You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
533 lines
15 KiB
533 lines
15 KiB
import * as tf from '@tensorflow/tfjs';
|
|
import { setWasmPaths } from '@tensorflow/tfjs-backend-wasm';
|
|
import modelJSON from './model.json';
|
|
import * as ccl from './ccl';
|
|
|
|
const charset = [' ', '0', '2', '4', '5', '8', 'A', 'D', 'G', 'H', 'J', 'K', 'M', 'N', 'P', 'R', 'S', 'T', 'V', 'W', 'X', 'Y'];
|
|
let weightsData: Uint8Array; // base64 encoded weights
|
|
let model: tf.LayersModel;
|
|
|
|
tf.enableProdMode();
|
|
|
|
const wasmToUrl = (wasm: any) => {
|
|
const blb = new Blob([wasm], { type: 'application/wasm' });
|
|
return URL.createObjectURL(blb);
|
|
};
|
|
|
|
const backendloaded = (async () => {
|
|
try {
|
|
// dead code elimination should occur here
|
|
// eslint-disable-next-line camelcase
|
|
if (execution_mode === 'userscript' || execution_mode === 'test') {
|
|
weightsData = (await import('./group1-shard1of1.bin')).default;
|
|
const tfwasmthreadedsimd = (await import('./tfjs-backend-wasm-threaded-simd.wasm')).default;
|
|
const tfwasmsimd = (await import('./tfjs-backend-wasm-simd.wasm')).default;
|
|
const tfwasm = (await import('./tfjs-backend-wasm.wasm')).default;
|
|
setWasmPaths({
|
|
'tfjs-backend-wasm.wasm': wasmToUrl(tfwasm),
|
|
'tfjs-backend-wasm-simd.wasm': wasmToUrl(tfwasmsimd),
|
|
'tfjs-backend-wasm-threaded-simd.wasm': wasmToUrl(tfwasmthreadedsimd)
|
|
});
|
|
} else {
|
|
weightsData = new Uint8Array(await (await fetch(chrome.runtime.getURL('./group1-shard1of1.bin'))).arrayBuffer());
|
|
const args = {
|
|
'tfjs-backend-wasm.wasm': chrome.runtime.getURL('tfjs-backend-wasm.wasm'),
|
|
'tfjs-backend-wasm-simd.wasm': chrome.runtime.getURL('tfjs-backend-wasm-simd.wasm'),
|
|
'tfjs-backend-wasm-threaded-simd.wasm': chrome.runtime.getURL('tfjs-backend-wasm-threaded-simd.wasm')
|
|
};
|
|
setWasmPaths(args);
|
|
}
|
|
const l = await tf.setBackend('wasm');
|
|
console.log('tf backend loaded', l);
|
|
} catch (err) {
|
|
console.log('tf err', err);
|
|
}
|
|
})();
|
|
|
|
function toggle(obj: HTMLElement, v?: any) {
|
|
if (v) obj.style.display = '';
|
|
else obj.style.display = 'none';
|
|
}
|
|
|
|
const iohander: tf.io.IOHandler = {
|
|
load: function () {
|
|
return new Promise((resolve, reject) => {
|
|
resolve({
|
|
modelTopology: modelJSON.modelTopology,
|
|
weightSpecs: modelJSON.weightsManifest[0].weights as tf.io.WeightsManifestEntry[],
|
|
weightData: weightsData.buffer,
|
|
format: modelJSON.format,
|
|
generatedBy: modelJSON.generatedBy,
|
|
convertedBy: modelJSON.convertedBy
|
|
});
|
|
});
|
|
}
|
|
};
|
|
|
|
async function load() {
|
|
model = await tf.loadLayersModel(iohander);
|
|
return model;
|
|
}
|
|
|
|
function black(x: number) {
|
|
return x < 64;
|
|
}
|
|
|
|
// Calculates "disorder" of the image. "Disorder" is the percentage of black pixels that have a
|
|
// non-black pixel below them. Minimizing this seems to be good enough metric for solving the slider.
