Research On Fine-grained Image Classification Based On Improved Neural Network

Fine-grained image classification technology,as one of the most important research contents in the field of machine vision,focuses on realizing more detailed classification between species.In contrast,the goal of coarse-grained classification is to distinguish different species.The goal of fine-grained classification is to improve the granularity of classification among subcategories.With the increasing demand for species classification,traditional classification methods can no longer meet the demand.Therefore,the research focus gradually shifted to how to improve the accuracy of fine-grained classification.Fine-grained image classification task has the phenomenon of small difference between classes,large difference within classes,and not distinct distinguishing characteristics of objects to be recognized.In order to focus on solving the problem that the prediction accuracy of fine-grained image recognition model is not high enough,this paper studies how to improve the model structure,improve the feature extraction framework,obtain the important feature information of the object to be recognized,and effectively improve the classification accuracy.Two methods based on traditional deep convolutional neural network are proposed.The main improvements of the model are as follows:Method 1: Based on VGG-16 convolutional neural network model,the new model is named New VGG-16.The optimization mainly includes the use of the improved sort?pool2d pooling layer.Add the BN layer after the convolution layer.Replace the original full connection layer with a convolution layer.The model architecture built based on Tensor Flow and combined with the transfer learning and training mode can train the model with better effect under the condition of selecting small batch data sets,and improve the generalization performance of the model in fine-grained image classification task.The experiment was based on 10 rare monkey images as a data set.Compared with the traditional model,the model can effectively improve the accuracy and applicability of image classification,and the accuracy can reach 95.75%.Method 2: Based on Bilinear CNN model,the convolutional part of Dense Net121 with compact structure was selected as feature extraction module.Using the improved Relu-and-Soft Plus activation function.Combined with the attention mechanism,the spatial and channel attention modules are introduced to extract multidimensional detailed feature information of objects.Add a convolution layer to adjust the dimension of feature graph.Grouping strategy was adopted to reduce the number of model parameters.After bilinear pooling,the global maximum pooling layer is used to process bilinear feature vectors and obtain significant features.The model was mainly based on five flower data sets samples,and was matched with Flavia plant data sets for training and testing.Experimental results show that compared with the original model,the training efficiency is improved,and the classification accuracy can be improved to 96.869%.For the same monkey data set,the prediction accuracy of this model is better than that of the improved model New VGG-16,which further improves the ability of processing fine-grained classification tasks.Combined with the animal and plant data sets selected by this study,it can provide important support for the ecological protection of rare species.Further promote the development of visual recognition technology,improve the accuracy of fine-grained image classification tasks,and promote artificial intelligence technology into an important stage of in-depth adjustment,speed innovation and cross-border fusion.