Efficient Invoice Interpretation: Practical and AI-Powered MicroService for Automated Data Extraction

Sammanfattning

This thesis details the development, implementation, and optimization of a local microservice designed to enhance invoice processing at Hulo, a company that manages au- tomatic payment invoices. This is conducted through Artificial Intelligence (AI) and Machine Learning (ML) techniques, focusing particularly on Computer Vision and Object- detection. Motivated by the need to address the inefficiencies of traditional manual processing which is error-prone, slow, and costly, this project employs a structured methodological approach. Initial data gathering, meticulous data labeling, and extensive training and validation of an AI model using tools like Roboflow [1] are key components. The chosen You Only Look Once (YOLO) model [2], known for its robust object detection capabilities, demonstrates significant improvements in recognizing and classifying invoice data through successive training phases. Challenges such as class imbalances are tackled through dataset enrichment and augmen- tation techniques, enhancing the model’s robustness and generalization capability across diverse invoice formats. The system effectively transforms invoices into structured JSON data, thus automating and streamlining business practices. The thesis encapsulates the project’s journey from conceptualization to deployment, highlighting strategic solutions for data quality enhancement and computational efficiency. The challenges and opportunities of the transition of this microservice into a production environment, where it undergoes rigorous real-world testing, ensures robustness, data security, and operational reliability. The successful implementation of this project underscores AI’s role in improving business processes and aligning with industry trends toward digital automation and intelligent systems.