Research And Implementation Of Grape Image Recognition For Mobile Devices Based On Deep Learning

Driven by the wave of artificial intelligence,the intelligent management of orchards is imperative.Among them,the automatic identification and detection of fruits is an important prerequisite for intelligent planting and picking.In the natural environment,the detection of grapes in orchards has problems such as leaves,mutual blocking of grapes and uneven light irradiation.The more prominent problem for the detection of green grapes is that the foreground（green grapes）and background（grape leaves）are similar,making manual Conventional detection methods for feature extraction are difficult to accurately detect,and detection methods based on convolutional neural networks can automatically extract richer features and use end-to-end training and detection methods to meet the needs of grape orchard scene detection.This article takes grapes in grape orchard scenes as study objects,uses deep convolutional neural networks to automatically detect orchard grapes in natural scenes,and designs orchard grape detection systems to achieve fast and efficient detection of orchard grapes,which is the intelligence of grapes picking lays a solid foundation.The main research work and conclusions are as follows:Establishment of a grape orchard dataset.By shooting 3223 images of 11 kinds of grapes in the orchard and collecting them manually,the grape image data set was established.In order to further expand the data set and increase the diversity of training data,the data set was amplified using horizontal flip,translation,random block sampling and other methods,and a total of 15,334 training data and 325 test data were obtained.Faster R-CNN orchard grape detection method based on ROI Align.First,the feature extraction network in Faster R-CNN is improved,and the residual network is used to replace the VGG16 feature extraction network in Faster R-CNN.The results show that using a residual network Res Net101 with more layers for the feature extraction network is better than using VGG16 and Res Net50,m AP_{0.5:0.05:0.95}is 10.5%and 5.1%higher than that respectively.Secondly,ROI Align is used to replace ROI Pooling,so as to avoid the pixel loss problem caused by ROI Pooling in two quantization processes.The experimental results show that the Faster R-CNN network using ROI Align is superior to ROI Pooling in detection accuracy,m AP_{0.5:0.05:0.95}up to 0.9102.SSD orchard grape detection method based on Mobile Net.In order to meet the needs of mobile terminal recognition in complex scenes of grape orchards and reduce the dependence on high-performance GPUs,this paper fuses SSD networks with lightweight feature extraction networks,and negative samples（background areas）was much larger than positive samples（grape）,Focal loss was used as the classification loss function.The experimental results show that when using Mobile Net V2 as the feature extraction network,using deep separable convolution,and Focal loss ofγ=2,α=0.75 as the loss function,the model’s m AP_{0.5:0.05:0.95}up to 0.8803,and the single image test speed on the Xiaomi M8 Lite is 100ms,which meets the real-time detection needs of grape orchards.In order to adapt to the multi-scale size detection requirements in grape images,the FPN network and the Mobile Net V2＿SSDLite network are fused.The m AP_{0.5:0.05:0.95}reaches 0.9054,and the single-image test speed on Xiaomi M8 Lite is 153ms.Design and implementation of orchard grape detection system.This paper designs and develops a mobile orchard grape recognition system based on the Android operating system.It is divided into two modes:offline detection and online detection,which realizes orchard grape detection and other functions.The Mobile Net V2＿SSD model was deployed on the Jetson Nano.After using multi-threaded processing tasks,the detection speed of a single image was 43ms,which laid a solid foundation for the subsequent research on picking robots.The test results show that the mobile terminal orchard grape recognition system meets the actual needs of the grape orchard in terms of accuracy and real-time detection speed,and has certain theoretical and practical application value.