Fine-grained Image Classification Model Based On Improved Light-CNNs

The purpose of the fine-grained image classification task is to distinguish the subcategories under the same category.Because different subcategories have highly similar visual content,feature differentiation is not obvious,so how to find the discriminant feature region is the difficulty of this task.At present,in the fine-grained image classification model based on deep learning,it also develops from the shallow level to the deeper level model,which has a large number of parameters and large model memory.The lightweight convolutional neural network proposed by quantization and weight pruning of the traditional convolutional network can solve the efficiency and parameter problems of the deeper convolutional neural network.So this paper mainly studies using lightweight convolution neural network for fine-grained image classification task,the traditional granular model running efficiency and the problem of the large amount of calculation,and focused on the difficulty of fine-grained image classification for existing lightweight convolution neural network was improved,the final design out the model of the recognition rate can be better than the existing fine-grained classification model,and fewer,prediction speed,more easy to deploy.The work completed in this paper mainly includes:1.Lightweight convolution neural network was adopted instead of existing finegrained model such as VGG,Res Net such calculations too much,a lot of occupy the main classification of memory network,and based on the idea of transfer learning,three preliminary training lightweight convolution neural network Squeeze Net,Shuffle Net,Mobile Net,through the network for training mode were obtained by these three lightweight convolution neural network is given priority to dry the fine-grained image classification model of the network.The accuracy,number of parameters,memory size and prediction speed of the three classification models on the test set were compared.It is concluded that the fine grained image classification model based on lightweight network can better solve the problems of computation and operation efficiency of the existing model.2.A lightweight fine-grained image classification model based on multi-scale fusion and joint training is proposed to improve the recognition rate of lightweight networks.According to the response characteristics of the high and low layers of the convolutional neural network to the recognition object,the characteristics of the bottom layer and the middle layer of the network were pyramid-fused,and the classifier and loss function were added to each layer for joint training.Finally,the accuracy rate of Squeeze Net improved by this method increased by 4.39% on the CUB bird dataset,5% on the FGVC-Aircraft dataset,and 92.9% on the Stanford Cars dataset,which was higher than some excellent fine-granular algorithms currently.The accuracy of Shuffle Net improved by this method on the above three fine-grained data sets is also improved by 5.2%,0.77% and 4%respectively.3.From the perspective of attention mechanism,proposed to CBAM,SE two attention mechanism module in three different ways is embedded into the Fire of the Squeeze Net module,and the improved network for bilinear fusion,through experiment contrast and visual analysis,the improved Squeeze Net in CUB birds,FGVC-Aircraft Aircraft,Stanford Cars these three fine-grained data set on the highest accuracy of 86.79%,89.41%,94.30%.4.Finally,the model proposed in this paper is compared with the current excellent fine-grained image classification models.Finally,a lightweight fine-grained image classification model with better recognition rate is designed,or is equivalent to the existing fine-grained image classification models.In terms of model performance,the number of parameters of the proposed model is less than 5M,the inference time of a single image reaches millisecond level,and the model memory has more advantages than the existing fine-grained models.