Executive Summary

The research required an improved License Plate Recognition system in Austria, Switzerland and the USA, which in turn was provided by Infosearch BPO’s annotation services. Challenges, solutions, and achievements of this case are described to reveal the crucial importance of high-quality data annotation in machine learning.

Background

As the requirements of practical transportation management and video surveillance for automobiles have risen, calls for integrating high-quality LPR systems have also followed. These systems depend on the data used for training machine learning models and consequently, data annotation is a major step.

Project Overview

Client: A leading technology company that designs and develops artificial intelligent based solutions for security.

Objective: To create an LPR system that will recognize and read license plates of different countries, in any positions and illumination.

Challenges

1.Diverse License Plate Formats: The client required to host many numbers plates style of many countries.

2.Varying Environmental Conditions: Unfavourable conditions such as lighting, weather and angle were observed during data capture and as such proper annotation was done.

3.High Volume of Data: The project concerned the annotation of tens of thousands of images, which needed to be done quickly without sacrificing regard to quality.

Annotation Services Provided

Infosearch BPO implemented a multi-faceted approach to address the client's needs:

1.Customized Annotation Guidelines: Established rather specific reference points that annotators had followed in order to be consistent labelling various license plate styles and formats.

2.Skilled Workforce: A pool of senior trained annotators with work experience in previous LPR projects was recruited to optimize the results.

3.Quality Control Mechanisms: Implemented a dual system of quality control, which was in three folds:

• Periodic checks of the densely annotated data.
• Feedback mechanisms to enable annotators increase accuracy.
• Differentiation of the annotation tools that allowed easy review and editing of the annotations.

4.Rapid Turnaround Times: Utilized efficient organization of work procedures and adopted specific project management tools aimed at meeting the time submits of the annotated datasets.

Implementation Process

1.Data Collection: The client offered a heterogeneous dataset including images of vehicles in different contexts.

2.Initial Training: Annotators also participated in training sessions where they were enlightened on details of LPR with particular emphasis on issues concerning accuracy.

3.Annotation Phase: The team started enhancing images – marking all the license numbers, then sorting by country, angle, or physical exposure.

4.Quality Assurance: Subsequently to the first run of annotations, a quality control team scrutinized the undertaken activity, made feedback and required corrections.

Outcomes

• Increased Model Accuracy: The high-quality annotated data made a great positive impact to the performance of the LPR models and the overall accuracy rate of the models reached more than 95%.
• Efficiency Gains: Fast data annotation cycle allowed clients to save time on their project and deliver their product much sooner than originally planned.
• Scalability: The processes that were implemented by Infosearch BPO also indicated flexibility – more so, the client can address other projects with similar needs in the future.

Conclusion

The partnering between Infosearch BPO and the client was an LPR project that highlights the role of accurate data annotation in machine learning. The respective measures were not only used to solve present issues but also to establish a basis for the following projects proving that Infosearch BPO’s annotation services are highly valuable.

Future Directions

For the future, Infosearch BPO has plans to increase the use of annotation automation like the AI-supported annotators in other future annotation activities. Continual improvement strategy is proposed here to address the emerging needs of clients within this machine learning segment.