The Research Of Deep Learning Based Flower Recognition Model

In the real world,since the speices of flowers characterized with a great variety,even experts with much experience also with the help of multi-features of color,texture,stem,shapes,and so on,to accurately distinguish their category.A portable equipment with the identification of the species of flower is very useful for a person with little knowledge of flowers.Thus,an accurate model of identification of the species of flowers is very essential for the flower recoginition task.With details,flower recoginition is one of the task to distinguish the species of flowers.Since the species of flowers with a great of variety,it is a hard task to distinguish the species by just using the manual features.Recently,with the development of deep learning technique,it can help us to discovery much more features hidden in the image and then help us to improve the precious of recoginition.Thus,motived by the advantage deep learning technique on feature extraction and the depiction of handcraft features on image detail,we build our flower recoginition model based on the features extracted from the following three types: feature extracted by using deep learning,dual-view features,and multi-modal features.First,considering the effective of feature selection by attention,we proposed a attention-driven flower recoginition model.Here,we use the attention mechanism to select the features extracted by CNN for flower classification.Second,considering the complementary between the RGB and the HSV views.We proposed a dual-stream network based flower recognition model.Firstly,we convert the RGB image into the HSV domain,and then extracted features from both the RGB and the HSV views.Lastly,we fuse these two type of features from RGB and HSV domains for the flower classification.Lastly,since the handcrafted features,such as the color,grayscale,contrast,texture,gradient,some important local part,and so on,can well depict the image detail,we designed an integrated learning framework to effectively fuse those handcrafted features.Finally,we conducted the flower classification by deep learning technique by using the fused features.Experiments showed that our proposed model is feasible and effective.Above all,our research of flower recoginition model are studied from the following three aspects: the attention-driven feature selection,the complementary among different color space,the depiction of handcrafted feature on image detail.Based on different aspects,we proposed three flower recoginition model,each of which can effectively improved the performance on flower classification.