Visual feature-based improved EfficientNet-GRU for Fritillariae Cirrhosae Bulbus identification

Fritillariae Cirrhosae Bulbus (FCB) as a well-known traditional Chinese Medicine (TCM), which is widely used for its ability of relieving cough and eliminating phlegm in cooking and treating. However, the adulteration by different species for economic profit has frequently been reported. Inspired by deep learning, a novel approach based on image captioning is proposed to achieve the accurate and fast identification of FCB: EGNet, via bridging between image visual features and word expression in Chinese Pharmacopoeia. In encoder module, Convolutional Block Attention Module (CBAM) and spatial attention module (SA) are introduced into EfficientNet-B0 to strengthen and focus on the unique features. For decoder module, due to the simpler structure and fewer parameters, gated recurrent unit (GRU) is applied for generating the correspondence and explanation with text descriptions. Simultaneously, the adaptive attention mechanism with a visual sentinel is inject into GRU for judging adaptively whether to rely on visual information or semantic information. Eventually, experiments confirm that the proposed EGNet outperforms competing methods. And it is superior in the highest identification accuracy of 99.0%, 99.3% and 99.4%, the best words matching completeness 91.1%, 92.2% and 91.6% for Lubei, Qingbei, and Songbei. This paper can significantly improve the accuracy of classification and the cost is low. It is proved to be an exceptional practice for the high-efficiency of TCM-discrimination and TCM-technology.