Micro-YOLOv8: a lightweight real-time object detector

Le résumé fourni par la source

Significant advancements have been made in object detection performance through the use of deep learning models. However, traditional models like Faster R-CNN and YOLO challenges in deployment on embedded mobile devices due to their large network size, which is constrained by limited computational resources and stringent power limitations. To address this issue, we draw upon the concept of Lightweight Deep Convolutional Neural Networks (LDCNN)s[6] and introduce a novel lightweight CNN-based object detection model, Micro-YOLOv8, which is derived from YOLOv8n. This model achieves a substantial reduction in the number of parameters and computational cost while detection preserving performance. We propose to replace convolutional layers in the YOLOv8 network with Generalized Sparse Convolution (GSConvns) and VoV-GSCSP, and introduce the Layer-Adaptive Magnitude-based Pruning (LAMP) score to minimize the number of parameters(Param.) and maximize the detection performance. Hence, the proposed Micro-YOLOv8 achieves a 49.09% parameter reduction and 55.56% lower computational cost while maintaining detection accuracy on both the WiderPerson(88.40%, AP50) and VOC datasets (77.26%, AP50), compared to the original YOLOv8n network.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.