Fruit Detection Based On Convolution Neural Network

With the continuous development of society,machine vision has already got application and achieved good results in some areas.At present,the various stages of the fruit product are continuously into industrialization.But the process of selling fruit still relies on manual weighing in the supermarket retail.That way is wastes human and the efficiency is not high.Once encountered that case which a large number of customers need to weigh,it is very easy to cause many people in line and inconveniences customers.So intelligent fruit sold will inevitably become the development trend of the future.This paper puts forward a kind of fruit detection algorithm based on convolution neural network.This algorithm can be used to implement intelligent weighing of fruit.In this paper,the research contents and contributions are as follows:(1)Establish a fruit image library which can train a convolution neural network effectively.And the image library contains a single fruit,multiple same fruits,multiple mixed fruit three types,which meets the actual situation.Home and abroad don't have the open firuit image library now.So the fruit image library has important research significance.(2)Selective search combined with entropy analysis of segmentation algorithm is proposed to deal with the detection and identification of multiple fruit images and multiple combinations of fruit images.The segmentation algorithm decomposes complex fruit maps into multiple target regions which can extract regions with valid information and improved ultimate recognition accuracy.(3)A convolution neural network architecture that is corresponding to multiple target regions achieved by the image segmentation with complex fruit is constructed.Using the regions to train the convolution neural network can both meet the requirements of large data of convolution neural network and increase the diversity of training image samples and improve the ability of trained convolutional neural network identification.(4)In the process that using convolution neural network architecture is proposed in this paper to detect fruit image,the redundant location boxes is eliminated by non-maximum suppression.Finally we get the fruit accurate localization.Experimental results show that the algorithm proposed in this paper on the established database has achieved 91.6%detection rate.Compared with the simple use of CNN for testing,the accuracy rate improved greatly and has a certain application value.