Abstract:

An improved Deeplab v3+ network model semantic segmentation algorithm with dynamic learning features is proposed for the problems of high-level labor cost and low-level intelligence in the management of domestic surgical instruments. In order to strengthen the effective feature learning of related tasks, the CBAM module of the attention mechanism is embedded into the Deeplab v3+ encoder and the high-level features of images are extracted through dense depth wise separable convolution and dilated convolution. Two low-level feature sources are added to the decoder, which retains important feature information and improves segmentation accuracy. The experimental results show that the MIoU, PA, Recall, and F-measure value of the improved network on the surgical instrument dataset are 0.854, 0.874, 0.872, and 0.873, respectively. Compared with other semantic segmentation networks, this improved network segmentation performance is better and has great engineering and practical value.

Key words: deep learning, automatic management of surgical instruments, semantic segmentation, attention mechanism

摘要：

针对当前国内手术器械管理耗费人力，智能化程度低的问题，提出一种动态学习特征的改进Deeplab v3+网络模型语义分割算法。为了加强相关任务有效特征学习，在Deeplab v3+模型编码端嵌入注意力机制CBAM模块并通过密集深度分离卷积和扩张卷积提取图像高层特征；在解码端增加两路低层特征来源，保留了重要特征信息，提高了分割准确率。实验结果表明，改进后网络在手术器械数据集上MIoU、PA、Recall、[F]值分别为0.854、0.874、0.872和0.873。相较于其他语义分割网络，改进网络分割性能更优，有极大的工程实用价值。

关键词: 深度学习, 手术器械自动化管理, 语义分割, 注意力机制