Identifying carrot appearance quality by an improved dense <scp>CapNet</scp>
Hongfei Zhu, Lianhe Yang, Yuxin Sun, Zhongzhi Han
Abstract
Abstract Carrot appearance quality is a significant concern of consumers and modern food process engineering. In the process of sorting in agricultural products, the agricultural products' appearance quality plays a particularly important role, so it is incredibly necessary to identify carrot appearance quality. In this study, we proposed an improved dense capsule network model (Modified‐DCNet) for carrot appearance quality identification. First, the initiator of Modified‐DCNet introduced the self‐attention layer that will reduce the interference of background information on the identification task. Specifically, the feature weight of the recognized area increased to reduce background information interference. Besides, the capsule layers of the improved model used a locally constrained dynamic routing algorithm to realize the sharing mechanism of capsule routing selection and transformation matrix in the local area, and this algorithm can reduce the size of network parameters and network training load. When the input image size is 227 × 227 pixels, the recognition accuracy of this model is 97.50%, and the parameter count is the only 5.2 M. The experimental results show that the improved dense capsule network model (Modified‐DCNet) has higher recognition accuracy and smaller parameter scales. Practical Applications In the design and manufacture of carrot sorter, it is vital to recognize the appearance quality of the carrot correctly. In this paper, Modified‐DCNet can achieve higher identification accuracy, dramatically reduce the burden of model parameter calculation. The carrot appearance plays a crucial role in quality assessment. Therefore, this work has practical application value.