| In order to solve the problem of low efficiency of traditional artificial wheat ears counting and yield prediction in field,this paper took wheat ears as the research sample and studied the real-time detection method of small targets and dense targets under deep learning.At first,the wheat ear images data set was established,including image segmentation,image marking and data enhancement.The second,we built the YOLOv4 network model based on Tensor Flow and adjusted it to improve the transfer learning.We introduce attention mechanism into the model to improve the attention of small dense ears and solve the problem of small shape and complex feature extraction.At the same time,we improved the prediction part of the model and modified the prediction structure by clustering.This increases the number of prior frames and solves the problem of dense and overlapping ears.We discuss the factors affecting the performance of the model through a lot of comparative experiments.In image segmentation we explore the influence of wheat ear images with different resolutions on network model detection results.The results show that the difference between foreground and background can be improved by changing the image resolution to determine the optimal pixel ratio of wheat ears,which has a significant effect on small dense wheat ears,and effectively improves the detection performance of small dense wheat ears.Data enhancement effectively expands the data set,increases the diversity of images,and solves the problems of small data set and insufficient model generalization ability during training.We also trained different models of YOLOv4,YOLOV4-tiny,and Faster R-CNN to compare with the improved model,and explored the influence of attention mechanism on the detection model as well as the influence of the anchors size and the number of anchors on the detection model.The results show that the attention mechanism can make the network obtain more larger receptive field to capture the global features of the target and increase the sensitivity of the small dense target.At the same time,clustering can make the anchors better match the size of the detected target,and increase the number of anchors to make the model prediction more accurate.Finally,in the test set,the detection accuracy of wheat ears in different periods,environments,varieties and sizes was 94.8%,and the detection speed of a single image was 57 FPS,which met the real-time detection of wheat ears.This results provide a technical support for wheat ear counting and yield prediction in the field. |