|
|
function calculateDisorder(imgdata: ImageData) {
|
|
const a = imgdata.data;
|
|
const w = imgdata.width;
|
|
const h = imgdata.height;
|
|
const pic: number[] = [];
|
|
const visited: number[] = [];
|
|
|
|
for (let c = 0; c < w * h; c++) {
|
|
if (visited[c]) continue;
|
|
if (!black(a[c * 4])) continue;
|
|
|
|
let blackCount = 0;
|
|
const items: number[] = [];
|
|
const toVisit: number[] = [c];
|
|
while (toVisit.length > 0) {
|
|
const cc = toVisit[toVisit.length - 1];
|
|
toVisit.splice(toVisit.length - 1, 1);
|
|
|
|
if (visited[cc]) continue;
|
|
visited[cc] = 1;
|
|
|
|
if (black(a[cc * 4])) {
|
|
items.push(cc);
|
|
|
|
blackCount++;
|
|
toVisit.push(cc + 1);
|
|
toVisit.push(cc - 1);
|
|
toVisit.push(cc + w);
|
|
toVisit.push(cc - w);
|
|
}
|
|
}
|
|
|
|
if (blackCount >= 24) {
|
|
items.forEach(function (x) {
|
|
pic[x] = 1;
|
|
});
|
|
}
|
|
}
|
|
|
|
let res = 0;
|
|
let total = 0;
|
|
for (let c = 0; c < w * h - w; c++) {
|
|
if (pic[c] !== pic[c + w]) res += 1;
|
|
if (pic[c]) total += 1;
|
|
}
|
|
|
|
return res / (total === 0 ? 1 : total);
|
|
}
|
|
|
|
// returns ImageData from captcha's background image, foreground image, and offset (ranging from 0 to -50)
|
|
function imageFromCanvas(img: HTMLImageElement, bg: HTMLImageElement, off: number) {
|
|
const h = img.height;
|
|
const w = img.width;
|
|
const th = 80;
|
|
const ph = 0;
|
|
const pw = 16;
|
|
const scale = th / h;
|
|
|
|
const canvas = document.createElement('canvas');
|
|
canvas.width = w * scale + pw * 2;
|
|
canvas.height = th;
|
|
|
|
const ctx = canvas.getContext('2d', { willReadFrequently: true })!;
|
|
|
|
ctx.fillStyle = 'rgb(238,238,238)';
|
|
ctx.fillRect(0, 0, canvas.width, canvas.height);
|
|
|
|
ctx.translate(canvas.width / 2, canvas.height / 2);
|
|
|
|
const draw = function (off: number) {
|
|
if (bg) {
|
|
const border = 4;
|
|
ctx.drawImage(
|
|
bg,
|
|
-off + border,
|
|
0,
|
|
w - border * 2,
|
|
h,
|
|
-w / 2 + border,
|
|
-h / 2,
|
|
w - border * 2,
|
|
h
|
|
);
|
|
}
|
|
ctx.drawImage(img, -w / 2, -h / 2, w, h);
|
|
};
|
|
|
|
// if off is not specified and background image is present, try to figure out
|
|
// the best offset automatically; select the offset that has smallest value of
|
|
// calculateDisorder for the resulting image
|
|
if (bg && off == null) {
|
|
let bestDisorder = 999;
|
|
let bestImagedata: ImageData | null = null;
|
|
let bestOff = -1;
|
|
|
|
for (let off = 0; off >= -50; off--) {
|
|
draw(off);
|
|
|
|
let imgdata = ctx.getImageData(0, 0, canvas.width, canvas.height);
|
|
const disorder = calculateDisorder(imgdata);
|
|
|
|
if (disorder < bestDisorder) {
|
|
bestDisorder = disorder;
|
|
draw(off);
|
|
imgdata = ctx.getImageData(0, 0, canvas.width, canvas.height);
|
|
bestImagedata = imgdata;
|
|
bestOff = off;
|
|
}
|
|
}
|
|
|
|
// not the best idea to do this here
|
|
setTimeout(function () {
|
|
const bg = document.getElementById('t-bg');
|
|
const slider = document.getElementById('t-slider') as HTMLInputElement;
|
|
if (!bg || !slider) return;
|
|
|
|
slider.value = '' + (-bestOff * 2);
|
|
bg.style.backgroundPositionX = bestOff + 'px';
|
|
}, 1);
|
|
draw(bestOff);
|
|
return bestImagedata;
|
|
} else {
|
|
draw(off);
|
|
return ctx.getImageData(0, 0, canvas.width, canvas.height);
|
|
}
|
|
}
|
|
|
|
// for debugging purposes
|
|
function imagedataToImage(imagedata: ImageData) {
|
|
const canvas = document.createElement('canvas');
|
|
const ctx = canvas.getContext('2d')!;
|
|
canvas.width = imagedata.width;
