基于改进YOLOX的水下垃圾检测算法

Underwater Garbage Detection Algorithm Based on Improved YOLOX

基于机器视觉的水下垃圾清理机器人已经成为修复海洋生态的一种有效手段,但是由于复杂的水下环境会造成采集图像的分辨率较低,导致垃圾检测精度较低。针对上述问题,提出一种基于改进YOLOX-S网络的水下垃圾检测算法,该算法通过采用空间到深度卷积模块代替下采样模块提高了图像中物体有效特征的提取能力,提升了其检测精度;主干网络引入空洞空间卷积池化金字塔模块增强了深层特征提取能力,以及颈部网络引入轻量化幽灵混洗卷积模块和Vov幽灵混洗跨阶段瓶颈模块获取了更多的多尺度特征信息,进一步提升检测精度。实验结果表明,在YOLOX网络中引入空间到深度卷积模块、幽灵混洗卷积模块和Vov幽灵混洗跨阶段瓶颈模块、空洞空间卷积池化金字塔模块均可提高YOLOX模型的检测精度。改进后YOLOX-S模型的平均精度均值(mean average precision,mAP)达到了67.4%,较原YOLOX-S模型提高了3.1%,有效提升了复杂海洋环境中的垃圾检测能力。

Machine vision-based underwater garbage cleaning robots have become an effective means to restore marine ecosystems.However,the complex underwater environment often leads to low-resolution image acquisition,resulting in low accuracy in garbage detection.To address this issue,this paper proposes an improved YOLOX-S network-based algorithm for underwater garbage detection.This algorithm replaces the down-sampling module with a spatial-to-depth convolution module to enhance the capability of extracting effective object features from images and improve detection accuracy.The introduction of the dilated spatial convolutional pyramid module in the backbone network enhances deep feature extraction,while the lightweight Ghost Shuffle convolution module and Vov Ghost Shuffle cross-stage bottleneck module in the neck network capture more multi-scale feature information,further improving detection accuracy.Experimental results demonstrate that incorporating the spatial-to-depth convolution module,Ghost Shuffle convolution module,Vov Ghost Shuffle cross-stage bottleneck module,and dilated spatial convolutional pyramid module into the YOLOX network can improve the detection accuracy of the YOLOX-S model.The improved YOLOX-S model achieves an average precision mean average precision(mAP)value of 67.4%,which is a 3.1%improvement over the original YOLOX-S model.This effectively enhances the garbage detection capability in complex marine environments.