Research On Fruit And Vegetable Image Segmentation And Feature Recognition Based On Deep Learning

China is a major producer of natural products and vegetables in the world,but due to the peculiarities of the plants and the complexity of the growth and development environment,the plants rely mainly on manual labor for harvest.With progressive urbanization,the steady rise in labor costs and the sharp decline in the rural workforce,the natural product vegetable industry has led to a labor shortage.Natural fruit and vegetable picking robots can replace manual labor so that the problem can be completely solved by fewer workers and low work efficiency,and they can provide solid professional support in the design and manufacture of fruit and vegetable picking robots.Based on the traditional Mask-RCNN algorithm,this thesis improves on the algorithm to optimize detection accuracy and uses the Darknet-53 Convolutional Neural Network method to detect the image features of fruits and vegetables.The main work of this thesis is as follows:To begin with,this thesis proposes a method for segmenting fruits and vegetables based on the RCNN mask.This thesis first introduces the idea and structure of the mask RCNN algorithm for image segmentation.Using the fruit and vegetable data set,a small image segmentation data set is created for fruit and vegetables.Using the data set to train the mask RCNN model,an image segmentation method for fruits and vegetables is obtained.This method segments the entire fruit and vegetable image.The results show that the mask RCNN algorithm has the problem of inaccurate segmentation in the image segmentation of fruits and vegetables.Second,this thesis proposes a multi-stage,function-fusion-based,mask RCNN method for segmenting fruits and vegetables.In view of the shortcomings of the original mask RCNN algorithm,the multi-stage feature fusion process is introduced to optimize the loss function,and the full link enhancement process is used to segment the fruit and vegetable images.In order to ensure sufficient training of the model,the method of data improvement is also used.The experimental results show that the improved mask RCNN method,which is based on a multi-stage feature fusion,has significantly improved the effect of image segmentation of fruits and vegetables.Finally,this thesis proposes a method for recognizing image features for fruits and vegetables based on Darknet’s Convolutional Neural Network.In accurately recognizing images using convolutional neural networks,it has been found that a good feature extraction network determines the accuracy of image classification and target recognition.Therefore,this thesis selects the Darknet-53 convolutional neural network framework with better classification and detection performance and replaces the BN normalization method used in the original network with the filter response normalization,under the condition of ensuring accuracy,reduce hardware requirements of network computing for devices and building the Darknet-53-FRN network.By optimizing the model structure and the parameters of the feature extraction network,we researched and contributed to the development of a fruit and vegetable image feature extraction network based on the folding neural network Darknet-53 in order to ensure an accurate recognition of fruit and vegetable images.