Abnormal area identification of corn ear based on semi‐supervised learning

Abstract

Abstract Screening corn ear is the key link in the breeding process of new varieties. But manual testing is difficult to measure the proportion of abnormal area. Meanwhile, the abnormal areas are mainly caused by mildew, moth and mechanical collision. A new refined semantic segmentation model was proposed based on the semi‐supervised learning method of generating antagonistic networks (GAN). Besides k‐means algorithm was used to remove a large amount of background information for data preprocessing. By using feature fusion and weighted loss function the model performance was improved. The introduction of transfer learning accelerated model convergence. Through the high‐throughput corn ear collection system, 1448 ear images (including abnormal conditions such as mildew, moth and mechanical damage) were collected and labelled. The proposed method was tested on real corn ear images with an accuracy of 0.950, mean precision of 0.933, mean IoU of 0.884, and FwIoU of 0.908. Experimental results show that the proposed method has better performance than general networks.