Fine-grained Image Recognition Based On Object-part Model

Image classification is one of the most important research topics in the filed of computer vision,With the pursuit of more precise target detection in production and life,the problem of fine grained image classification has become a more and more concern in recent years.Fine grained image classification task mainly refers to the fine classification task between different subclasses of the same large class.Compared with coarse grained classification tasks,the main challenge fine grained image classification faced is that the granularity of the category of the image is more finely grained.Different subclasses have very little difference in form,size and angle.Therefore,how to effectively detect foreground images and extract important local information from them is the key to success of fine grained image classification.For fine grained image classification tasks,according to the number of tag information used in the model,it is mainly divided into two categories:"strong supervised learning" and "weak supervised learning".The existing method is mainly to detect objects and components under the condition of component labeling,then to express and classify the features,but the labeling information of these components requires a lot of manpower,and the error rate of manual annotation is also high.This thesis first classifies fine grained images directly with CNN network,but the effect of classification is not very good,a fine grained image classification model based on object-parts is proposed on this basis.The main work of this article is as follows:(1)Object location model.Through a large number of experiments,it is found that the Convolution features in the network model have a greater response to the main objects in the image.For example,some Convolution features have a larger response to the bird’s head position and some of the Convolution features are more responsive to the bird’s body position.But the feature of the single channel Convolution layer is not very good to realize the location of the object.Based on the above findings,This thesis combines the convolution layer features of all channels in the Conv5 layer to construct the object location model and realize the location detection of the object.In the end,image classification is carried out by the acquired image combined with deep neural network.(2)Part selection model.A large number of image blocks can be obtained from the image from the bottom-up algorithm,Then it can be found that the convolution layer features in the network have clustering characteristics by the convolution layer visualization analysis.In this thesis,the part detector of the Convolution layer is constructed by spectral clustering method to screen the image blocks so as to get the component representation of the image.(fine-grained image classification is carried out by combining the obtained parts image with the deep neural network.(3)The above two models can achieve better classification results in fine-grained image recognition,but there is still a gap compared to the existing algorithm effect.So in the end,this thesis build a three flow network model to fine grained image classification by integrating the object location model,the component selection model and the original image feature.The network model can not only extract the overall feature of the image,but also retain the local feature of the image.the experiment proves that compared with the existing fine-grained recognition algorithm,the network model has a better classification effect.This model mainly uses the image level annotation information and has achieved a better classification effect on the CUB200-2011[1]data set.Because no additional manual annotation information is needed,the model has good robustness.This model can be applied to foreground detection,object segmentation and fine grained image classification.