Loading…

Study of YOLOv7 Model Lightweighting Based on Group-Level Pruning

With the improvement of urbanization and living standards, waste disposal has become an urgent environmental problem. In recent years, object detection networks based on deep learning have greatly improved classification accuracy and speed by learning feature expression and object detection through...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2024, Vol.12, p.96138-96149
Main Authors: Bai, Xuemei, Shi, Hongjin, Wang, Zhijun, Hu, Hanping, Zhang, Chenjie
Format: Article
Language:English
Subjects:
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:With the improvement of urbanization and living standards, waste disposal has become an urgent environmental problem. In recent years, object detection networks based on deep learning have greatly improved classification accuracy and speed by learning feature expression and object detection through deep neural networks. However, with the increase of model depth and number of parameters, object detection models based on deep neural networks require a lot of computing and storage resources, and are difficult to deploy on edge devices. To solve this problem, this paper proposes a lightweight garbage classification object detection algorithm based on improved YOLOv7. Firstly, the DW-SPPFCSPC module is constructed according to the depth-separable volume product, replacing the original SPPCSPC module. Second, the ECA channel attention mechanism was added to the backbone of YOLOv7. By paying more attention to important feature channels, the model can better capture important information in images and thus improve the performance of target detection. Finally, the network channel is pruned by analyzing the network dependency graph. Experimental results show that the model size is greatly reduced by this method. Compared with the original YOLOv7 model, the number of parameters is reduced to 6.05% and floating-point operations (FLOPs) are reduced to 1/16 with a loss of only 0.56% in average accuracy. The method is tested on VOC and H2O data sets, and good compression results are obtained.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2024.3423816