tensorflow is not my friend

Dockerfile

@@ -1,4 +1,4 @@
-FROM tensorflow/tensorflow:latest
+FROM python:3.11.1-bullseye
 
 COPY . /app
 

requirements.txt

@@ -1,14 +1,12 @@
 wget
 flask
-keras
-numpy
+Pillow
 psutil
 yt_dlp
-aiohttp
 urllib3
 requests
 py-cpuinfo
-opencv-python
+pytesseract
 ffmpeg-python
 flask_limiter
 SpeechRecognition

routes/aminoOSRapi/captcha_processor.py

@@ -1,111 +0,0 @@
-import cv2
-from numpy import asarray as np_as_array
-from numpy import all as np_all
-
-
-class CaptchaProcessor:
-
-    WHITE_RGB = (255, 255, 255)
-
-    def __init__(self, data: bytes):
-        self.img = cv2.imdecode(
-            np_as_array(bytearray(data), dtype="uint8"),
-            cv2.IMREAD_ANYCOLOR
-        )
-
-    def threshold(self):
-        self.img = cv2.threshold(self.img, 0, 255, cv2.THRESH_OTSU)[1]
-
-    def convert_color_space(self, target_space: int):
-        self.img = cv2.cvtColor(self.img, target_space)
-
-    def get_background_color(self) -> tuple:
-        return tuple(self.img[0, 0])
-
-    def resize(self, x: int, y: int):
-        self.img = cv2.resize(self.img, (x, y))
-
-    def save(self, name: str):
-        cv2.imwrite(name, self.img)
-
-    def get_letters_color(self) -> tuple:
-        colors = []
-        for y in range(self.img.shape[1]):
-            for x in range(self.img.shape[0]):
-                color = tuple(self.img[x, y])
-                if color != self.WHITE_RGB: colors.append(color)
-        return max(set(colors), key=colors.count)
-
-    def replace_color(self, target: tuple, to: tuple):
-        self.img[np_all(self.img == target, axis=-1)] = to
-
-    def replace_colors(self, exception: tuple, to: tuple):
-        self.img[np_all(self.img != exception, axis=-1)] = to
-
-    def increase_contrast(self, alpha: float, beta: float):
-        self.img = cv2.convertScaleAbs(self.img, alpha=alpha, beta=beta)
-
-    def increase_letters_size(self, add_pixels: int):
-        pixels = []
-        for y in range(self.img.shape[1]):
-            for x in range(self.img.shape[0]):
-                if self.img[x, y] == 0: pixels.append((x, y))
-        for y, x in pixels:
-            for i in range(1, add_pixels + 1):
-                self.img[y + i, x] = 0
-                self.img[y - i, x] = 0
-                self.img[y, x + i] = 0
-                self.img[y, x - i] = 0
-                self.img[y + i, x] = 0
-                self.img[y - i, x] = 0
-                self.img[y, x + i] = 0
-                self.img[y, x - i] = 0
-
-    # Отдаление символов друг от друга
-    # Может многократно повысить точность, но я так и не придумал правильную реализацию
-    def distance_letters(self, cf: float):
-        pixels = []
-        for y in range(self.img.shape[1]):
-            for x in range(self.img.shape[0]):
-                if self.img[x, y] == 0: pixels.append((x, y))
-        for y, x in pixels:
-            self.img[y, x] = 255
-            center = self.img.shape[1] / 2
-            z = self.img.shape[1] / x
-            if z >= 2: self.img[y, x - int((900 // x) * cf)] = 0
-            else: self.img[y, x + int((900 // x) * cf)] = 0
-
-    def slice_letters(self):
-        contours, hierarchy = cv2.findContours(self.img, cv2.RETR_TREE, cv2.CHAIN_APPROX_NONE)
-        letter_image_regions = []
-        letters = []
-        for idx, contour in enumerate(contours):
-            if hierarchy[0][idx][3] != 0: continue
-            (x, y, w, h) = cv2.boundingRect(contour)
-            if w / h > 1.5:
-                half_width = int(w / 2)
-                letter_image_regions.append((idx, x, y, half_width, h))
-                letter_image_regions.append((idx, x + half_width, y, half_width, h))
-            else:
-                letter_image_regions.append((idx, x, y, w, h))
-        letter_image_regions = sorted(letter_image_regions, key=lambda z: z[1])
-        for _, x, y, w, h in letter_image_regions:
-            frame = self.img[y:y + h, x:x + w]
-            if frame.shape[1] > 35: continue
-            frame = cv2.resize(frame, (20, 40))
-            frame = cv2.cvtColor(frame,  cv2.COLOR_RGB2BGR)
-            letters.append(frame)
-        return letters
-
-    def show(self):
-        cv2.imshow("Captcha Processor", self.img)
-        cv2.waitKey(0)
-
-    @classmethod
-    def from_file_name(cls, name: str):
-        file = open(name, "rb")
-        processor = cls(file.read())
-        file.close()
-        return processor
-
-

routes/aminoOSRapi/main.py

@@ -1,12 +1,81 @@
-from .utils import *
-import asyncio
+import requests
+import pytesseract
+from PIL import Image, ImageEnhance
 
