Improve OCR preprocessing and normalize merchant names

src/parser.js

@@ -78,7 +78,7 @@ export function parseStoreHeaderText(text) {
 
   return {
     side: match[1].toLowerCase(),
-    merchant: match[2].trim(),
+    merchant: match[2].replace(/\s+/g, ""),
     rawText: normalized,
   };
 }

src/vision.js

@@ -4,6 +4,16 @@ import { normalizeOcrText } from "./parser.js";
 let cvPromise;
 let workerPromise;
 let progressListener = () => {};
+const WHITE_TEXT_MIN_CHANNEL = 150;
+const WHITE_TEXT_MAX_SPREAD = 55;
+
+export function isWhiteTextPixel(red, green, blue, alpha = 255) {
+  return (
+    alpha >= 128 &&
+    Math.min(red, green, blue) >= WHITE_TEXT_MIN_CHANNEL &&
+    Math.max(red, green, blue) - Math.min(red, green, blue) <= WHITE_TEXT_MAX_SPREAD
+  );
+}
 
 export function onOcrProgress(listener) {
   progressListener = listener;
@@ -79,7 +89,6 @@ function prepareOcrCrop(sourceCanvas, rect) {
 
   const context = canvas.getContext("2d", { willReadFrequently: true });
   context.imageSmoothingEnabled = false;
-  context.filter = "grayscale(1) contrast(2)";
   context.drawImage(
     sourceCanvas,
     rect.x,
@@ -92,6 +101,23 @@ function prepareOcrCrop(sourceCanvas, rect) {
     canvas.height,
   );
 
+  const image = context.getImageData(0, 0, canvas.width, canvas.height);
+  for (let index = 0; index < image.data.length; index += 4) {
+    const value = isWhiteTextPixel(
+      image.data[index],
+      image.data[index + 1],
+      image.data[index + 2],
+      image.data[index + 3],
+    )
+      ? 0
+      : 255;
+    image.data[index] = value;
+    image.data[index + 1] = value;
+    image.data[index + 2] = value;
+    image.data[index + 3] = 255;
+  }
+  context.putImageData(image, 0, 0);
+
   return canvas;
 }
 

test/parser.test.js

@@ -12,6 +12,13 @@ import {
   rectFromPoints,
   resolveFieldRect,
 } from "../src/parser.js";
+import { isWhiteTextPixel } from "../src/vision.js";
+
+test("keeps white text pixels and removes colored or dark pixels", () => {
+  assert.equal(isWhiteTextPixel(220, 215, 195), true);
+  assert.equal(isWhiteTextPixel(210, 165, 80), false);
+  assert.equal(isWhiteTextPixel(120, 120, 120), false);
+});
 
 test("keeps only letters, digits and spaces in OCR item names", () => {
   assert.equal(cleanItemName(" ~~  Soulshot:   S-grade `  "), "Soulshot S grade");
@@ -139,5 +146,6 @@ test("parses trade side and merchant from the store header", () => {
     merchant: "Dwa",
     rawText: "Private Store({Sell) - Dwa",
   });
+  assert.equal(parseStoreHeaderText("Private Store(Buy) - Hik vision")?.merchant, "Hikvision");
   assert.equal(parseStoreHeaderText("Items on Sale"), null);
 });