Repository navigation
Expand file tree
/
Copy pathvision.html
More file actions
4 lines (4 loc) · 4.33 KB
/
Copy pathvision.html
File metadata and controls
4 lines (4 loc) · 4.33 KB
1
2
3
4
<!DOCTYPE html>
<html lang="en"><head><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1">
<title>Concrete Vision Sorter | Stephen Jones</title><meta name="description" content="Stephen Jones: robotics and mechatronics engineer. Georgia Tech BSME, OMSCS (Computational Perception and Robotics).">
<meta property="og:type" content="website"><meta property="og:title" content="Stephen Jones | Robotics Engineer"><meta name="theme-color" content="#7ec8f0"><link rel="icon" href="assets/opt/headshot.jpeg"><link rel="stylesheet" href="style.css"></head><body data-bg="sky"><main><section class="panel" data-bg="sky"><h1>Concrete vision sorter</h1><p class="lead">I rebuilt a published crack-detection approach with a lighter pipeline: a frozen VGG16 turns each image into 512 numbers, PCA shows how much of that is redundant, and a logistic regression makes the call.</p><div class="chips"><span class="chip"><b>99.5 to 99.8%</b> accuracy across runs</span><span class="chip">2,000 images, balanced</span><span class="chip"><b>~80%</b> fewer dimensions</span></div></section><section class="panel" data-bg="sky"><h2><svg class="ico" aria-hidden="true"><use href="#i-scan-search"/></svg>Pipeline</h2><ol class="steps"><li><b>Data:</b> 1,000 crack and 1,000 non-crack images, resized to 224x224x3.</li><li><b>Features:</b> pre-trained VGG16 (ImageNet, no top layers, average pooling) gives a 512-d vector per image.</li><li><b>Split:</b> 70% train, 30% test (600 held-out images).</li><li><b>PCA:</b> about 101-103 components explain 95% of variance, so roughly 80% of the dimensions are redundant.</li><li><b>Classifier:</b> logistic regression on the 512 features.</li></ol></section><section class="panel" data-bg="ocean2"><h2>Results</h2><div class="figs"><figure><div class="fi"><img decoding="async" src="assets/Vision%20Perception%20Project/Confusion%20Matrix.png" alt="Confusion matrix: 297 and 300 correct, 1 false positive, 2 false negatives." loading="lazy"></div><figcaption>Confusion matrix: 297 and 300 correct, 1 false positive, 2 false negatives.</figcaption></figure><figure><div class="fi"><img decoding="async" src="assets/Vision%20Perception%20Project/thethreedifferentcracksinoneimage.png" alt="Representative cracks from the three K-Means clusters." loading="lazy"></div><figcaption>Representative cracks from the three K-Means clusters.</figcaption></figure></div><table><thead><tr><th>Crack style (K-Means)</th><th>Test images</th><th>Accuracy</th></tr></thead><tbody><tr><th>Style 1 (thick, dark)</th><td>66</td><td>100%</td></tr><tr><th>Style 2 (jagged)</th><td>102</td><td>100%</td></tr><tr><th>Style 3 (hairline)</th><td>134</td><td>98.5%</td></tr></tbody></table><p>Clustering the crack images into three visual styles shows where the model is weakest: fine hairline cracks. That is the failure mode I would test first on new data.</p></section><section class="panel" data-bg="ocean2"><h2>Why the numbers move between runs</h2><img src="assets/gifs/coding_head_smash_computer.gif" alt="" style="float:right;width:64px;margin-left:1rem;image-rendering:pixelated"><p>Each run randomly samples 1,000 images per class and then splits them 70/30, so accuracy shifts by a few tenths of a percent depending on which images land in the test set. The confusion matrix shown is one run (597 of 600 correct). Across runs I have seen 99.5% to 99.8%, so I report a range instead of a single best result.</p><!-- TODO: replace with mean and standard deviation from crack_eval_multiseed.py --></section><section class="panel" data-bg="ocean2"><h2>What this does and does not show</h2><ul><li>Pre-trained features transfer well: no deep network training was needed.</li><li>The test set comes from the same dataset. The source paper also tested on pavement, concrete buildings, and brickwork, and accuracy dropped on unseen materials. That generalization test is the next step.</li><li>For a robot, the same pipeline could run on a camera feed to flag anomalies.</li></ul><div class="callout warn"><b>Source:</b> built from a graded ME 4710 homework (HW9) that re-implements ideas from a published crack-detection paper.</div></section><section class="panel pager"><a class="gel blue" href="me2110.html">Previous</a><a class="gel blue" href="index.html#projects">All projects</a><a class="gel blue" href="abs.html">Next</a></section></main><script src="main.js"></script></body></html>