Research And Implementation Of Immature Mango Detection Based On Convolutional Neural Networks

Automatic target detection of immature mango is premise of planting and management tasks,such as intelligent spraying,monitoring of growth situation,early yield estimation and so on.Computer vision technology provides an effective and convenient way for fruit target detection,and has become the mainstream detection technology all over the world.In orchard scenes,it is difficult to detect immature mangoes because of variability of light,complex background and high color similarity between mangoes and leaves.Especially,detection of occluded and overlapped mangoes is an extremely challenging task.Deep convolutional neural networks has been attracted much attention in recent years.Since it extracts more abundant and abstract convolution representations than traditional manual designed features,and has a more accurate and rapid target detection ability.In this paper,research of immature mango detection in orchard scenes based on deep learning methods of region proposal and regression was proposed.Moreover,an immature mango detection system for practical application was designed.The main research and innovation works are as follows:?1?Establishment of mango image dataset.It was obtained through field collection and handmade annotation.In order to reduce the influence of illumination in natural scenes,Adaptive Histogram Equalization was adopted to improve image quality and to enhance diversity.For natural scenes with occluded and overlapped mangoes,a method of strengthening convolutional representations of mango foreground region using foreground region labelling was proposed,avoiding the extraction of non-target features during network training.In order to further expand the scale of training data and improve its diversity,data size was augmented by horizontal flipping and multi-angle rotating with angles of±10°and±20°.?2?Mango target detection based on Faster R-CNN.In order to overcome the dependence of traditional detectors on manual designed features and problem of low accuracy,mango target detection based on Faster R-CNN was proposed.In experimental images,positive-negative class imbalance encountered because of the background areas was much larger than the target areas.Therefore,Focal loss function was introduced to avoid the influence.Anchor parameters of scale and width to height ratio were optimized by mathematical statistics in order to enhance the region generation ability.The results of Faster R-CNN model showed that,the precision of deeper VGG-16 network with 13 convolutional layers was 14.16%and 12.64%higher than that of ZF and VGG-CNN-M-1024 networks respectively.In addition,the precision and recall rate reached up to 94.55%and 93.17%with a detection speed of 11 fps,at the scale of{64²,128²,256²},width to height ratio of{0.5,0.6,0.7,1,2}and focal loss parameters of?=1,?=0.35.?3?Immature mango detection based on improved YOLOv2.YOLOv2 is a regression based detection algorithm,which simultaneously predict location and class of targets in an image,and has faster detection speed.In order to improve the rapidity and accuracy of mango detection,an improved YOLOv2 model was proposed by introducing dense connectivity structures and optimizing multi-scale training strategy.In aspect of base networks,Tiny-yolo with 9 convolutional layers was selected.In order to enhance the use of shallow layer features and to abate gradient disappearance,the 7^th convolutional layer was replaced by a self-designed dense block.The result showed that precision and recall rate increased by 0.9%and1.34%respectively after improvement.In aspect of training,multi-scale strategy was selected to adapt to mangoes with different scales in natural scenes.Through comparison test,input scale was increased from 448² to 512²,and training scale was increased to{384²,416²...672²}.At a detection rate of 83 fps,the precision and recall rate reached up to 97.02%and95.1%respectively.The performance was better than Faster R-CNN,YOLO and Adaboost model.?4?Design and implementation of mango detection system.In order to provide intelligent and precise auxiliary services for agricultural planting and management,a simple and easy-to-use mango detection system was designed through Microsoft Visual Studio 2013platform and OpenCV 2.4.9 library.The system could be divided into four modules:Model loading,Data acquisition,Intelligent detection and Result output.Therefore,a function of real-time mango target detection was feasible.The information,such as detection result,loading model,total number of targets,detection time,task status and device name and so on,was visualized on the interface,providing users with intuitive and clear grasp of detection situations.