In order to classify the appearance quality level of Fritillaria thunbergii, the F. thunbergii dataset was constructed with the DigiEye system followed by an image annotation tool. Several statistical learning and object detection algorithms were selected to train and test the F. thunbergii dataset. The results showed that the model trained by the YOLO-X of YOLO (you only look once) series had relatively better performance. In addition, to optimize YOLO-X, according to the unique features of F. thunbergii dataset, a dilated convolution structure was embedded into the end of the backbone feature extraction network of YOLO-X as it could improve the model sensitivity to the dimension feature. The mean average precision (mAP) of the improved model was raised to 99.01%; the average precision (AP) for superfine, level one, level two, moth-eaten, mildewed, and broken F. thunbergii were raised to 99.97%, 98.33%, 98.47%, 98.71%, 99.73%, and 98.85%, respectively; and the weighted harmonic mean of precision and recall (F₁) were raised to 0.99, 0.92, 0.94, 0.97, 0.99, and 0.97, respectively. The tune-up in this study enhanced the detection performance of the model without increasing the number of parameters, computational complexity, or major changes to the original model. This study provides a scientific basis for the subsequent construction of F. thunbergii detection platform.

In order to solve the precision and real-time problems of parental discrimination in the processes of hybrid rice breeding and pollination, an improved DeepLabV3+ hybrid rice breeding parental discrimination semantic segmentation model based on a fully convolution neural network was proposed. The lightweight MobileNetV2 structure of the backbone network was used to replace the Xception structure of the original DeepLabV3+ backbone network, which is more suitable for the application on mobile devices. An extraction method of low-level features with close connection was proposed. The lower-level information and higher-level information were preliminarily concated as the input of the original lower-level information, which enabled the network to obtain more intensive information, thus enhancing the ability of the network to extract details. The results showed that the improved DeepLabV3+ network model had higher segmentation precision for parents of hybrid rice seed production than the original DeepLabV3+ network model, and reduced the model training time and image predictive time. Compared with other mainstream network models and advanced network models, it is found that the accuracy of different parameters of improved DeepLabV3+ network model is improved. This study provides a reference for the development of deep learning in the field of agricultural visual robots.