| Corn is one of the important crops in our country.During the growth of corn,the images of corn plants can be captured via an unmanned aerial vehicle.For these images,the computer vision technology can be adopted to count the plants to estimate the total corn output,and formulate a breeding plan for the next year.However,the actual captured corn plant images may be affected by some unfavorable factors,such as illumination reflection,background similarity,etc.The degraded images will decrease the accuracy of plant counting,and affect the correct display of the distribution of corn plants.In order to deal with the above problems,two woks have been carried out in this thesis.(1)A corn plant counting algorithm is proposed based on an improved YOLOv3.First,for the corn plant images in the training set,the diversity is enhanced through the data expansion realized by mirror flip and rotation operation.At the same time,the histogram equalization is adopted to enhance the image contrast and reduce the impact of illumination reflection.In order to improve the detection accuracy,an improved YOLOv3 plant detection network is designed based on feature fusion operation and residual module.Finally,the generated prediction bounding box is screened through the non-maximum suppression algorithm.Furthermore,the remaining detection results are counted to obtain the corn plant number.The experimental results on two corn plant data sets with different illumination intensity verify the effectiveness of the proposed algorithm.(2)A corn plant counting algorithm is proposed based on scale-aware and channel attention.First,the problem of corn plant counting is considered as a mapping relationship between the input image and the density map.In order to generate high-quality prediction density maps,the scale-aware context module and the channel attention module are used and merged.The scaleaware context module is adopted to extract multi-scale context information in the image,and the channel attention module is utilized to capture the channel dependency to alleviate background error estimation.Secondly,in order to improve counting sensitivity,a hybrid loss function is adopted based on absolute count loss and density map loss.Finally,the predicted number of corn plants is obtained by summing the probability values of the pixel points in the generated density map.In addition,the distribution of corn plants in different regions can be easily observed via the predicted density map.Compared with other counting algorithms,the experimental results demonstrate the superiority of the proposed algorithm. |
