Research On Fruit Recognition Technology Based On Multi-scale Convolutional Neural Network

As an important part of computer vision task,the development of image recognition has been concerned by scholars at home and abroad.The fruit image recognition technology can promote the development of modern smart agriculture and smart city,so it has high research value.At present,the commonly used fruit recognition methods can not meet the needs of fast and accurate recognition in production and life,so it is necessary to study a new fruit image recognition technology.With the rapid development of deep learning technology in recent years,convolutional neural network(CNN)has performed extremely well in the field of image recognition.Therefore,based on the in-depth study of CNN theory,this paper applies the classical CNN model to the fruit image recognition,and optimizes and improves the CNN model according to the actual problems,so as to improve the performance of fruit recognition.The main work of this paper is as follows:(1)This paper systematically analyzes the basic process and algorithm of traditional fruit image classification and recognition methods,and uses small-scale fruit image set to test these methods.The experimental results show that the recognition effect of traditional fruit image recognition methods based on machine learning algorithm is at a low level,and CNN is needed for fruit image recognition experiments.(2)In order to meet the data requirements of CNN model training,this paper uses data enhancement technology and tfrecord structure to expand and make fruit image data set,which is used to accelerate data reading and model training.After that,VGG16 is fine tuned and trained by using transfer learning method.The accuracy rate of fruit recognition is 97.17%,which is much higher than the traditional fruit recognition method.(3)Aiming at the problems of VGG16 model file is too large and the recognition accuracy is not high enough,this paper designs a smaller scale improved CNN model referring to VGG structure.The experimental results show that the accuracy rate of the improved CNN model is 98.13%,which is higher than VGG16 model,and the model parameters are less.(4)Aiming at the problems of traditional CNN such as VGG,such as large parameters,slow convergence speed and long training time,this paper designs a multi-scale CNN model,and improves and optimizes the model from three aspects: activation function,weight initialization method and optimization function.The experimental results show that the recognition accuracy of the fruit recognition method based on multi-scale CNN and its improved model is 98.27% and 98.81% respectively,which is further improved compared with the traditional CNN.Through the analysis and comparison of various experimental results,it can be seen that the effect of the fruit recognition method based on CNN is far better than the traditional method,but there are still some shortcomings.In view of these shortcomings,the multi-scale CNN designed in this paper not only improves the accuracy of fruit recognition,but also greatly improves the convergence speed of the model,showing the strong performance and superiority of fruit recognition technology based on multi-scale CNN.