Research On Fine-Grained Image Classification Based On Deep Neural Network

Fine-grained image classification has always been an important research topic in the field of computer vision.There is a strong research demand in both industry and academia.With the rapid development of deep neural network technology and the emergence of large-scale training data sets,common image recognition tasks have achieved great success.However fine-grained classification recognition still has a lot of room for improvement.The difficulty of fine-grained classification mainly lies in the small difference between classes,the large difference within classes and the non-prominent target objects.So how to accurately extract the location of the target objects in the image and the distinguishing features of the target objects is the key to solve the task of fine-grained classification.Although the traditional deep convolution neural network can effectively mine the basic contour,texture and other feature information of the image,it is not enough to distinguish the fine-grained image.Bilinear convolutional neural network is a model to solve the problem of weak supervision fine-grained classification,but there are many problems,such as too many parameters,low prediction accuracy and so on.Aiming at the above two problems,this thesis puts forward corresponding improvement methods.The main work of this thesis is as follows:(1)In order to solve the problem that there are too many parameters in bilinear convolution network,which is not conducive to model training and deployment,this thesis proposes a bilinear convolution model based on feature grouping.The experimental results show that the model can effectively reduce the parameters to 1 / 3 of the original model at the expense of a little precision.(2)By investigating the application of attention mechanism in image field,a new model improvement scheme is proposed by combining bilinear convolution model with two different attention modules.The improved model can effectively extract the location information and key distinguishing features of the target in the fine-grained image.The experimental results show that the improved model based on attention mechanism greatly improves the prediction accuracy of the model.(3)An industrial data set is constructed,and the task of automobile brand recognition is realized by using the improved two models.The prediction accuracy is more than 90%.The performance of the fusion model is compared and analyzed.In this thesis,two kinds of weakly supervised fine-grained classification models are proposed,which are compared and analyzed on three open datasets.The practicability and effectiveness of the improved model are proved.In addition,this thesis constructs a high-quality vehicle brand data set,using the improved model to achieve the task of vehicle brand recognition.