The grade of wheat quality depends on the proportion of unsound kernels. Therefore, the rapid detection of unsound wheat kernels is important for wheat rating and evaluation. However, in practice, unsound kernels are hand-picked, which makes the process time-consuming and inefficient. Meanwhile, methods based on traditional image processing cannot divide adherent particles well. To solve the above problems, this paper proposed an unsound wheat kernel recognition algorithm based on an improved mask RCNN. First, we changed the feature pyramid network (FPN) to a bottom-up pyramid network to strengthen the low-level information. Then, an attention mechanism (AM) module was added between the feature extraction network and the pyramid network to improve the detection accuracy for small targets. Finally, the regional proposal network (RPN) was optimized to improve the prediction performance. Experiments showed that the improved mask RCNN algorithm could identify the unsound kernels more quickly and accurately while handling adhesion problems well. The precision and recall were 86% and 91%, respectively, and the inference time on the test set with about 200 targets for each image was 7.83 s. Additionally, we compared the improved model with other existing segmentation models, and experiments showed that our model achieved higher accuracy and performance than the other models, laying the foundation for wheat grading.

In the detection of global wheat heads, it is easy to give rise to difficulties due to different wheat varieties, planting densities and growth periods of wheat plants in different countries. In addition, the illumination conditions of the image collection and the complex background of field will also reduce the detection accuracy. It is also hard to accurately detect targets that are occluded and partially displayed in the image. To solve the above problems, in this paper, an improved YOLOv5 algorithm that integrates separable convolution and attention mechanisms is proposed. Firstly, the number of CSP modules of YOLOv5 is reduced to shrink memory consumption. Subsequently, vanilla convolutions in the CSP are replaced by separable convolutions which is also added to the fusion path and to reduce the redundant information of the feature map, so as to reduce the complexity of the model. In addition, the co-attention mechanism is added in backbone. Finally, the feature fusion module was adjusted to make the high-level features fuse more low-level information. Compared with the original algorithm, results show that the mAP of the improved algorithm reaches 93.8% which is 4.2% higher than that of the YOLOv5 algorithm, and the FPS is 27.4 which is 1.3 higher than YOLOv5. YOLOv7 is emphatically compared during model evaluation, other YOLO series and mainstream detection algorithms are also compared, and results show that our model has the best inference time and the best accuracy when dealing with high pixel images.