This repository contains supporting tools and machine learning code for Cropilot, an AI-powered application for automatically cropping scanned documents.
Cropilot utils provides the models and scripts used to detect pages in scanned books, newspapers, periodicals, and other printed media. These components support the main Cropilot application by predicting page locations, rotation angles, and crop instructions that can later be reviewed in the Cropilot editor or applied automatically.
The repository contains the training code for the computer vision models used by Cropilot. If you want to train a model from scratch, see the model training tutorial:
Train a Cropilot model from scratch
Cropilot currently uses two models to generate page predictions.
Cropilot uses a fine-tuned YOLO model based on YOLO11s to detect the number and position of pages in each scan.
RotateNET is a ResNet-based model that predicts the rotation angle of each detected page so the final crop can be properly aligned.
The cropilot_api_tools directory contains utility scripts for working with the Cropilot API. These scripts are intended for larger batch workflows that go beyond the simple upload flow available in the web UI.
Cropilot Tools can be used to:
- Upload scan batches to the Cropilot editor.
- Download Cropilot-generated crop instructions.
- Apply crop instructions to original image files.
- Fine-tune custom models for your own document datasets.
See the Cropilot Tools documentation for setup and usage instructions: