Mixed YOLOv3-LITE: A Lightweight Real-Time Object Detection Method.

Area of Science:

• Computer Vision
• Deep Learning
• Embedded Systems

Background:

• Mobile and embedded devices have limited computational power and high energy consumption, hindering complex AI tasks.
• Real-time object detection is crucial for many applications but often requires significant computational resources, typically GPUs.

Purpose of the Study:

• To develop a lightweight, real-time object detection network optimized for non-GPU and mobile devices.
• To improve the balance between detection precision and processing speed for edge computing applications.

Main Methods:

• Introduced Mixed YOLOv3-LITE, a network based on YOLO-LITE, incorporating residual blocks (ResBlocks) and parallel high-to-low resolution subnetworks.
• Utilized shallow and narrow convolutional layers to enhance network depth while preserving shallow network characteristics.
• Optimized the network architecture for efficient performance on devices without Graphics Processing Units (GPUs).

Main Results:

• The Mixed YOLOv3-LITE model size is 20.5 MB, significantly smaller than YOLOv3 (91.70%), tiny-YOLOv3 (38.07%), and SlimYOLOv3-spp3-50 (74.25%).
• Achieved a mean average precision (mAP) of 48.25% on the PASCAL VOC 2007 dataset, a 14.48% improvement over YOLO-LITE.
• Obtained an mAP of 28.50% on the VisDrone 2018-Det dataset, outperforming tiny-YOLOv3 (18.50%) and SlimYOLOv3-spp3-50 (2.70%).

Conclusions:

• Mixed YOLOv3-LITE demonstrates superior efficiency and performance for object detection on mobile and non-GPU devices.
• The proposed architecture effectively balances detection accuracy and speed, addressing the limitations of current edge AI hardware.
• This lightweight network is suitable for real-time object detection tasks in resource-constrained environments.