Object Recognition Of Fragrant Pear In Natural Environment Based On Deep Learning

Now,artificial intelligence technology has become more and more mature,and its research results have been applied to various fields.If artificial intelligence technology is applied to agriculture,it can not only greatly improve the agricultural development,but also liberate the hands of farmers.There are many disadvantages in the traditional manual picking technology,such as:(1)The manual picking brings high cost;(2)It is possible to bring safety problems to farm workers in the picking work;(3)High time cost in the picking work.Therefore,in the era of artificial intelligence,it is imperative to research the picking machines.Because the traditional fruit picking machine mainly uses the basic characteristics of fruit such as color and shape,therefore,the traditional fruit object recognition methods mainly include the gray threshold method,the shape feature location method and the color feature extraction method.For the traditional fruit recognition method,it has poor recognition effect for the occluded fruit and the fruit color is similar to the leaf color.Therefore,experts and scholars from home and abroad are generally use deep learning method to solve the problem of fruit object recognition in recent years.For the pear picking industry,because the color of the fruit is similar to that of the leaves when the pear is ripe,accordingly,the object recognition method based on the color feature is not very good for the pear recognition.Secondly,due to the serious occlusion between fragrant pears and between fragrant pears and leaves,therefore,the result of fruit recognition based on shape feature is not ideal for the fruit with serious occlusion.finally,because of the crisp texture of fragrant pear,it is easy to cause fruit damage when picking,so improving the accuracy of the target contour of fragrant pear is the key to improve the working quality of fragrant pear picker.Based on the above,this paper proposes a deep learning model Mask R-CNN method based on transfer learning to identify the fragrant pear target.First of all,the network is preliminarily trained through the data set of Kaggle and 9600 Fruit Pictures downloaded on the Internet,then the data set of mature fragrant pear in the natural environment is trained based on the pre trained network,and finally the network model of accurate recognition of fragrant pear target is obtained.The accuracy and recall rate are used to evaluate the pictures whose target is three meters away.Because of the distance problem,the contour of the pear fruit is not easy to be accurately identified,so the target within three meters is evaluated by the segmentation accuracy.The experimental results show that the average segmentation accuracy of the image is98.02% on the basis of using the fruit image data set for pre training without adding noise to the same target,which is 4.30% higher than using the coco data set for pre training(the accuracy is 93.72%).In addition,the model also has a good recognition effect for the occluded fruit.Without adding noise,the average segmentation accuracy of the occluded fruit is 95.28%,which is only 2.74% higher than that of the occluded fruit.Therefore,the model has a good robustness for pear recognition and segmentation.