Grading evaluation of haploid fertility restoration traits based on inception-ResNet in maize

Double haploid (DH) technology can significantly shorten the breeding cycle and improve the breeding efficiency, and it is favored by breeders. The metrics for evaluating the effect of haploid genome doubling mainly include anther emergence and ear seed setting. The evaluation of fertility restoration ability is mainly conducted through visual inspection at present, which is time-consuming, and easy to be affected by fatigue, resulting in errors and inconsistencies. Therefore, it is urgent to develop efficient and accurate evaluation technology to reduce the field work burden of researchers. In this work, we propose a grading evaluation model (Maize-IRNet) of haploid anther emergence and ear seed setting based on Inception-ResNet. Firstly, the modules of Stem and Inception-ResNet are utilized for image feature extraction and multi-scale feature learning. Then, the Reduction module is used for spatial downsampling and feature compression, and the global attention mechanism (GAM) is used to enhance the recognition of key regions of the image. The experimental results show that the Maize-IRNet’s classification accuracy of haploid ear seed setting and anther emergence is 84.2% and 84.0%, which is higher than six baseline methods (VGG11_bn, ResNet50, ResNet101, ViT-Base-16, gMLP, MLP-Mixer). In order to facilitate the practical application for breeding researchers, we have developed a mobile application that integrates the Maize-IRNet model. This study helps to achieve high-throughput collection of fertility restoration phenotypes, improves the evaluation efficiency of fertility restoration, reduces breeding costs, and provides technical support for the promotion of engineering breeding of DH technology.