|
3 | 3 | <head> |
4 | 4 | <meta charset="UTF-8" /> |
5 | 5 | <meta name="viewport" content="width=device-width, initial-scale=1.0"/> |
6 | | - <title>Camera + Visual Search</title> |
| 6 | + <title>Camera Visual Search (Cached Embeddings)</title> |
| 7 | + |
7 | 8 | <script src="https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@4.11.0/dist/tf.min.js"></script> |
8 | 9 | <script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/mobilenet@3.1.0/dist/mobilenet.min.js"></script> |
| 10 | + |
9 | 11 | <style> |
10 | 12 | body { font-family: Arial, sans-serif; text-align: center; padding: 1em; } |
11 | | - video, canvas { width: 100%; max-width: 400px; } |
12 | | - #result-img { margin-top: 1em; max-width: 400px; } |
| 13 | + video, canvas, img { width: 100%; max-width: 400px; } |
| 14 | + #status { margin: 1em 0; font-weight: bold; } |
13 | 15 | </style> |
14 | 16 | </head> |
15 | 17 | <body> |
16 | 18 |
|
17 | | - <h1>Visual Search from Camera</h1> |
| 19 | + <h1>Visual Search</h1> |
| 20 | + <p id="status">Loading model…</p> |
18 | 21 |
|
19 | 22 | <video id="video" autoplay playsinline></video> |
20 | 23 | <br /> |
21 | | - <button id="capture-btn">Take Photo</button> |
22 | | - |
23 | | - <canvas id="capture-canvas" style="display:none;"></canvas> |
24 | | - |
25 | | - <h2>Most Similar Match:</h2> |
26 | | - <img id="result-img" alt="Matching image will appear here"> |
27 | | - |
28 | | - <script> |
29 | | - let model; |
30 | | - const video = document.getElementById("video"); |
31 | | - const canvas = document.getElementById("capture-canvas"); |
32 | | - const ctx = canvas.getContext("2d"); |
33 | | - const resultImg = document.getElementById("result-img"); |
34 | | - |
35 | | - // 1️⃣ Start camera |
36 | | - async function startCamera() { |
37 | | - const stream = await navigator.mediaDevices.getUserMedia({ video: true }); |
38 | | - video.srcObject = stream; |
39 | | - } |
| 24 | + <button id="capture-btn" disabled>Take Photo</button> |
| 25 | + |
| 26 | + <canvas id="canvas" style="display:none;"></canvas> |
| 27 | + |
| 28 | + <h2>Best Match</h2> |
| 29 | + <img id="result-img" /> |
| 30 | + |
| 31 | +<script> |
| 32 | +let model; |
| 33 | +let collectionEmbeddings = []; // { url, embedding } |
| 34 | + |
| 35 | +const video = document.getElementById("video"); |
| 36 | +const canvas = document.getElementById("canvas"); |
| 37 | +const ctx = canvas.getContext("2d"); |
| 38 | +const statusEl = document.getElementById("status"); |
| 39 | +const resultImg = document.getElementById("result-img"); |
| 40 | +const captureBtn = document.getElementById("capture-btn"); |
| 41 | + |
| 42 | +/* ------------------ Camera ------------------ */ |
| 43 | +async function startCamera() { |
| 44 | + const stream = await navigator.mediaDevices.getUserMedia({ video: true }); |
| 45 | + video.srcObject = stream; |
| 46 | +} |
| 47 | + |
| 48 | +/* ------------------ Model ------------------ */ |
| 49 | +async function loadModel() { |
| 50 | + model = await mobilenet.load({ version: 2, alpha: 1.0 }); |
| 51 | + console.log("MobileNet loaded"); |
| 52 | +} |
| 53 | + |
| 54 | +/* ------------------ JSON ------------------ */ |
| 55 | +async function loadCollection() { |
| 56 | + const res = await fetch( |
