摘要
在竹集成材的生产流程中,为了减少竹材加工过程中产生的浪费,需要精确获取竹材端面的内周长与厚度。而生长过竹梢的长竹条由于其纤维结构上的特征也无法进行粗刨。为提高竹产业生产效率,降低竹材浪费,提出了一种融合改进后的空洞空间卷积池化金字塔(ASPP)与双注意力机制(CBAM)的轻量级端到端网络模型。该网络既能够得到竹梢的精确位置,又能够提取出竹材端面的内外轮廓为后续计算竹材内轮廓周长与厚度提供条件。主干特征提取网络由4个卷积模块(block)组成,共实现16倍的下采样,每个模块内搭建残差结构来缓解梯度消失的问题;上采样阶段采用两次4倍的上采样实现端到端的输出,在两次上采样前分别加入改进后的空洞空间卷积池化金字塔与双注意力机制以提高网络输出的精度。该研究在训练阶段针对不同输出任务采用了不同的损失函数。试验表明,所提出的方法在分割竹材端面任务中交并比达到96.11%,竹梢位置检测任务中距离误差为3.09%,每秒传输帧数达到114.21。与LEDNet、BiSeNet-V2、RegSeg分割网络相比,本研究所提方法能够更好地平衡检测精度与检测速度。
In the production process of bamboo laminated lumber,in order to reduce the waste generated in the process of bamboo subdivision and rough section,it is necessary to determine the inner contour perimeter and thickness of the bamboo end surface.The long bamboo strip with bamboo branch also cannot carry out the rough section because of the feature on its fiber structure.In order to improve the production efficiency of the bamboo industry and reduce the waste of bamboo materials,this study proposed a lightweight end-to-end network model,which could not only obtain the precise position of bamboo branches,but also segment the inner and outer contours of bamboo end faces,providing conditions for subsequent calculation of the bamboo inner contour and thickness.The network can be divided into two stages,i.e.,down-sampling and up-sampling.In the down-sampling stage,a backbone feature extraction network was built by stacking 3×3 convolutional layers to achieve 16 times down-sampling.Each convolutional layer would be followed by a BN layer and a PReLU activation function.In order to alleviate the problem of gradient disappearance,this study also built a residual structure in the backbone.In the up-sampling stage,two times of four times up-sampling were used to achieve end-to-end output.Before the first up-sampling,the feature map was sent to the serial-parallel ASPP.Serial-parallel ASPP was improved on the basis of ASPP.It changed the convolution kernel of the first branch to 3×3 to further expand the receptive field.In order to alleviate the“grid effect”caused by the dilated convolution,the parallel dilated convolution branch was reduced to two,each with serial two dilated convolutions.Experiments proved that the improved ASPP can further improved the network accuracy.Before the second up-sampling,it was fused with the CBAM mechanism,which consisted of two parts,i.e.,the spatial attention module and the channel attention module.This research used different loss functions for different output tasks in the training sta
作者
石烨炜
鲍光海
SHI Yewei;BAO Guanghai(College of Electrical Engineering and Automation,Fuzhou University,Fuzhou 350108,China)
出处
《林业工程学报》
CSCD
北大核心
2023年第5期138-145,共8页
Journal of Forestry Engineering
基金
福建省科技计划项目(2018H0014)。
