基于跨模态特征融合的RGB-D花椒图像显著性检测  

RGB-D Pepper Image Saliency Detection Based on Cross-modal Feature Fusion

作　　者：李节 孙成龙 王逸涵 杨前 李柏林[1] LI Jie;SUN Chenglong;WANG Yihan;YANG Qian;LI Bailin(School of Mechanical Engineering,Southwest Jiaotong University,Chengdu 610031,China)

机构地区：[1]西南交通大学机械工程学院,四川成都610031

出　　处：《机械制造与自动化》2024年第6期211-217,共7页Machine Building & Automation

基　　金：四川省科技计划重点研发项目(2021YFN0020)。

摘　　要：针对现有显著性检测模型无法有效地协同花椒枝干彩色图像和深度图像特征,建立基于注意力的RGB-D图像花椒枝干显著性检测模型。由两个单流卷积网络分别提取彩色和深度图像特征;设计基于空间和通道注意力机制的跨模态融合模块,用于融合多尺度的彩色流和深度流特征;研发多尺度监督机制,用于缓解由于采用最近邻域上采样的解码方式导致边缘预测不准确的问题。实验结果表明:该方法的平均精确度、平均召回率、综合评价指标和平均绝对误差均优于对比显著性目标检测方法。To address the inability of existing saliency detection models to utilize the features of pepper branch color images and depth images effectively,an attention-based RGB-D image pepper branch saliency detection model is proposed.Color and depth image features are extracted separately by two single-stream convolutional networks.A cross-modal fusion module based on spatial and channel attention mechanisms is designed to fuse multi-scale color stream and depth stream features.A multi-scale supervision mechanism is developed to alleviate the inaccurate edge prediction caused by the use of nearest-neighbor upsampling decoding.Experimental results show that the average accuracy,average recall rate,comprehensive evaluation index and average absolute error of the proposed method are all superior to the compared salient object detection methods.

关 键 词：花椒自动化采摘 图像处理 RGB-D显著性目标检测 跨模态融合 注意力机制 多尺寸监督 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]