Fine Grained Image Classification Algorithm Based On Deep Learning

Deep learning is one of the important methods to study image classification.As an important subclass of image classification task,fine-grained image classification as its research object is under the same categories of objects of each subclass,has many application scenarios,at the same time,because the subclass class similarity between small and large difference of the same subclass in classification of high difficulty,average depth of convolution neural network to classification task on fine-grained image has excellent classification performance,Aiming at these problems,this paper constructs a weakly supervised fine-grained image classification model,and conducts experiments on the benchmark data set.The specific research contents are as follows:Aiming at the problem that only a single convolutional layer is used for feature extraction and the fine-grained image features are not effectively analyzed,the two-way neural network is used to extract different semantic information of the fine-grained image in parallel,through the polynomial kernel low-bit mapping and then the two-way feature is processed Combine to obtain discriminative features and improve the network model of feature extraction.A basic model for optimizing the residual block is proposed.By strengthening the information flow in the residual block,the network can extract features more effectively.Experiments on different fine-grained data sets have achieved good results.By using the learning rate finder and the learning rate controller,find the best learning rate in Epoch as the initial hyperparameter and do multi-period control and single-period control on the learning rate to make the loss function drop quickly and quickly fit the parameters Space,experiments show that it effectively accelerates the convergence of the model.In order to improve the classification accuracy and reduce the computational complexity,this paper introduces an asymmetric multi-branch network,first uses a 1×1 convolution kernel filter to discriminate the extracted features,and then separates branches and pays attention to local features and global features.Adding different branches using a weighting strategy and using the proposed optimized residual block basic model as the feature extractor of the network.The overall experiment shows that the classification accuracy on different fine-grained image data sets has been significantly improved.