Research On Fine-grained Image Classification Based On Deep Learning

Image classification occupies an important position in today’s computer vision field.General target classification has almost no problems in the current era of rapid development of deep learning.Fine-grained image classification has the feature of small differences between classes and large differences within classes.It has become a difficult task in the field of image classification.As an important research content in the field of computer vision,which have great significance to be able to distinguish the fine-grained image in both practical applications of life and the development of artificial intelligence.In recent years,researchers have also conducted a lot of research based on deep learning.Although some research results have been obtained,there is still a lot of space for improvement.In order to continue to optimize the effect of deep learning on fine-grained image classification,based on a large number of research results of image classification,this paper studies and improves the current deep learning image classification algorithms.The specific contributions are as follows:1)Aiming at the problems that fine-grained images has large class differences within classes.Expand the receptive field of the convolution kernel feature descriptor by adding more detailed weight information to the deformable convolutional neural network,which is used to highlight the areas that contribute more to the target classification.2)For the phenomenon of data sets that require a large amount of labeled information for strong supervised learning,in order to reduce the difficulty of data set construction,decrease the computational complexity,and lessen the disturbance to the network that may be caused by manually set labeled boxes.Using active learning capabilities of generative and adversarial networks and excellent image modeling capabilities for target classification images to achieve active learning of image features,by changing the input method of generating the adversarial network to balance the authenticity and diversity of the generated samples,Introduce the idea of image restoration,and use the random input method of the generative adversarial network that combines image feature points and random noise to reduce the training difficulty of the generative and adversarial network.3)Considering that the large number of current deep learning methods makes it difficult to apply in real life,this paper improves on the basis of k-means clustering,adjusts the initial centroid values and quantization values setting method,and based on the idea that image information on the same area with stronger similarity,which also contributes similarly to the target classification.Then the image is divided into little regions to further quantify the parameters.In addition,the advantage of strong generalization ability of extreme learning machines is adopted to solve the issue of low parameter precision.In summary,the various improved methods proposed in this paper start from three perspectives : the image,weakly supervised learning,and network compression,which not only improves the accuracy of fine-grained image classification tasks,but also makes it possible to apply it in real life.Through a large number of experiments on multiple fine-grained image datasets,it is shown that the proposed method has a certain degree of improvement over the current mainstream deep learning image classification methods.At the same time,the detection speed is greatly improved.Thus,it will make some contributions to the research and development of fine-grained image classification.