This app uses OCR to remove text from an image. You might use this as a stand-alone microservice, or as a pre-processing step in a computer vision pipeline. This tutorial is an adaptation of this post .
Define the environment You’ll start by creating a Beam app definition. In this file, you’re defining a few things:
The libraries you want installed in the environment
The compute settings (some of the CV operations are heavy, so 16Gi of memory is a safe choice)
from beam import App, Runtime, Image, Output app = App( name = "rmtext" , runtime = Runtime( cpu = 1 , memory = "16Gi" , image = Image( python_packages = [ "numpy" , "matplotlib" , "opencv-python" , "keras_ocr" , "tensorflow" , ], commands = [ "apt-get update && apt-get install -y libgl1" ], ), ), ) Removing the text from an image You’ll use the code below to accomplish the following:
Identify text in the base-64 encoded image and create bounding boxes around each block of text
Add a mask around each box of text
Paint over each text-mask to remove the text
We’ve added an app.run() decorator to remove_text. This decorator will allow us to run this code on Beam, instead of your laptop.
from beam import App, Runtime, Image, Output import base64 import matplotlib.pyplot as plt import keras_ocr import cv2 import math import numpy as np app = App( name = "rmtext" , runtime = Runtime( cpu = 1 , memory = "16Gi" , image = Image( python_packages = [ "numpy" , "matplotlib" , "opencv-python" , "keras_ocr" , "tensorflow" , ], ), ), ) def midpoint ( x1 , y1 , x2 , y2 ): x_mid = int ((x1 + x2) / 2 ) y_mid = int ((y1 + y2) / 2 ) return (x_mid, y_mid) @app.run () def remove_text ( ** inputs ): # Grab the base64 from the kwargs encoded_image = inputs[ "image" ] # Convert the base64-encoded input image to a buffer image_buffer = base64.b64decode(encoded_image) pipeline = keras_ocr.pipeline.Pipeline() # Read the image img = keras_ocr.tools.read(image_buffer) # Generate (word, box) tuples prediction_groups = pipeline.recognize([img]) mask = np.zeros(img.shape[: 2 ], dtype = "uint8" ) for box in prediction_groups[ 0 ]: x0, y0 = box[ 1 ][ 0 ] x1, y1 = box[ 1 ][ 1 ] x2, y2 = box[ 1 ][ 2 ] x3, y3 = box[ 1 ][ 3 ] x_mid0, y_mid0 = midpoint(x1, y1, x2, y2) x_mid1, y_mi1 = midpoint(x0, y0, x3, y3) thickness = int (math.sqrt((x2 - x1) ** 2 + (y2 - y1) ** 2 )) cv2.line(mask, (x_mid0, y_mid0), (x_mid1, y_mi1), 255 , thickness) img = cv2.inpaint(img, mask, 7 , cv2. INPAINT_NS ) img_rgb = cv2.cvtColor(img, cv2. COLOR_BGR2RGB ) # Save the generated image to the Beam Output path cv2.imwrite( "output.png" , img_rgb) if __name__ == "__main__" : input_image = "./coffee.jpeg" with open (input_image, "rb" ) as image_file: encoded_image = base64.b64encode(image_file.read()) remove_text( image = encoded_image) You can run this code on Beam by running beam run:
beam run app.py:remove_text Make sure to include a sample image in your working directory, and update the script with the path. In this example, I’m using this image as a sample:
Deployment If you’re satisfied with this function and want to deploy it as an API, you can do so by updating the decorator:
Just replace @app.run() with @app.task_queue()
# This function will be exposed as a web API when deployed! @app.task_queue () def remove_text ( ** inputs ): ... You can deploy this app by running:
You’ll call the API by copying the task queue URL from the dashboard.
Since this task runs asynchronously, you’ll use the /v1/task/{task_id}/status/ API to retrieve the task status and a link to download the image output.
This will return a response, which contains:
Task ID
The start and end time
A dictionary with pre-signed URLs to download the outputs
{ "task_id" : "edbcf7ff-e8ce-4199-8661-8e15ed880481" , "started_at" : "2023-04-24T22:44:06.911920Z" , "ended_at" : "2023-04-24T22:44:07.184763Z" , "outputs" : { "my-output-1" : { "path" : "output_path" , "name" : "my-output-1" , "url" : "http://data.beam.cloud/outputs/6446df99cf455a04e0335d9b/hw6hx/hw6hx-0001/edbcf7ff-e8ce-4199-8661-8e15ed880481/my-output-1.zip?..." } }, "status" : "COMPLETE" , } Enter the outputs url in your browser to download the image. You’ll see that the text has been removed:
