Research On Flower Image Classification Algorithms Based On Deep Learning

Flower classification is the most basic work in the field of botany,which is mainly based on the morphological characteristics of flowers.There are 369,000 known flowers species,which makes flowers one of the most prosperous organisms in the world.Therefore,the classification process expends much time and energy for plant taxonomists.With the development of digital imaging technology,it is very easy for people to shoot clear flower images through mobile phones,cameras and other devices.So classifying flower automatically based on images is a research hot spot.In contrast to coarse-grained image classification,flower image classification is a fine-grained classification problem.The difficulty lies in the great similarity between different flowers,and there is great difference between the same flowers.Aiming at the above difficulties,based on the in-depth analysis of domestic and overseas research results,this paper proposes two flower image classification algorithms based on deep learning.Our main research contents can be summarized as follows:1.Several methods of flower image classification are summarized,including methods based on manual features and methods based on deep learning.And the advantages and disadvantages of the above methods are analyzed.2.For the problem that the convolution neural network is difficult to realize the optimization of network's parameters and is easy to appear the over-fitting phenomenon in the case of small sample datasets,in this paper,a novel algorithm based on saliency map and principal component analysis network(PCANet)is proposed.Our algorithm consists of two phases: flower region extraction and flower feature learning.In the first phase,the saliency map is utilized to extract region of flower.Exact region of flower can reduce the influence of background and highlights the contribution of image details.In the second phase,the extracted flower regions are input into PCANet to learn flower features automatically.And softmax classifier is used to estimate the category of flower.Experimental results based on Oxford 17 flowers dataset show that the proposed method can obtain better classification accuracy under the condition of small sample datasets.3.For the problem that the lack of flower images' annotations and the high cost of labeling make fine-grained image classification methods based on the deep learning unable to locate the target area well in the case of large sample datasets,anunsupervised flower image classification method based on selective convolution descriptor aggregation is proposed.Firstly,a normalization method based on keeping the aspect ratio is used for normalizing the size of flower images.Then we utilize the VGG-16 model pre-trained by ImageNet to obtain the flower image's feature and locate the significant region based on the high activation in the feature map,so keep useful convolution features according to the location information.Then all selected convolution features are aggregated to form a low dimensional feature vector.Lastly,the feature vector is entered into the softmax layer for classification.The method can locate significant region without supervision and select useful feature of flower image.Experimental results based on Oxford 102 flowers dataset show that the proposed method can obtain better classification accuracy under the condition of large sample datasets.