 def apipredict(request):
     try:
         if request.method == 'POST': url = request.form['url']
         else: url = request.args['url']
-        if url.strip() in ['', None]: raise Exception()
+        if url.strip() in ['', None]:
+            raise Exception()
     except: return {"status": "error", "details": { "error_code": 101, "error_details": "No link provided" }}
-    loop = asyncio.new_event_loop()
-    coroutine = predict(url)
-    return loop.run_until_complete(coroutine)
+
+    WALK_Y = 33
+
+    WHITE_RGB = (255, 255, 255)
+
+    def replaceColor(image: Image, forReplace: tuple, newColor: tuple):
+        for y in range(image.size[1]):
+            for x in range(image.size[0]):
+                if pix[x, y] == forReplace:
+                    image.putpixel((x, y), newColor)
+
+    image = Image.open(requests.get(url, stream=True).raw)
+
+    pix = image.load()
+
+    garbageColor = image.getpixel((0, 0))
+
+    replaceColor(image, garbageColor, WHITE_RGB)
+
+    colors = []
+
+    for i in range(249):
+        pixel = image.getpixel((i, WALK_Y))
+        if pixel != WHITE_RGB:
+            colors.append(pixel)
+
+    captchaLettersColor = max(set(colors), key = colors.count)
+
+    for y in range(image.size[1]):
+        for x in range(image.size[0]):
+            if pix[x, y] != captchaLettersColor:
+                image.putpixel((x, y), WHITE_RGB)
+
+    image = ImageEnhance.Contrast(image).enhance(500)
+
+    for y in range(image.size[1]):
+        for x in range(image.size[0]):
+            if pix[x, y] == captchaLettersColor:
+                p1 = pix[x + 1, y]
+                p2 = pix[x - 1, y]
+                p3 = pix[x, y + 1]
+                p4 = pix[x, y - 1]
+                p5 = pix[x + 2, y]
+                p6 = pix[x - 2, y]
+                p7 = pix[x, y + 2]
+                p8 = pix[x, y - 2]
+                if p1 != captchaLettersColor:
+                    image.putpixel((x + 1, y), captchaLettersColor)
+                if p2 != captchaLettersColor:
+                    image.putpixel((x - 1, y), captchaLettersColor)
+                if p3 != captchaLettersColor:
+                    image.putpixel((x, y + 1), captchaLettersColor)
+                if p4 != captchaLettersColor:
+                    image.putpixel((x, y - 1), captchaLettersColor)
+                if p5 != captchaLettersColor:
+                    image.putpixel((x + 2, y), captchaLettersColor)
+                if p6 != captchaLettersColor:
+                    image.putpixel((x - 2, y), captchaLettersColor)
+                if p7 != captchaLettersColor:
+                    image.putpixel((x, y + 2), captchaLettersColor)
+                if p8 != captchaLettersColor:
+                    image.putpixel((x, y - 2), captchaLettersColor)
+
+    image = ImageEnhance.Contrast(image).enhance(1200)
+
+    resized = image.resize((image.size[0] * 5, image.size[1] * 5))
+
+    decoded = pytesseract.image_to_string(resized, config = "--psm 13 --oem 3 -c tessedit_char_whitelist=23456789", lang = "eng")
+
+    return {"solution": decoded}

routes/aminoOSRapi/model.h5

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:792c015158ffcfaadbb2a65fef9623af7fa1d243e3e1f915444f86c40049ea13
-size 3730536

routes/aminoOSRapi/utils.py

@@ -1,43 +0,0 @@
-from keras.models import load_model
-from aiohttp import ClientSession
-from numpy import expand_dims as np_expand_dims
-from .captcha_processor import CaptchaProcessor
-from asyncio import get_running_loop
-
-model = load_model("/app/routes/aminoOSRapi/model.h5")
-
-
-async def get_binary_from_link(link: str) -> bytes:
-    async with ClientSession() as session:
-        return await (await session.get(link)).read()
-
-
-async def predict(url: str, recursion: int = 0) -> dict:
-    binary = await get_binary_from_link(url)
-    processor = CaptchaProcessor(binary)
-    processor.replace_color(processor.get_background_color(), processor.WHITE_RGB)
-    processor.replace_colors(processor.get_letters_color(), processor.WHITE_RGB)
-    processor.convert_color_space(6)
-    processor.threshold()
-    try:
-        processor.increase_letters_size(2)
-    except IndexError:
-        return await predict(url, recursion + 1)
-    letters = processor.slice_letters()
-    if len(letters) != 6: return await predict(url, recursion + 1)
-    shorts = []
-    final = ""
-    letters_solving = [
-        get_running_loop().run_in_executor(None, model.predict, np_expand_dims(letter, axis=0))
-        for letter in letters
-    ]
-    letters_solving = [await result for result in letters_solving]
-    fulls = [list(map(lambda x: float(x), letter[0])) for letter in letters_solving]
-    for prediction in fulls: shorts.append(prediction.index(max(*prediction)))
-    for short in shorts: final += str(short)
-    return {
-        "prediction": final,
-        "letters_predictions": shorts,
-        "full_prediction": fulls,
-        "recursion": recursion
-    }