|
|
canvas.height = imagedata.height;
|
|
ctx.putImageData(imagedata, 0, 0);
|
|
|
|
const image = new Image();
|
|
image.src = canvas.toDataURL();
|
|
return image;
|
|
}
|
|
|
|
function toMonochrome(px: Uint8ClampedArray) {
|
|
const ret = Array<number>(px.length >> 2);
|
|
for (let i = 0; i < px.length; i += 4) {
|
|
ret[i >> 2] = +(px[i] < 128);
|
|
}
|
|
return ret;
|
|
}
|
|
|
|
const greedyCTCDecode = (yPred: tf.Tensor<tf.Rank>) => tf.tidy(() => yPred.argMax(-1).arraySync());
|
|
|
|
function imgDisp(pixConv: (f: ArrayLike<number>, w: number, h: number, s: Uint8ClampedArray) => ArrayLike<number>, img: ArrayLike<number>, w: number, h: number, t?: boolean) {
|
|
const dt = new ImageData(w, h);
|
|
const rgba = pixConv(img, w, h, dt.data);
|
|
const imgres = imagedataToImage(dt);
|
|
document.body.appendChild(imgres);
|
|
if (t) {
|
|
imgres.style.transform = 'rotate(90deg) scaleY(-1)';
|
|
}
|
|
}
|
|
|
|
let colors = [
|
|
[255, 0, 0], // Red
|
|
[0, 255, 0], // Green
|
|
[0, 0, 255], // Blue
|
|
[255, 255, 0], // Yellow
|
|
[255, 0, 255], // Magenta
|
|
[0, 255, 255], // Cyan
|
|
[128, 0, 0], // Dark Red
|
|
[0, 128, 0], // Dark Green
|
|
[0, 0, 128], // Dark Blue
|
|
[128, 128, 0], // Olive
|
|
[128, 0, 128], // Purple
|
|
[0, 128, 128], // Teal
|
|
[192, 192, 192], // Silver
|
|
[128, 128, 128], // Gray
|
|
[255, 165, 0], // Orange
|
|
[0, 128, 64] // Medium Sea Green
|
|
];
|
|
|
|
const monoToPalette = (p: number[][], max: number) =>
|
|
function (arr: ArrayLike<number>, w: number, h: number, res: Uint8ClampedArray) {
|
|
let choice = p.slice(0);
|
|
const choices = new Map<number, number[]>();
|
|
for (let i = 0; i < arr.length; ++i) {
|
|
let col: number[];
|
|
if (choices.has(arr[i])) {
|
|
col = choices.get(arr[i])!;
|
|
} else {
|
|
col = choice.shift()!;
|
|
choices.set(arr[i], col);
|
|
if (choice.length == 0)
|
|
choice = p.slice(0);
|
|
}
|
|
[res[i * 4], res[i * 4 + 1], res[i * 4 + 2]] = col;
|
|
res[i * 4 + 3] = 255;
|
|
}
|
|
return res;
|
|
};
|
|
|
|
|
|
function monoToRgba(arr: ArrayLike<number>, w: number, h: number, res: Uint8ClampedArray) {
|
|
for (let i = 0; i < arr.length; ++i) {
|
|
res[i * 4] = res[i * 4 + 1] = res[i * 4 + 2] = arr[i] * 255;
|
|
res[i * 4 + 3] = 255;
|
|
}
|
|
return res;
|
|
}
|
|
|
|
function processCTCDecodedSequence(decodedSequence: number[], blankLabel = 0) {
|
|
const result = [];
|
|
let prevLabel = blankLabel;
|
|
|
|
for (const label of decodedSequence) {
|
|
if (label !== blankLabel && label !== prevLabel) {
|
|
result.push(label);
|
|
}
|
|
prevLabel = label;
|
|
}
|
|
|
|
return result;
|
|
}
|
|
|
|
function indicesToSymbols(decodedIndices: number[]) {
|
|
return decodedIndices.map(index => charset[index - 1] || '');
|
|
}
|
|
|
|
async function predict(img: HTMLImageElement, bg: HTMLImageElement, off: number) {
|
|
if (!model) {
|
|
model = await load();
|
|
}
|
|
const image = imageFromCanvas(img, bg, off);
|
|
if (!image)
|
|
throw new Error("Failed to gen image");
|
|
const mono = toMonochrome(image.data);
|
|
console.log(mono.reduce((a, b) => a + b), 0);
|
|
const labels = ccl.connectedComponentLabeling(mono, image.width, image.height);
|
|
const props = ccl.computeBounds(labels, image.width, image.height);
|
|
|
|
const sortedByArea = Object.entries(props).sort((a, b) => a[1].area - b[1].area);
|
|
const n = 8;
|
|
let eightBiggest = sortedByArea.slice(0, -n);
|
|
//const filtered = new Float32Array(80 * 300);
|
|
|
|
// TODO: maybe centering?
