基于轻量化网络的海底垃圾检测算法SLD-Net  

SLD-Net:Seabed Garbage Detection Algorithm Based on Lightweight Network

作　　者：周华平 汪佳伟 ZHOU Huaping;WANG Jiawei(College of Computer Science and Engineering,Anhui University of Science and Technology,Huainan Anhui 232001)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001

出　　处：《湖北理工学院学报》2024年第1期41-46,共6页Journal of Hubei Polytechnic University

摘　　要：设计了一种轻量级的两阶段海洋垃圾检测算法SLD-Net(Seabed Litter Detection Network)。一方面,在预处理阶段构建了一个由颜色转换子模块和图像去噪子模块组成的图像增强模块MDIE,提高了图像质量;另一方面,采用轻量化的MobilenetV2作为目标检测的骨干网络,并引入一种改进的双路FPN结构,对深层特征图进行增强,提高对小目标的检测能力。经图像增强和目标检测网络联合训练后,进一步提升了模型的精度。在J-EDI海洋垃圾数据集上进行实验验证,mAP和速度分别达到了94.5%和65 FPS,且模型参数量仅有5.4 M,表明SLD-Net算法在精度、速度和参数量上达到了很好的效果。A lightweight two-stage marine garbage detection algorithm SLD Net(Sealed Litter Detection Network)has been designed,on one hand,an image enhancement module,MDIE,consisting of a color conversion sub-module and an image denoising sub-module is constructed in the pre-processing stage,which improves the image quality;on the other hand,the lightweight MobilenetV2 is adopted as the backbone network for target detection,and an improved two-way FPN structure is introduced to augment the deep feature maps and improve the detection of small targets.The accuracy of the model is further improved by the joint training of image enhancement and target detection networks.Experimental validation on the J-EDI marine litter dataset,the mAP and speed reach 94.5%and 65 FPS,respectively,and the number of model parameters is only 5.4 M,indicating that the SLD-Net algorithm achieves good results in terms of accuracy,speed and number of parameters.

关 键 词：神经网络 海底垃圾检测 轻量化网络 图像增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]