Rapid detection of turtle cracks in corn seed based on reflected and transmitted images combined with deep learning method

Abstract

Cracks in corn seeds greatly affect their storage tolerance and germination ability. These cracks are categorized as single, double, and turtle cracks. The vigor experiment showed that turtle cracks lead to the poorest seed vigor, with significantly lower germination rates after artificial aging. Traditional methods to detect turtle cracks, like lightboxes, staining, and colorimetry, are inefficient or costly. Therefore, a detection method for turtle cracks of corn seeds based on reflected light and transmitted light images combined with deep learning was proposed. The transmitted and reflected light images on embryo and endosperm surfaces of 2000 corn seeds with different crack were collected. A Based on Self-Attention Modul Convolutional Neural Support Vector Networks (SACNSVN) model was proposed to detect the turtle cracks seeds, and compared with the model established by LSTM, CNN, BLR, SVM, and MLP. The effects of embryo surface, endosperm surface, varieties and maturity on turtle crack detection were investigated. The results showed that: (1) The transmitted light image had a significant correlation with 62 and 58 indicators of cracks and turtle cracks, respectively, which better reflected the crack and turtle crack seeds than reflected light image; (2) The accuracy of turtle cracks detection model based on embryo endosperm surface combined dataset were higher than embryo surface, endosperm surface, and embryo endosperm surface mixed dataset; (3) The SACNSVN model provides the optimal accuracy with 82.6%; (4) Variety greatly affects turtle crack detection in corn seeds, while maturity level has a minor impact, and SACNSVN exhibits strong generalization ability. This study provides significant technical support for optimizing the seed processing process and improving the seed quality.