Fine-grained Recognition Via Pose Alignment And Part Based Representation For Poyang Lake Birds Population

Poyang Lake,which is located in the northern part of Jiangxi Province,is an important freshwater wetland in China.It is a paradise for birds.But with the change of environment,the number of birds is decreasing year by year.Since 1998,in order to better protect and monitor the birds,the Jiangxi Provincial Forestry Department on a regular basis for the development of the annual survey of waterfowl in Poyang Lake,thus showing the importance of the protection of birds in Poyang Lake.However,most of studies are in the number of birds or the kinds of birds,and gathering specimens of each species etc.Due to large amount of human,and time-consuming,it's hard to provide the basis for the national protection of endangered birds in real-time monitoring.Though CUB has 200 kinds of bird,it can't cover all birds.So,we set a Poyang Lake birds' dataset in the first step.For the marked national level and second level protected birds,we combine computer visual technology and propose a classification methods of combining posture alignment and position expression of birds,and provide basis for the realization of automatic monitoring and protection of endangered birds.Image classification was a direction of computer visual,it has two main task,one is traditional image recognition,and another is based on image recognition of fine-grained.In the traditional image recognition task,the difference of the target is very large,such as bicycles,people,trees,airplanes,etc.In contrast,fine-grained image recognition tasks mainly to distinguish sub-class in the categories,such as to distinguish grus-grus and grus-vipio.The difference between the objects it identifies only exists in a few subtle areas,for example,beak or wings.However,these subtle differences is easy covered by background,position and illumination.The general research methods are similar to traditional image recognition methods,they extract whole features for identification,but due to ignore the details,it can't get better results.Some recognition researches extract part feature through the detecting salient regions,but these larger range salient regions can't well express the difference around targets.Many researches have shown that the feature of the target component is very important in the fine grain recognition task.Therefore,in this paper,according to the establishment of the database,we propose a new recognition method for fine-grained object.Firstly using k-means to cluster pose,so that similar pose can be attributed to one class and their difference can well show.Next,according to the components' label,use HOG,LDP,HSV to extract the image components' shape,texture and color features,and combined with the whole image feature to do classification.Experiments show that our method can achieve better classification results.