| With the progress of science and technology,our life has gradually become intelligent,"unmanned supermarket" is becoming more and more popular.When users buy fruit and vegetable goods in "unmanned supermarket",it is necessary to automatically detect and classify fruits and vegetables.But the popular bar code scanning method requires human participation,which is not only inefficient but also a waste of manpower and material resources.The traditional method of object detection usually needs to traverse the whole image,and the accuracy is not high.Although the current methods have improved in accuracy,the actual availability is not high because of the huge models.Therefore,the detection and classification algorithms of fruits and vegetables need to be improved in terms of accuracy,speed and availability.This thesis mainly studies the technology of fruits and vegetables detection and classification based on deep learning.In order to solve the problem of low accuracy and poor usability,we present a new method of fruit and vegetable detection and classification based on improved Faster R-CNN.In this method,we design a cascade region proposal network to improve the position accuracy of the proposal region.Then we concate the context information around the candidate area to improve the classification accuracy when extracting the feature of proposal regions.In addition,we use the idea of MobileNet to compress the neural network model with a depth separation convolution approach into the original 1/10,increase the availability of the model.At the same time,we build a rich fruit and vegetable dataset,and use this dataset to train a fruit and vegetable detection and classification neural network model.Finally,based on the above method,we implement a system of fruit and vegetable detection and classification on Android mobile terminal,and test the system in practical scenarios.The experimental results show that the system has high accuracy,speed and compatibility. |