|
|
//imgDisp(monoToPalette(colors, Math.max(...new Set(labels))), labels, image.width, image.height);
|
|
|
|
for (const [label, region] of eightBiggest) {
|
|
//if ((region.maxRow - region.minRow) <= 20) {
|
|
// continue;
|
|
//}
|
|
|
|
for (let y = region.minRow; y <= region.maxRow; ++y) {
|
|
for (let x = region.minCol; x <= region.maxCol; ++x) {
|
|
if (labels[y * image!.width + x] === +label) {
|
|
labels[y * image!.width + x] = 0;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
eightBiggest = sortedByArea.slice(-n);
|
|
//imgDisp(monoToPalette(colors, Math.max(...new Set(labels))), labels, image.width, image.height);
|
|
|
|
for (const [label, region] of eightBiggest) {
|
|
if ((region.maxRow - region.minRow) > 20) {
|
|
continue;
|
|
}
|
|
|
|
for (let y = region.minRow; y <= region.maxRow; ++y) {
|
|
for (let x = region.minCol; x <= region.maxCol; ++x) {
|
|
if (labels[y * image!.width + x] === +label) {
|
|
labels[y * image!.width + x] = 0;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
//imgDisp(monoToPalette(colors, Math.max(...new Set(labels))), labels, image.width, image.height);
|
|
|
|
for (const [label, region] of eightBiggest) {
|
|
if ((region.maxRow - region.minRow) <= 20) {
|
|
continue;
|
|
}
|
|
|
|
for (let y = region.minRow; y <= region.maxRow; ++y) {
|
|
for (let x = region.minCol; x <= region.maxCol; ++x) {
|
|
if (labels[y * image!.width + x] === +label) {
|
|
labels[y * image!.width + x] = 1;
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
const filtered2 = tf.tensor3d(labels, [image.height, image.width, 1]).concat(tf.zeros([80, 300 - image.width, 1]), 1);
|
|
//imgDisp(monoToPalette(colors, Math.max(...new Set(labels))), labels, image.width, image.height);
|
|
//const tensor = tf.tensor3d(filtered, [80, 300, 1], 'float32');
|
|
//const tr = [1, 0, 2];
|
|
//console.log(tensor.shape, tensor.transpose(tr).shape);
|
|
const prediction = model.predict(filtered2.transpose([1, 0, 2]).expandDims(0));
|
|
|
|
let d: tf.TypedArray;
|
|
|
|
if (!Array.isArray(prediction)) {
|
|
const v = greedyCTCDecode(prediction) as number[][];
|
|
console.log(v);
|
|
const s = processCTCDecodedSequence(v[0], charset.length + 1);
|
|
return indicesToSymbols(s).join('').trim();
|
|
} else
|
|
throw new Error("unexpected inference");
|
|
|
|
// createSequence(d);
|
|
return '';
|
|
}
|
|
|
|
function createSequence(prediction: any) {
|
|
const csl = charset.length;
|
|
const sequence: Record<string, number>[] = [];
|
|
|
|
// for each prediction
|
|
for (let pos = 0; pos < prediction.length; pos += csl) {
|
|
// look at the probabilities for the 22 token characters
|
|
const preds = prediction.slice(pos, pos + csl);
|
|
const max = Math.max(...preds);
|
|
|
|
const seqElem: Record<string, number> = {};
|
|
|
|
for (let i = 0; i < csl; i++) {
|
|
const p = preds[i] / max; // normalize probability
|
|
const c = charset[i + 1];
|
|
|
|
if (p >= 0.05) { // if it's probable enough
|
|
seqElem[c || ''] = p; // save its probability, to give alternative solutions
|
|
}
|
|
}
|
|
|
|
sequence.push(seqElem);
