Green Citrus Detection In Field Using Deep Learning

In the process of fruit production,making full use of information and intelligent technology is of great significance for fruit growth situation detection,intelligent spraying water and fertilizer,yield estimation and automatic picking.Accurate and fast object detection of green fruit is one of the key technologies to realize the intelligent production of orchard.In an unstructured field environment,it is difficult to detect green citrus because of variability of light,complex background and variable fruit morphology.Deep learning can learn powerful feature representations from large amounts of image data,and object detection methods based on deep learning have high detection accuracy.In this research,deep learning technology is used to study green citrus detection.The main research contents are as follows:(1)Establish different fruit data sets.Green citrus dataset is obtained through field collection and handmade annotation.The green citrus images are randomly divided into training set and test set according to a certain proportion.In order to increase the amount of data during model training,the training set were enhanced by brightness adjustment,rotation transformation,horizontal mirroring and random erasure.At the same time,an apple data set is collected from the public literature for testing.(2)A detection method of green citrus based on improved Tiny-YOLOv3 is proposed.In field environment,the background of the scene is complex,the number of Tiny network layers is small,and the extracted features are not abstract enough.So the detection result of Tiny-YOLOv3 model is poor.A two-step convolution layer is adopted to Tiny network to replace the maximum pooling layer for down-sampling to reduce the loss of the target feature.Attention modules and multi-layer dense blocks are added to improve the feature extraction capability of Tiny network.The results of feature image visualization show that the improved feature extraction network can better suppress the background interference of leaves and branches.The results show that the TADYv3 model can alleviate the problems of overlap and occlusion,and improve the detection accuracy of green citrus in field.(3)Research on video detection method based on TADYV3.The TADYV3 model is difficult to detect the green citrus due to single visual angle and severe occlusion.Multi-view information of citrus fruits is obtained by video.The target of green citrus is tracked by Kalman filtering and Hungarian matching.The video detection of green citrus is realized.The effects of different target detection models and interval frames on target tracking results are compared.The experimental results show that video detection method based on TADYV3 not only further alleviates the problems of fruit overlap and occlusion,but also overcomes the problem that fruits are not visible from a single perspective.Real-time detection of green citrus in field is realized.