A Fine-grained Image Recognition Method Based On Convolutional Neural Network

Image recognition is one of the core research direction in the field of computer vision,which can be divided into general image recognition and fine-grained image recognition.General image recognition refers to the recognition of types of objects such as cars,people and dogs,while fine-grained image recognition refers to the recognition of different subcategories of the same large category,such as the different types of birds and different types of cars.Compared with general images,fine-grained images are difficult to be recognized due to data characteristics such as small inter-class variance,large inter-class variance and a small percentage of targets in images.Although the current fine-grained image recognition algorithms have achieved a certain recognition accuracy,the existing algorithms only use the whole image features for recognition,and the image features contain a large number of background features and ignore the shallow distinguishable features such as color and shape,which results in low recognition accuracy.To solve these problems,this paper proposes a fine-grained image recognition method based on convolutional neural network,and makes detailed analysis and experiments.The research contents of this paper include the following points.Firstly,this paper proposes a fine-grained image recognition based on feature fusion to solve the problem that the existing methods ignore the effect of shallow features on fine-grained image recognition.By analyzing the import role of shallow features in fine-grained image recognition,an adaptive weight feature fusion method is proposed.The shallow and deep features are fused by this method,then are classified.Adaptive weights can give full play to the role of shallow features in recognition.In addition,the proposed method does not change the number of parameters and is highly transferable.Experimental results verify that the proposed method is effective.Secondly,this paper proposes a fine-grained image recognition method based on multi-classification loss function aiming at the problem that existing algorithms only use the whole image features to classify,which contain a lot of background noise information.By analyzing the disadvantages of image features recognition and the advantages of multi-level(image-level,object-level and pixel-level)feature recognition,using multi-level features to design multi-class loss function for multi-grain recognition.The proposed method performs weakly supervised learning only under the image category label information,prompts the network to pay attention to and learn multi-level information,obtain distinguished fine-grained features,and improve recognition accuracy.Finally,this paper fuses and optimizes the feature fusion recognition method and the method based on multi-class loss function,and proposes a fine-grained image recognition method based on multi-level fusion learning.This method uses adaptive weights to fuse different feature information of the target area.Experimental results verify that the proposed method further improves the accuracy of fine-grained image recognition.