Computer Engineering and Applications ›› 2024, Vol. 60 ›› Issue (3): 88-99.DOI: 10.3778/j.issn.1002-8331.2308-0333

• Special Issue on Object Detection • Previous Articles     Next Articles

Improved Lightweight Underwater Target Detection Algorithm of YOLOv7

XIN Shi’ao, GE Haibo, YUAN Hao, YANG Yudi , YAO Yang   

  1. School of Electronic Engineering, Xi’an University of Posts and Telecommunications, Xi’an 710121, China
  • Online:2024-02-01 Published:2024-02-01

改进YOLOv7的轻量化水下目标检测算法

辛世澳,葛海波,袁昊,杨雨迪,姚洋   

  1. 西安邮电大学 电子工程学院,西安 710121

Abstract: Aiming at the problems of target false and missing detection caused by limited memory and computing power of underwater equipment and complex underwater environment, a lightweight underwater target detection method YOLOv7-SDBB is proposed. The ShuffleNetv2 lightweight network is introduced on the backbone network of YOLOv7 to reduce the parameter amount and calculation amount of the feature extraction network. The D-ELAN and D-MPConv modules are designed to further realize the network lightweight and improve the model detection speed. Due to the phenomenon of false and missing detection is prone to occur during underwater detection, BiFPN is used to perform multi-scale feature fusion and integrate deep feature information. In view of the problem of feature information loss caused by BiFPN feature fusion, the BiFormer attention mechanism is used to retain key information and improve target detection accuracy. The experimental results show that the accuracy of the improved model on the URPC2020 dataset has increased by 2.7 percentage points, the amount of parameters and calculations have decreased by 20.3% and 41.7% respectively, and the detection speed has increased to 100.9?FPS, realizing a good balance between the speed and accuracy of underwater target detection.

Key words: lightweight network, underwater target detection, YOLOv7, sparse attention mechanism

摘要: 针对水下设备内存和计算能力有限和水下环境复杂造成的目标错检和漏检问题,提出一种轻量级水下目标检测方法YOLOv7-SDBB。在YOLOv7的骨干网络上引入ShuffleNetv2轻量级网络,降低特征提取网络的参数量和计算量;设计了D-ELAN和D-MPConv模块,在进一步实现网络轻量化的同时提高模型检测速度;由于水下检测过程中容易出现错检、漏检的现象,利用BiFPN(bidirectional feature pyramid network)进行多尺度特征融合,融合深层的特征信息;针对BiFPN特征融合导致的特征信息丢失的问题,采用BiFormer注意力机制保留关键信息,提高目标检测精度。实验结果表明,改进后模型在URPC2020数据集上的精度提高了2.7个百分点,参数量和计算量分别下降了20.3%和41.7%,检测速度提升至100.9?FPS,从而验证了提出的算法在精度和速度之间取得了很好的平衡。

关键词: 轻量级网络, 水下目标检测, YOLOv7, 稀疏注意力机制