Research On Target Counting And Plant Detection Algorithm Based On Deep Learning

It is necessary to count large-scale interested targets quickly in related areas of management,industry and daily life.For example,counting the crowd is a necessity in the certain scenes of early warning for urban emergency,public management and traffic monitoring;in modern agricultural production,it is necessary to monitor the whole process of crop growth and estimate the yield,that is,through field crop image monitoring,identification and counting,to achieve the purpose of phenotypic analysis and yield estimation,which is because that the accuracy of yield estimation directly affects the grain futures price and is also an important basis to ensure national food security.With the development of artificial intelligence technology,computer vision system based on deep learning algorithm has made a breakthrough in the application of target counting,object detection and other fields.It also provides technical support for accurate yield estimation and phenotypic research of crops.Firstly,this research focuses on the crowd counting algorithm based on convolutional neural network.To solve the problem of the drastic scale change of pedestrian counting in the actual scene,we improved the classical congestion scene counting network(CSRNet)by following processes.In the front section of feature extraction network,by combined with spatial pyramid feature extraction algorithm,the multi-scale feature information enhancement module is designed to extract and enrich the multi-scale feature information in the image.In the back section of network,the feature fusion structure is designed so that the dilated convolution network with different hole rates can be adopted to integrate multi-scale features and to generate corresponding density maps.The network is trained and tested on Shanghai-Tech public population counting dataset.By compared with other five classical counting algorithms based on density maps,the experimental results show that the proposed algorithm achieves higher counting accuracy and generates high-quality density estimation maps.Secondly,this thesis studies on the crop counting algorithm by chosen wheat yield counting as an example.In dealing with the problem of disorderly distribution of wheat ears in field environment,a wheat ear counting algorithm based on context feature perception is proposed.Both high and low-level semantic features obtained by the multi-scale feature information enhancement module are re-integrated,and a new feature fusion network model is established.This model can maintains the full information of the image and can lift the weight of the partial information,as a result,enhance the expression ability of the network to the fused features.Experiments on the marked GWHD wheat ear dataset show that the proposed algorithm has higher counting accuracy than several classical counting algorithms.On the basis of completing the tasks of counting and yield estimation,in order to realize the accurate positioning of wheat ears and provide support for further phenotypic research,this thesis developed a wheat ear object detection algorithm based on no anchor frame,which not only reduces the super parameters,but also avoids the complex calculation related to the anchor frames.Anchor free object detector named Fovea Box is introduced to calculate the cost matrix in order to adapt to the changes of image scale,and the corresponding loss function is optimized.Two open wheat ear detection datasets,GWHD and WEDD,are used to train and test the algorithm.The experimental results show that the algorithm improves the accuracy of wheat ear object detection in real field environment.