Identification of Hollow Edamame Using HSI Based on Deep Learning

Identification of hollow edamame with pod remains a challenge due to its similarity in the appearance with normal edamame with pod. In this study, Hyperspectral reflection/transmission imaging (RHSI/THSI) were used to obtain hyperspectral reflection/transmission image. The reflection characteristic images (RIs) and transmission characteristic images (TIs) were extracted by SC (SPA and CARS) combined with image quality assessment coefficient (IQAc). Aiming at the problem of low grayscale in TIs, targeted optimization was carried out. SSD, YOLOv5 and YOLOv8 were constructed based on RIs, TIs and optimized transmission characteristic images (OTIs) for the Identification of hollow edamame. The research results show that all models can efficiently identify hollow edamame, with the recognition accuracy reaching over 95%. In comparison, THSI performs better than RHSI in discriminating hollow edamame. Among all the models, OTIs-YOLOv8 has the highest Identification precision, with a precision of 99.57% and a recall of 99.9%. The results indicate that the optimized transmission characteristic images exhibit more significant advantages than the reflection characteristic images in the discrimination of hollow edamame. The research of this study provides a solid methodological approach and reliable experimental basis for the quality evaluation of edamame and the development of related rapid detection equipment.