|
|
}
|
|
|
|
return sequence;
|
|
}
|
|
|
|
|
|
async function imageFromUri(uri: string) {
|
|
if (uri.startsWith('url("')) {
|
|
uri = uri.substr(5, uri.length - 7);
|
|
}
|
|
// eslint-disable-next-line camelcase
|
|
if (execution_mode !== 'test' && !uri.startsWith('data:')) {
|
|
return null;
|
|
}
|
|
|
|
const img = new Image();
|
|
await new Promise((r) => {
|
|
img.onload = r;
|
|
img.src = uri;
|
|
});
|
|
return img;
|
|
}
|
|
|
|
async function predictUri(uri: string, uribg: string, bgoff: string | null) {
|
|
const img = await imageFromUri(uri);
|
|
const bg = uribg ? await imageFromUri(uribg) : null;
|
|
const off = bgoff ? parseInt(bgoff) : null;
|
|
|
|
return await predict(img!, bg!, off!);
|
|
}
|
|
|
|
const solveButton = document.createElement('input');
|
|
solveButton.id = 't-auto-solve';
|
|
solveButton.value = 'Solve';
|
|
solveButton.type = 'button';
|
|
solveButton.style.fontSize = '11px';
|
|
solveButton.style.padding = '0 2px';
|
|
solveButton.style.margin = '0px 0px 0px 6px';
|
|
solveButton.style.height = '18px';
|
|
solveButton.onclick = async function () {
|
|
solve(true);
|
|
};
|
|
|
|
const altsDiv = document.createElement('div');
|
|
altsDiv.id = 't-auto-options';
|
|
altsDiv.style.margin = '0';
|
|
altsDiv.style.padding = '0';
|
|
|
|
let storedPalceholder: string;
|
|
|
|
let overrides = {};
|
|
|
|
function placeAfter(elem: HTMLElement, sibling: HTMLElement) {
|
|
if (elem.parentElement !== sibling.parentElement) {
|
|
setTimeout(function () {
|
|
sibling.parentElement?.insertBefore(elem, sibling.nextElementSibling);
|
|
}, 1);
|
|
}
|
|
}
|
|
|
|
let previousText: string | null = null;
|
|
async function solve(force?: any) {
|
|
const resp = document.getElementById('t-resp') as HTMLInputElement;
|
|
if (!resp) return;
|
|
|
|
const bg = document.getElementById('t-bg');
|
|
if (!bg) return;
|
|
|
|
const fg = document.getElementById('t-fg');
|
|
if (!fg) return;
|
|
|
|
const help = document.getElementById('t-help');
|
|
if (!help) return;
|
|
|
|
await backendloaded;
|
|
|
|
placeAfter(solveButton, resp);
|
|
placeAfter(altsDiv, help);
|
|
|
|
// palememe
|
|
setTimeout(function () {
|
|
toggle(solveButton, bg.style.backgroundImage);
|
|
}, 1);
|
|
|
|
const text = fg.style.backgroundImage;
|
|
if (!text) {
|
|
altsDiv.innerHTML = '';
|
|
return;
|
|
}
|
|
|
|
if (text === previousText && !force) return;
|
|
previousText = text;
|
|
|
|
altsDiv.innerHTML = '';
|
|
if (!storedPalceholder)
|
|
storedPalceholder = resp.placeholder;
|
|
resp.placeholder = 'solving captcha...';
|
|
|
|
overrides = {};
|
|
|
|
const result = await predictUri(
|
|
text,
|
|
bg.style.backgroundImage,
|
|
force ? bg.style.backgroundPositionX : null
|
|
);
|
|
resp.placeholder = storedPalceholder;
|
|
|
|
resp.value = result;
|
|
}
|
|
|
|
const observer = new MutationObserver(async function (mutationsList, observer) {
|
|
solve(false);
|
|
});
|
|
|
|
//window['solve'] = solve;
|
|
|
|
observer.observe(document.body, {
|
|
attributes: true,
|
|
childList: true,
|
|
subtree: true
|
|
});
|
|
|