May

Automated solutions for the detection of road damage and the efficiency of internal processes

Project overview

May, a technology-driven company that specializes in road maintenance, wanted to create an advanced Image classification agents to detect road damage and provide product recommendations for quick fixes, minimizing the need for on-site experts. The aim was to reduce repair times and overall costs. In addition, May wanted to improve internal workflows through automated tools for software development, project management and tender creation.

This case study looks at the development of the road damage detection system using machine learning (ML) and how May has optimized internal processes with automated tools and significantly increased efficiency.

Image classification agent for detecting road damage

Target:

May needed a robust image classification system capable of recognizing different types of road damage (potholes, cracks, etc.) and providing workers with recommendations for repair products. This system had to be fast, accurate and usable by non-experts.

Challenges:

Precise detection of road damageThe system had to be able to distinguish between different types of damage.
Product recommendations: Automated suggestions for suitable repair materials had to be made.
User-friendliness for road workersWorkers needed an uncomplicated user interface that provided suggestions in real time.

Solution:

The solution was to develop an image classification agent using a machine learning model, in particular a YOLO (You Only Look Once)-architecture, which was trained on a balanced and extended data set for detecting road damage. The system integrated product recommendations for fast repairs.

development workflow:

Load and pre-process data record: Images of various road damages were loaded from organized subdirectories based on class designations. With the ImageDataGenerator of TensorFlow, the images have been converted for consistency to 640 x 640 Pixel reduced.
Balancing the data set with SMOTE: To compensate for class imbalances (i.e. fewer examples of rare damage types), the Synthetic Minority Over-sampling Technique (SMOTE) was used to generate synthesized images for underrepresented classes. These synthesized images were then reapplied to 640 x 640 to ensure a uniform size.
Data splitting: The data set was created in 70 % Training, 20 % Validation and 10 % test kits to ensure a balanced distribution across each class for training, validation and testing.
Data Augmentation: : The training set was created using the imgaug-library underwent extensive data enrichment. This included transformations such as rotating, flipping, scaling, cropping, adding noise and adjusting contrast and brightness. These enrichments helped to improve model generalization by diversifying the data set.
YOLO model training: The extended data set was used to create a YOLOv8 model for image classification. The fine-tuning was carried out on a pre-trained YOLO model for 100 epochs with a stack size of 32. Dropout was applied to prevent overfitting and the performance of the model was monitored on the validation set.
Model evaluation and exportAfter training, the performance of the model was assessed using the validation and test sets. The final model, once deemed satisfactory, was made available for deployment in the ONNX format exported.

Key technologies:

YOLOv8: Used for efficient image classification and transfer learning.
SMOTEIs used to equalize the data set and generate synthetic data for minority classes.
imgaugImproved training data diversity through augmentation.
TensorFlow/KerasIs used for image preprocessing and model management.

Result:

Improved accuracy: The final model achieved a 95 percent accuracy in the detection and classification of road damage.
Faster repairs: Road workers received real-time product suggestions so they could make immediate repairs without expert guidance.
Cost efficiency: : The reduced need for experts on site led to a 30 percent cost reduction during maintenance work.

Conclusion

By implementing advanced automated solutions for both external customers and internal processes, May has significantly improved its operational efficiency. The YOLO-based image classification system for detecting road damage enabled quick solutions without the need for expert supervision and reduced costs by 30 %. Internally, automation tools optimized software development, project management and proposal preparation, resulting in faster development cycles and better project outcomes.

Main advantages: