AUTHOR=Wu Chenjie , Xie Zhijun , Chen Kewei , Shi Ce , Ye Yangfang , Xin Yu , Zarei Roozbeh , Huang Guangyan 

TITLE=A Part-based Deep Learning Network for identifying individual crabs using abdomen images

JOURNAL=Frontiers in Marine Science

VOLUME=Volume 10 - 2023

YEAR=2023

URL=https://www.frontiersin.org/journals/marine-science/articles/10.3389/fmars.2023.1093542

DOI=10.3389/fmars.2023.1093542

ISSN=2296-7745

ABSTRACT=Crab (e.g., swimming crab and mud crab) is famous for its high nutritional value but is difficult to preserve. Thus, the traceability of the crab is vital in food safety. Existing deep learning methods can be applied to identify individual crabs. However, there is no work that identified individual crabs by using abdomen images. In this paper, we provide a novel Part-based Deep Learning Network (PDN) to reliably identify an individual crab from its abdomen images captured under various scenes.  In our PDN, we develop three non-overlapping partitions of the abdomen image, each emphasizing the extraction of different morphological characteristics of the crab's abdomen. We further develop three overlapping partition strategies to avoid the edge features of every partition being ignored and improve reliability. In particular, our PDN adopts the attention mechanism of the partition feature to strengthen the contribution of parts with high feature information, thus improving the individual identification accuracy. We create a swimming crab (Crab-201) dataset with 201 swimming crab abdomen images and a more complex mud crab dataset (Crab-146) collected by ourselves to train and test the proposed PDN. Experimental results show that the proposed PDN using the overlapping partition strategy is better than the non-overlapping partition strategy. The edge texture of the abdomen has more identifiable features than the sulciform texture of the lower part of the abdomen. It also demonstrates that the proposed PDN_OS3 (emphasizing the edge texture of the abdomen with overlapping partition strategies) is more reliable and accurate than the counterpart methods to identify an individual crab.