Tiyapunjanit P.Siammai T.Klaythong K.Jantrachotechatchawan C.Duangrattanalert K.Mahidol University2023-12-092023-12-092023-01-012023 3rd International Conference on Emerging Smart Technologies and Applications, eSmarTA 2023 (2023)https://repository.li.mahidol.ac.th/handle/20.500.14594/91348Detecting susceptible shrimp larvae presents a significant challenge that requires dedicated effort, skill, and expertise. In this study, we propose an advanced approach that combines probabilistic deep learning with transfer learning and deep metric learning using a triplet loss function. To obtain an optimal solution, we employed 5-fold cross-validation to conduct a rigorous comparison among different variations of the model. Our results showed that integrating transfer learning and deep metric learning provides significant improvements to our system. Through our experiment, we observed that DenseNet121 appeared to become more effective when incorporated with the proposed techniques, achieving an accuracy of 92%, sensitivity of 87%, and specificity of 97%. Importantly, after tuning, the model gained an exciting ability to generalize to previously unseen datasets, surpassing an accuracy of 90% on different background settings (i.e., varying colors and textures). These findings have undoubtedly demonstrated the efficacy of our methodology in accurately and robustly identifying susceptible shrimp larvae. Source code and sample data for our study can be downloaded directly from our GitHub repository (https://github.com/patipond-tiy/vannameivision).Computer ScienceConference PaperSCOPUS10.1109/eSmarTA59349.2023.102934952-s2.0-85178165590