A fine-tuned YOLOv5 deep learning approach for real-time house number detection

Area of Science:

• Computer Vision
• Deep Learning
• Object Detection

Background:

• Detecting small objects like house numbers in natural images is challenging due to blur and depth.
• Convolutional Neural Network (CNN) models are prevalent in object detection but struggle with small, variable-sized objects.

Purpose of the Study:

• To improve real-time house number detection from natural images.
• To evaluate the effectiveness of fine-tuning common CNN models for this task.

Main Methods:

• Applied classical CNN models: Faster R-CNN, MobileNet, YOLOv4, YOLOv5, and YOLOv7.
• Introduced a novel fine-tuning technique to enhance the performance of these CNN models.
• Conducted experiments on real-world data from Kayseri province.

Main Results:

• Fine-tuning improved the F1 scores across all tested CNN models.
• The fine-tuned YOLOv5 achieved the highest F1 score (0.972), outperforming other models.
• YOLOv7 remained the fastest model (0.009s), but fine-tuned YOLOv5 offered superior accuracy.

Conclusions:

• Fine-tuning is an effective strategy for enhancing CNN-based house number detection.
• The proposed fine-tuned YOLOv5 approach offers a strong balance of high accuracy and acceptable speed.
• This research contributes to more robust real-time object detection systems for natural scenes.