| 57 | + "https://johnstack.github.io/JavaScript-Sandpit/ng_sandbox/collection.sample.json" |
| 58 | + ); |
| 59 | + const json = await res.json(); |
| 60 | + return json.items; // adjust if structure differs |
| 61 | +} |
| 62 | + |
| 63 | +/* ------------------ Embeddings ------------------ */ |
| 64 | +async function getEmbedding(img) { |
| 65 | + const tensor = tf.browser.fromPixels(img) |
| 66 | + .resizeNearestNeighbor([224, 224]) |
| 67 | + .toFloat() |
| 68 | + .expandDims(); |
| 69 | + |
| 70 | + const embedding = model.infer(tensor, "conv_preds"); |
| 71 | + const data = embedding.dataSync(); |
| 72 | + |
| 73 | + tensor.dispose(); |
| 74 | + embedding.dispose(); |
| 75 | + |
| 76 | + return data; |
| 77 | +} |
| 78 | + |
| 79 | +/* ------------------ Similarity ------------------ */ |
| 80 | +function cosineSimilarity(a, b) { |
| 81 | + let dot = 0, magA = 0, magB = 0; |
| 82 | + for (let i = 0; i < a.length; i++) { |
| 83 | + dot += a[i] * b[i]; |
| 84 | + magA += a[i] * a[i]; |
| 85 | + magB += b[i] * b[i]; |
| 86 | + } |
| 87 | + return dot / (Math.sqrt(magA) * Math.sqrt(magB)); |
| 88 | +} |
| 89 | + |
| 90 | +/* ------------------ Preload Embeddings ------------------ */ |
| 91 | +async function preloadEmbeddings() { |
| 92 | + statusEl.textContent = "Loading image collection…"; |
| 93 | + |
| 94 | + const items = await loadCollection(); |
| 95 | + |
| 96 | + for (let item of items) { |
| 97 | + const img = new Image(); |
| 98 | + img.crossOrigin = "anonymous"; |
| 99 | + img.src = item.url; // adjust key if needed |
| 100 | + |
| 101 | + await new Promise(resolve => img.onload = resolve); |
| 102 | + |
| 103 | + const embedding = await getEmbedding(img); |
| 104 | + collectionEmbeddings.push({ |
| 105 | + url: item.url, |
| 106 | + embedding |
| 107 | + }); |
40 | 108 |
|
41 | | - // 2️⃣ Load MobileNet |
42 | | - async function loadModel() { |
43 | | - model = await mobilenet.load(); |
44 | | - console.log("MobileNet loaded"); |
45 | | - } |
| 109 | + statusEl.textContent = `Indexed ${collectionEmbeddings.length}/${items.length}`; |
| 110 | + } |
46 | 111 |
|
47 | | - // 3️⃣ Fetch JSON collection |
48 | | - async function loadCollection() { |
49 | | - const res = await fetch( |
50 | | - "https://johnstack.github.io/JavaScript-Sandpit/ng_sandbox/collection.sample.json" |
51 | | - ); |
52 | | - const json = await res.json(); |
53 | | - return json.items; // adjust based on JSON structure |
54 | | - } |
| 112 | + statusEl.textContent = "Ready! Take a photo."; |
| 113 | + captureBtn.disabled = false; |
| 114 | +} |
55 | 115 |
|
56 | | - // 4️⃣ Compute embedding |
57 | | - async function getEmbedding(imgElement) { |
58 | | - const logits = model.infer(imgElement, "conv_preds"); |
59 | | - return logits.dataSync(); // embedding vector |
60 | | - } |
| 116 | +/* ------------------ Search ------------------ */ |
| 117 | +async function findBestMatch(capturedImg) { |
| 118 | + const captureEmbedding = await getEmbedding(capturedImg); |
61 | 119 |
|
62 | | - // 5️⃣ Compute cosine similarity |
63 | | - function cosineSim(a, b) { |
64 | | - let dot = 0, magA = 0, magB = 0; |
65 | | - for (let i = 0; i < a.length; i++) { |
66 | | - dot += a[i] * b[i]; |
67 | | - magA += a[i] * a[i]; |
68 | | - magB += b[i] * b[i]; |
69 | | - } |
70 | | - return dot / (Math.sqrt(magA) * Math.sqrt(magB)); |
