| The accurate count of fruits is the key of intelligent agricultural yield estimation,which has a high reference value for evaluating crop growth and exploring the problems existing in the planting industry.The preservation period of fruits is short,the counting of young fruits before harvest can effectively estimate the final yield,guide agricultural growers to adjust the process plan,reduce the labor cost of harvest and guarantee economic benefits.With the development of computer vision technology,automatic fruit counting using image information processing technology is a hot topic in current research.Fruit targets in natural environment are affected by many factors such as light condition and environmental occlusion,so it is difficult to achieve accurate counting.At the same time,fruits of mature plants vary in shape and size,and the existing image counting method requires manual realtime adjustment and optimization of model parameters,which increases the production cost.In this paper,a method to count fruit images of plants in different complex environments is proposed,which can realize the automatic counting of fruit images of mango,apple,etc.In this method,the improved context information model was used to refine the fruit feature information,and the multi-scale features were fused with the feature fusion model to achieve the purpose of fruit count.At the same time,in order to reduce the accuracy loss caused by the truth graph method in the supervision of network training,this paper adopts the paired network supervision method based on the Hungarian algorithm to optimize the training process of the network,so as to achieve higher counting precision and more accurate counting results.The main work of this paper is as follows:(1)For fruit images collected in orchards under different lighting and shooting conditions,this paper proposes a fruit image counting model,which has three typical characteristics: a)It is an automatic counting model based on deep learning architecture;b)It has strong robustness to images with complex background and uneven illumination;c)It has good generalization ability for different varieties of fruit and can be applied to a variety of different fruit counts.(2)This paper proposes an improved context information fusion structure,which can pay attention to fruit features of different scales and fuse multi-scale feature information in a more detailed manner,effectively improving the problem of insufficient detail feature information extracted by the front-end network.(3)This paper conducted comparative experiments with other methods on the published plant fruit data set(including apple and mango),and the experimental results showed that the method in this paper could achieve satisfactory results on different plant fruit images,and the counting accuracy and universality were improved to a certain extent. |
