摘要
脐橙瑕疵检测突出问题是脐橙的果梗、脐部与瑕疵难以区分。针对这一问题,提出一种利用深度学习物体检测技术对脐橙的果梗脐部进行检测的算法。该模型以顺序卷积与跳跃式卷积共同提取深度特征;融合注意力机制加强待检测物体位置权重,在权重重分配的特征层上进行多尺度上下层信息融合,使用融合后的特征层进行默认框提取;对训练得到的模型进行模型压缩,进一步提升模型时间性能。实验结果表明,基于该模型能够准确实时识别定位出果梗、脐部不会与瑕疵产生误判,模型检测正确检测率达到90.6%,单幅图片预测时间降低为15 ms。
The prominent problem in the detection of navel orange is that the stem end,blossom end and defect of navel orange are difficult to distinguish. For this problem,we proposed an algorithm for detecting navel orange stem end and blossom end by using deep learning object detection technology. The model extracted depth features together with sequential convolution and skip-connectional convolution. The fusion attention mechanism strengthened the position weight of the object to be detected,and multi-scale upper and lower information fusion was performed on the feature layer of weight redistribution. The feature layer performed default box extraction. The model obtained by training was model-compressed to further improve the model time performance. The experimental results show that based on the model,the fruit stem end can be accurate and real-time identified,and the blossom end is not misjudged with the defect. The correct detection rate of the model detection is 90.6%,and the prediction time of the single picture is reduced to 15 ms.
作者
杜雨亭
李功燕
许绍云
Du Yuting;Li Gongyan;Xu Shaoyun(School of Microelectronics,University of Chinese Academy of Science,Beijing 100049,China;Institute of Microelectronics of Chinese Academy of Sciences,Beijing 100029,China)
出处
《计算机应用与软件》
北大核心
2019年第7期208-212,共5页
Computer Applications and Software
基金
国家重点研发计划项目(2018YFD0700300)
关键词
卷积神经网络
脐橙
物体检测
注意力机制
Convolutional neural network
Navel orange
Object detection
Attention mechanism