71 | | - } |
| 120 | + let bestScore = -1; |
| 121 | + let bestMatch = null; |
72 | 122 |
|
73 | | - // 6️⃣ Find nearest |
74 | | - async function findNearest(captureImg, collection) { |
75 | | - const captureEmbedding = await getEmbedding(captureImg); |
76 | | - |
77 | | - let bestSim = -1; |
78 | | - let bestImg = null; |
79 | | - |
80 | | - for (let item of collection) { |
81 | | - // load collection image |
82 | | - const tempImg = new Image(); |
83 | | - tempImg.crossOrigin = "anonymous"; |
84 | | - tempImg.src = item.url; // adjust if JSON key differs |
85 | | - |
86 | | - await new Promise((resolve) => { |
87 | | - tempImg.onload = resolve; |
88 | | - tempImg.onerror = resolve; // skip broken links |
89 | | - }); |
90 | | - |
91 | | - const emb = await getEmbedding(tempImg); |
92 | | - const sim = cosineSim(captureEmbedding, emb); |
93 | | - |
94 | | - if (sim > bestSim) { |
95 | | - bestSim = sim; |
96 | | - bestImg = item.url; |
97 | | - } |
98 | | - } |
99 | | - return bestImg; |
| 123 | + for (let item of collectionEmbeddings) { |
| 124 | + const score = cosineSimilarity(captureEmbedding, item.embedding); |
| 125 | + if (score > bestScore) { |
| 126 | + bestScore = score; |
| 127 | + bestMatch = item.url; |
100 | 128 | } |
101 | | - |
102 | | - |
103 | | - // 🔘 Capture button logic |
104 | | - document.getElementById("capture-btn").addEventListener("click", async () => { |
105 | | - canvas.width = video.videoWidth; |
106 | | - canvas.height = video.videoHeight; |
107 | | - ctx.drawImage(video, 0, 0); |
108 | | - |
109 | | - const dataUrl = canvas.toDataURL("image/jpeg"); |
110 | | - const captureImg = new Image(); |
111 | | - captureImg.src = dataUrl; |
112 | | - |
113 | | - captureImg.onload = async () => { |
114 | | - resultImg.src = "Searching..."; |
115 | | - const collection = await loadCollection(); |
116 | | - const nearestUrl = await findNearest(captureImg, collection); |
117 | | - |
118 | | - if (nearestUrl) { |
119 | | - resultImg.src = nearestUrl; |
120 | | - } else { |
121 | | - resultImg.alt = "No similar image found"; |
122 | | - } |
123 | | - }; |
124 | | - }); |
125 | | - |
126 | | - |
127 | | - // Initialize |
128 | | - startCamera(); |
129 | | - loadModel(); |
130 | | - </script> |
| 129 | + } |
| 130 | + |
| 131 | + return bestMatch; |
| 132 | +} |
| 133 | + |
| 134 | +/* ------------------ Capture ------------------ */ |
| 135 | +captureBtn.addEventListener("click", async () => { |
| 136 | + canvas.width = video.videoWidth; |
| 137 | + canvas.height = video.videoHeight; |
| 138 | + ctx.drawImage(video, 0, 0); |
| 139 | + |
| 140 | + const img = new Image(); |
| 141 | + img.src = canvas.toDataURL("image/jpeg"); |
| 142 | + |
| 143 | + img.onload = async () => { |
| 144 | + statusEl.textContent = "Searching…"; |
| 145 | + const match = await findBestMatch(img); |
| 146 | + resultImg.src = match; |
| 147 | + statusEl.textContent = "Match found!"; |
| 148 | + }; |
| 149 | +}); |
| 150 | + |
| 151 | +/* ------------------ Init ------------------ */ |
| 152 | +(async function init() { |
| 153 | + await startCamera(); |
| 154 | + await loadModel(); |
| 155 | + await preloadEmbeddings(); |
| 156 | +})(); |
| 157 | +</script> |
131 | 158 |
|
132 | 159 | </body> |
133 | 160 | </html> |
0 commit comments