CCNNet: a novel lightweight convolutional neural network and its application in traditional Chinese medicine recognition
Gang Hu, Guanglei Sheng, Wang Xiaofeng, Jinlin Jiang
Abstract
Abstract With the development of computer vision technology, the demand for deploying vision inspection tasks on edge mobile devices is becoming increasingly widespread. To meet the requirements of application scenarios on edge devices with limited computational resources, many lightweight models have been proposed that achieves good performance with fewer parameters. In order to achieve higher model accuracy with fewer parameters, a novel lightweight convolutional neural network CCNNet is proposed. The proposed model compresses the modern CNN architecture with “bottleneck” architecture and gets multi-scale features with downsampling rate 3, adopts GCIR module stacking and MDCA attention mechanism to promote the model performance. Compares with several benchmark lightweight convolutional neural network models on CIFAR-10, CIFAR-100 and ImageNet-1 K, the proposed model outperforms them. In order to verify its generalization, a fine-grained dataset for traditional Chinese medicine recognition named “TCM-100” is created. The proposed model applies in the field of traditional Chinese medicine recognition and achieves good classification accuracy, which also demonstrates it generalizes well. The bottleneck framework of the proposed model has some reference values for the design of lightweight model. The proposed model has some promotion significance for classification or recognition applications in other fields.