A segmented model for automobile appearance detection based on improved YOLOv11-seg

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Conventional models struggle with vehicle appearance segmentation in complex scenes due to poor small-target detection and feature fusion.
• Limited recognition capabilities hinder accurate identification of vehicle components in cluttered environments.

Purpose of the Study:

• To develop an enhanced vehicle appearance segmentation model for improved detection in intricate settings.
• To address limitations in small-target recognition and multi-scale feature fusion in existing models.

Main Methods:

• Utilized the YOLOv11-seg framework with a novel MCALayerPlus module for concurrent multi-scale target processing.
• Implemented an improved ShapeIoU loss function incorporating size-sensitivity and category-aware shape penalty.
• Extracted multi-scale features to suppress false detections in cluttered backgrounds.

Main Results:

• Achieved state-of-the-art performance on a specialized automotive dataset.
• Obtained a mean Average Precision (mAP@0.5) of 94.09% and mAP@0.5:0.95 of 77.12%.
• Demonstrated high precision (91.31%) and recall (90.75%) with a lightweight model (5.75 MB) and fast inference (45.3 FPS).

Conclusions:

• The proposed model significantly enhances vehicle appearance segmentation accuracy and efficiency.
• The MCALayerPlus module and improved ShapeIoU loss contribute to superior performance in complex scenarios.
• The model's lightweight and high-speed characteristics make it ideal for real-time intelligent transportation systems.