基于多尺度特征深度神经网络的不同产地山楂细粒度图像识别

Fine-grained image recognition of Cratargi Fructus from different origin based on multi-scale feature deep neural network

摘要:

中药是中医治疗疾病的主要途径，也是我国中医药事业传承与创新发展的物质基础，其真伪优劣也会直接影响中医临床的疗效，因此研究科学合理且高效的中药材质量检测方法符合当前行业热点.山楂作为中国著名的药食两用类药材，在烹饪和治疗中具有保护心血管、降低血压的作用，被广泛应用；但由于自然环境与栽培条件的不同，不同产地的山楂易被混淆从而对品质产生影响.尽管化学、生物鉴定的方法广泛而重要，但专业门槛高，耗时较长；且传统图像处理方法容易受外在环境因素干扰，可靠性差.因此亟待研究快速准确的方法以实现山楂产地的精准鉴别；受CoAtNet与Swin-Transformer网络启发，本文结合MBConv模块中深度可分离卷积网络对局部信息建模的特点与Swin Transformer模块多层次结构可弥补网络非局部性损失的特性，提出一种多尺度特征的混合神经网络模型，通过获取图像不同层级特征，将获取的形状、颜色与纹理等浅层特征作为先验知识与高层级语义信息进行特征融合，研究了一种快速有效的识别方法以实现对不同产地山楂的有效鉴别；此外，本文提出一种新的局部空间注意力机制，通过形成通道注意力模块联合空间注意力模块的新结构，实现对图像细粒度特征的聚焦与学习.实验结果表明，本文所提出的方法有最高的鉴别准确率为89.306%，优于其他基线模型.实践证明，本文的研究提高中药材鉴别的科技水平，拓宽传统中医药的研究思路.

Abstract:

Traditional Chinese medicine (TCM) is the primary approach for treating diseases and is also the he foundation for the development and innovation of TCM, the authenticity of TCM directly impacts the clinical efficacy. Therefore, scientific, reasonable, and efficient quality detection of TCM is a pressing research topic. Cratargi Fructus (CF) as a well-known edible food in China, which has been widely used for the ability of protecting cardiovascular and lowering blood pressure in cooking and treatment. However, it is reported that the difference in natural environment and cultivation conditions affects the CF’s quality and CF from different origins are easily confused, thus, the species authentication is necessary. Although physicochemical, biological, and manual identification methods are widely used, they have a high professional threshold and are inefficient. Image processing methods are easily affected by environmental factors, which reduces their reliability. Thus, there is an urgent need to study fast and accurate methods for the identification of CF. Inspired by CoAtNet and Swin-Transformer networks, we have proposed a hybrid neural network model with multi-scale features, combining the local information of the deep separable convolution network in MBConv and the non-local loss of the multi-level structure in Swin Transformer. By acquiring different features, the superficial features including shape, color and texture as prior knowledge have fused the high-level semantic information. A fast and effective recognition method is developed to realize the effective identification of CF from different origin. Furthermore, a new global spatial attention mechanism is introduced, which can focus and learn the fine-grained features of images by forming a new structure of channel attention module and spatial attention module. Our experimental results demonstrate that our proposed method has the highest identification accuracy of 89.306%, which outperforms other baseline models. This study highlights the potential for improving the scientific and technological level of TCM identification and broadening research on TCM.