基于深度学习的半监督图像标注系统设计与实现

Design and implementation of semi-supervised image labeling system based on deep learning

关键词:  图像标注  深度学习  目标检测  食用菌  半监督

Key words:  image labeling  deep learning  object detection  edible mushroom  semi-supervised

摘要:

针对深度学习研究中标注训练样本费时费力的问题，以食用菌为研究对象，设计一种基于深度学习的半监督图像标注方法。该方法将深度学习目标检测模型与迭代图像标注工作有效结合，采用“检测模型训练—目标自动检测—人工标注修正—检测模型更新”的迭代操作，实现半监督方式的图像标注。基于所设计的方法构建了半监督图像标注系统，在试验中对系统进行性能评测和分析。结果表明:迭代更新后的检测模型在测试集上的检测准确率为98.1%，召回率为88.5%，平均准确率为88.3%；利用所构建的半监督图像标注系统可以实现15 s/幅的标注速度，单幅图像的标注耗时仅为纯手工标注耗时的2.5%，图像标注时间代价大幅降低。研究结果为深度学习研究中的训练样本标注提供了高效的标注方法和工具，有助于提高图像标注效率，减少人力成本投入。

Abstract:

In order to tackle the problem of time-consuming and labor-intensive training samples labelling in deep learning, a semi-supervised image labeling method based on deep learning for edible mushroom was designed. The method effectively combines the object detection model with iterative image labeling, and realizes the semi-supervised image labeling through the iterative operations of “detection model training-automatic object detection-manual labels correction-detection model updating”. Based on this method, a semi-supervised image labeling system was constructed and the performance of the system was tested and analyzed in the experiments. The experimental results show that the detection precision, recall and average precision of the model after iteratively updated achieves 98. 1%, 88. 5% and 88. 3% on the test set, respectively. Using the constructed semi-supervised image labeling system, the labeling speed achieves 15 seconds per image and the labeling time of a single image is only 2. 5% of that of pure manual labeling, which greatly reduces the time cost. The research achievement provided an efficient labeling method and tool for training samples labeling in deep learning study, which is helpful to improve the efficiency of image labeling and reduce labor cost.