Checkbox-Text Extraction: YOLO Checkbox Detector

The Document Understanding Subnet employs a custom-trained YOLO Checkbox Detector built on YOLOv8-large, optimized specifically for checkbox detection across a broad range of document types. This YOLOv8 model was trained on a unique dataset of over 10,000 document images, encompassing both scanned and standard formats. The model was rigorously tested on a challenging test set of 300 diverse images to benchmark its performance against leading AI solutions. The results were impressive, as shown below:

MODEL	F1 - SCORE
Azure Form Recognizer	0.72
GPT-4 Vision	0.63
YOLO Checkbox Detector	0.88

With an F1-Score of 0.88, the YOLO Checkbox Detector sets a new benchmark in checkbox detection, offering unmatched accuracy and reliability for document processing tasks.

YOLOv8 Architecture

YOLOv8 is a significant evolution within the YOLO (You Only Look Once) series of object detection models, introducing notable improvements in speed and accuracy. Its architectural enhancements make YOLOv8 a strong choice for real-time object detection in complex document environments.

• Enhanced Backbone and Feature Extraction: YOLOv8 includes a more sophisticated backbone network and a refined feature extraction process, allowing for better spatial and contextual understanding of the document image. This enhanced architecture results in more precise bounding box predictions for checkboxes and other elements, ensuring high accuracy.

• Optimized Performance and Efficiency: Designed for both high frame rates (often surpassing 60 FPS) and efficient operation on low-resource devices, YOLOv8 utilizes techniques like model pruning, quantization, and optimized convolution. These optimizations make it well-suited for real-time applications, enabling the subnet to deliver rapid checkbox detection without compromising on accuracy.

• Advanced Loss Function: YOLOv8 integrates a loss function that combines localization loss, confidence loss, and classification loss, each weighted by α, β, and γ, respectively. This setup balances the detection accuracy of the bounding boxes, improves classification confidence, and maintains alignment with ground truth coordinates, all of which are critical for high-stakes applications in document understanding.

Figure 6: Illustration of YOLOv8 architecture.

Additionally, YOLOv8 introduces enhanced Intersection over Union (IoU) metrics to further improve the precision of checkbox bounding boxes, ensuring that each detected checkbox is aligned accurately with its associated text.