摘要
由于背景环境复杂,检测物体易受部分遮挡、天气以及光线变化等因素的影响,传统目标检测方法存在提取特征难、检测准确率低、检测耗时长等缺陷.为了改善传统目标检测方法存在的缺陷,实现快速准确的目标检测,提出了一种基于快速区域卷积神经网络(faster regions with convolutional neural network,Faster-RCNN)算法的轻量化改进方法,即针对算法Inception-V2特征提取网络进行轻量化改进,并以带泄露线性整流(leaky rectified linear unit,Leaky ReLU)作为激活函数,解决使用线性整流(rectified linear unit,ReLU)激活函数存在的神经元输入为负数时输出为0的问题.基于上述改进方法,选择沙滩废弃物的检测为案例以验证方法的有效性,并且结合不同特征提取网络在检测沙滩废弃物时的表现,对比了SSD(single shot multibox detector)与Faster-RCNN算法.实验结果表明:所提改进算法在实际检测中有较好的综合性能,且相比原算法Faster-RCNN_Inception-V2,轻量化改进后的Inception-V2特征提取网络卷积计算量减少51.8%,模型训练耗时缩短了9.1%,检测耗时减少了10.9%,各类别AP的平均值(mean average precision,mAP)增加了1.02%,可见所提的改进方法能够有效提高目标检测的准确率,减少检测耗时,并在沙滩废弃物检测上得到成功应用,为海滨城市的沙滩清理维护提供了技术支持与保障.
Affected by adverse factors including complex environmental background,partially blocked objects,changes in weather and light,the traditional object detection method endures defects of difficult feature extraction,low detection accuracy,long detection time among others.To remedy these defects and to achieve rapid and accurate object detections,we propose a lightweight improvement method.It is based on faster regions with convolutional neural network(Faster-RCNN)algorithm,in which Inception-V2 is used as feature extraction network and the lightweight improvement on Inception-V2 network is made.When the rectified linear unit(ReLU)activation function is used in the Faster-RCNN algorithm,a defect is observed.Namely,if the input value of the neuron is negative,the output value becomes all 0,and Leaky ReLU is taken as the activation function.Based on these improved methods aforementioned,we select the detection of beach wastes as a case to verify the effectiveness of the improved algorithm.We then compare the performance of single shot multibox detector(SSD)and Faster-RCNN combined with different feature extraction networks.Experimental results show that the proposed algorithm comprehensively outperforms traditional methods in actual detection.After lightweight improvement of Inception-V2 feature extraction network used in Faster-RCNN,the convolution computation amount of Inception-V2 feature extraction network decreases by 51.8%;model training time decreases by 9.1%;detection time decreases by 10.9%;and mAP increases by 1.02%.Therefore,the proposed algorithm can effectively improve the accuracy of object detection,reduce detection time,and can be successfully applied to beach-waste detection,providing technical support and maintenance for beach cleaning and maintenance in coastal cities.
作者
龚圣斌
王少杰
侯亮
张荣辉
林晓涵
吴彬云
GONG Shengbin;WANG Shaojie;HOU Liang;ZHANG Ronghui;LIN Xiaohan;WU Binyun(School of Aerospace Engineering,Xiamen University,Xiamen 361102,China)
出处
《厦门大学学报（自然科学版）》
CAS
CSCD
北大核心
2022年第2期253-261,共9页
Journal of Xiamen University：Natural Science
基金
国家自然科学基金(51905460,51975495)
厦门市重大科技项目(3502Z20191019)。
