory is a failure (noise too small), and the CP penalty belongs at the transmitter (receiver-side application shifts 0 dB BER 0.0786 -> 0.102).mas_image_metrics: per-object matching by argmax over truth labels, not by index; a mask with the right area in the wrong place must score IoU 0.mas_split_seed:noOverlap/noGap/reproducibleasserted;classBalanceis a number to read, not a boolean to trust.
Validation
Choose relevant checks beyond the tools:
- Image dimensions, pixel range, mask area, IoU, accuracy, F1.
- Training loss decreases on a smoke subset.
- Inference output has expected class names, boxes, or mask shape.
- Exported figure or video is nonempty — necessary, never sufficient.
Reproducibility
Record random seed, split policy, pretrained model version, GPU availability, and dataset hash or URL.
Tests: tests/test_signal_vision.m, tests/test_signal_contract.m (42/42) and
tests/test_vision_contract.m (103/103) — every metric carries a negative control, and
expected values come from hand algebra, closed forms or MATLAB design generators, never
from the code under test.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| planning-and-task-breakdownaddyosmani/agent-skillsBreaks work into ordered tasks. Use when you have a spec or clear requirements and need to break work into implementable tasks. Use when a task feels too large to start, when you need to estimate scope, or when parallel work is possible. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| screenpipethedotmack/claude-memSearch screen and audio history recorded by Screenpipe and optionally save selected context into claude-mem. Use when the user asks to recall desktop activity, a meeting, or something seen in another app through Screenpipe. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |
| contact-widgetnexu-io/open-designSelf-contained floating chat widget with welcome screen, social links, meeting button, and message input. Single HTML file, zero dependencies. | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |
| meeting-notesnexu-io/open-designMeeting notes page — title bar with attendees, agenda checklist, decisions block, action items table with owners + dates, and a "next meeting" footer. Use when the brief mentions "meeting notes", "minutes", "1:1 notes", "all-hands recap", or "会议纪要". | Stars: 108k | Updated | Claude CodeCodex | Verified | Install |