Research And Implementation On Flower Image Recognition Based On Deep Learning

As an important branch of machine learning,deep learning has achieved some good results in the classification,positioning,detection and segmentation of computer vision in these ten years.The flower images is typical fine-grained images with many complicated features,which are difficult to extract and identify.Identifying flower images in the traditional way is labor intensive and ineffective.This paper uses the deep learning method to study the classification and detection of flower images.Through the experimental verification of the existing algorithms,we analyzed the advantages and disadvantages in detail,and proposed the improved algorithms.The main research contents of the thesis are as follows:(1)We studied the basic structure of convolutional neural networks,dimension reduction and feature extraction,and the weights sharing and sparse connectivity.And then the back propagation process was analyzed and deduced in detail.At the same time,we analyzed and compared a variety of different optimization methods in detail.(2)The existing common image classification algorithms based on convolutional neural networks were studied and compared experimentally.Aiming at the problem of parameter redundancy and destruction of images space structure caused by the fully connected layer,an effective improvement using spatial pyramid pooling is proposed.Not only improves the test accuracy by 2.14%,but also the training time is reduced to 22.6%.Although the cross entropy loss is a loss between classes,there is no requirement for the distance between the vector representations of each class in essentially.Therefore,in order to improve the distance loss between the class features,the triplet loss and the improved center loss function are used to improve the accuracy,and experiment shows,the final training accuracy is improved by 3%.(3)The YOLO algorithm is used to detect the flower images.Firstly the YOLO algorithm was introduced,analyzing the advantages and disadvantages of it,and then we improved the algorithm by multi-network cascade optimization method,changing the original single network to a three-level serial network.This not only improves the mAP of detection by 5.2,but also has a good detection effect for smaller flower targets.