基于深度神经网络的藏药材识别分类研究

Research on Classification and Recognition of Tibetan Medicinal Materials Based on Deep Neural Network

为了实现对藏药材快速、高效、准确的鉴定识别,本文采用深度学习的方式提出了一种对称卷积核多分支残差块。该块不仅可以根据藏药材的差异动态调节卷积的参数,而且可以在更细粒度的级别上提取多尺度空间信息,并形成长距离的信道依赖性,然后将此块进行堆叠,形成了一个新型的网络结构——深度对称卷积核神经网络。实验结果表明,该网络在相同条件下取得了参数量-计算量-准确率之间的平衡,具有更好的性能。

In order to achieve rapid, efficient and accurate identification and identification of Tibetan medicinal materials. In this paper, we propose a symmetric convolution kernel multi branch residual block based on deep learning. This block can not only dynamically adjust the convolution parameters according to the differences of Tibetan medicinal materials, but also extract multi-scale spatial information at a finer granularity level and form channel dependence for long distance. Then stack this block, a new network structure-deep symmetric convolution kernel neural network is formed. The experimental results show that under the same conditions, the network achieves a balance between parameter quantity, computer quantity and accuracy, and achieves better performance.