Research On Fruit And Vegetable Image Recognition Method Based On Deep Learning

Computer vision technology is an extension of biological vision system,combined with image processing,signal processing,computer information processing,neural network and other basic technology and continuous development of computer technology.Computer vision technology is widely used in various fields,including agriculture,with good results and great potential.In the existing research on agricultural products using computer vision,fruit and vegetable image recognition is an important part of agricultural products research,and is the key technology to realize the automatic classification of fruits and vegetables agricultural products.The main challenge of fruit and vegetable image recognition is that there are many kinds of fruits and vegetables,lack of a large number of labeled data,and it is difficult to achieve fruit and vegetable image classification by supervised learning method.In view of the above problems,the fruit and vegetable image data set is constructed to classify the fruit and vegetable image.The main conclusions are as follows:1)Data fusion under various collection methods can break the information barrier between different recognition systems,establish a new recognition model,and ensure the comprehensiveness,complementarity and accuracy of fruit and vegetable data classification.In this paper,fruit recognize,a multi view open data set of fruit and vegetable images is established by fusing kaggle high-definition Logitech webcam,and web crawler technology is used to obtain the fruit and vegetable image processed by computer,and the original fruit and vegetable image data set is obtained n order to solve the problem of data imbalance in the fusion network crawling data and fruit and vegetable photo data,random under sampling method is used to achieve sample balance and expand the data set.The experimental data set F&V is obtained by removing the edge data so as to retain as many sample features as possible.2)Aiming at the difficulty of feature extraction in traditional machine learning methods,this paper uses deep convolution neural network to map digital image information to label output directly,and realizes end-to-end learning.The global pooling technology is used to replace the fully connected network,which reduces the number of fully connected network features,thus simplifying the scale and complexity of the network,which is helpful to improve the training efficiency of the network and the estimation efficiency of the model.From the perspective of selecting optimizer,three kinds of optimizers(SGD,SGD momentum,Adam)are used to get the optimal solution in the learning process to reduce the gradient and finally complete the training.The experimental results show that Adam training can achieve the fastest convergence by iteratively updating the network weights through adaptive learning rate.3)In this paper,a method of fruit and vegetable image classification with specific migration mode is proposed,and the recognition models of FVM1,FVM2,FVM2-1,FVM2-2,FVM2-3 and FVM3 are established.Six models were tested on the F&V test data set.The accuracy rate of positive samples,recall rate of positive samples,average accuracy rate,F1-score and kappa coefficient were selected to evaluate the established fruit and vegetable recognition model.The results show that compared with the direct migration method,the specific method can achieve higher image recognition accuracy,and the improvement is more than 2%.The FVM2-2 optimization based on Inception-V3 has the best effect,and its average accuracy rate has increased by 7.33%,and the average accuracy rate of FVM3 based on Resnet50 has increased by 3.78%.The results show that training the network with specific migration mode can improve the ability of feature expression and generalization of the network.