Low Resolution Image Classification Based On Convolutional Neural Network With Attention

In recent years,with the rapid development of deep learning,methods based on deep neural networks have been widely used in computer vision and have demonstrated performance superior to all other methods.In some tasks such as image classification or recognition,models based on deep neural networks can even outperform human performance.Although deep models are achieving increasingly impressive results,these results are often generated from high-quality image data.Such high-quality data may be easy to obtain in certain scenarios,such as facial recognition or medical diagnosis,but in other practical applications,we may only be able to obtain low-resolution images due to factors such as insufficient device performance,long distance,or poor imaging environment.These low-resolution images are characterized by a low resolution and noise,which results in a significant challenge for image classification tasks in low-resolution scenes.To address the problems posed by low-resolution images,this paper focuses on the following two issues:1.To address the problem of low feature points in low-resolution images,this paper proposes an improvement to the original random data augmentation algorithm.The improved algorithm dynamically adjusts hyperparameters by evaluating the accuracy or loss value of the current round of the dataset to determine whether to adjust data augmentation and how to adjust it.This algorithm enables data processing to be dynamically performed during model training,thereby improving the quality of training data and enabling models to be trained on more diverse data distributions.2.To address the problem of complex model overfitting in low-resolution images,this paper proposes a neural network model equipped with a global multidimensional attention mechanism(Global-Multidimensional-Attentionmechanism Neural Network,GMANet)based on residual networks and dense connection networks.The model extracts features from local and global image samples through the connection between the basic network and the self-attention network layer,and finally obtains multiple output results after different network blocks,calculating the model-weighted cross-entropy loss function according to a certain weight.The model’s feature extraction considers both local and global features,allowing the model to learn more useful features and alleviating the problem of model overfitting.The GMANet model using the random data augmentation algorithm achieved Top1 accuracy rates of 98.17%,94.67%,and 81.08% on the publicly available datasets SVHN,CIFAR10,and CIFAR100,respectively.Compared to the highest Top1 accuracy rates of the three baseline networks Res Net50,Wide Res Net28,and Dense Net121,the results showed an improvement of 0.92%,5.17%,and 5.09%,respectively.This result demonstrates that our algorithm and model can effectively improve the classification performance of low-resolution images.In summary,the research in this paper aims to provide new ideas and methods for addressing the problems posed by low-resolution images,which can cope with more diverse data distributions while improving model generalization ability.We believe that these achievements can provide some feasible solutions for similar problems.