Research On Few-shot Image Classification And Its Applications Based On Multi-scale Information

In recent years,deep learning methods have achieved great success in the fields of computer vision,which can provide new ideas for the research of disaster prediction methods.However,due to the uncertainty and danger of natural disasters,it is extremely difficult to obtain disaster images.In addition,the amount of parameters of a deep learning model is generally large,and a large amount of data is usually required to train the model.If only a small amount of data is used to train the model,overfitting will often occur and the performance will be unsatisfactory.Therefore,how to use a small number of samples to train a model with acceptable performance,namely few shot learning,has become a topic that needs urgent research.Metric learning is one of the main research directions in the field of few shot learning,but the current algorithms based on metric learning such as prototypical networks and relation networks have the following two deficiencies: First,the algorithms only use cosine similarity or 3×3 convolution to calculate the similarity between samples,which may result in a lower classification accuracy when the targets are in different sizes;second,the metric learning algorithms that care about the category of the object in the image are difficult to solve the problem of disaster image classification that pays more attention to the disaster state of the object in the image.Aiming at the first shortcoming,this paper deeply researches the metric learning algorithm suitable for comparing objects of different sizes,and proposes a few shot object image classification model based on multi-scale comparison; on this basis,for the second shortcoming,in the specific application scenarios of disaster image classification with a small number of samples,a few shot disaster image classification model based on multi-scale comparison and attention mechanism is proposed.The specific work of this paper is as follows:1.In response to the large size gap of the objects in the images,by systematically exploring the impact of multi-scale information on the model effect,a few shot object image classification model based on multi-scale comparison is proposed,referred to as multi-scale comparison network.The multi-scale comparison network first uses a4-layer encoder to encode support samples and testing samples to obtain their feature maps,then deep splices the feature maps,and finally uses a comparator comprising two layers of multi-scale comparative modules and two fully connected layers to derive the similarity between support samples and testing samples,and according to the similarity the categories of testing samples are obtained.Experimental results on two benchmark data sets ——Omniglot and Mini-Imagenet prove the effectiveness of the multi-scale comparison network,and compared with the relational network,the multiscale comparison network has an accuracy improvement of about 2% in all the indicators of all experiments on the Mini-Imagenet dataset.Finally,we explored the impact of multi-scale information,including multi-scale feature extraction,multi-scale feature comparison,the number of scales and the increase in parameters on classification accuracy.2.For the existing few shot image classification algorithm that cares about the category of the object in the image,it is not applicable to the disaster image classifica-tion that pays more attention to the disaster that occurs in the image,a few shot disaster image classification model based on multi-scale comparison and attention mechanism is proposed.The model first uses a feature extraction network with the residual structure to prevent network degradation to extract features from the support set and test set images,and then deep splices the feature maps,sends them to the attention mechanism network that can weaken the information of irrelevant objects to get the attention weight corresponding to the feature maps,then multiply the attention weight and the feature maps in the plane dimension by element to obtain the optimized feature maps,and finally the feature maps are sent to a disaster comparison network that pays more attention to the pixel-level comparison of object states to obtain the similarity of the two images,and then the category of the disaster is obtained.Combined with practical applications,we built a few shot disaster image data set and conducted experiments on it,which proved the superiority of the small sample disaster image classification model.Compared with relation networks and multi-scale comparison networks,the accuracy of the model is improved.Finally,the scale quantity selection experiment and the ablation experiment are carried out to further explore the role of each module in the few shot disaster image classification model.