Design And Implementation Of Image Classification Algorithm Based On Multivariate Unbalanced Small Dataset

People hope that machines can help us automatically recognize and process redundant images in our work and life,thereby improving work efficiency and quality of life.To learn autonomous judgment and expression,machines need to be trained with a large amount of data,and there are certain requirements for the quality of the dataset.In practical research,specialized datasets are usually required for training.However,the datasets required for specialized training inevitably have problems with insufficient or imbalanced data due to the insufficient number of samples available in reality.Small unbalanced data sets have a certain impact on image classification algorithms,such as overfitting,underfitting,etc.,which leads to a decline in the accuracy of the algorithm.In order to better address the challenges posed by small imbalanced datasets to classification algorithms,this article mainly focuses on two innovative works:creating specialized datasets to fill the gaps in existing specialized datasets,which can also help improve the accuracy and robustness of the algorithm;Implementing image classification algorithms for small imbalanced datasets can be applied in the research of small imbalanced image classification,helping to solve relevant practical problems.(1)Create a specialized dataset.This dataset consists of two parts,including the target dataset and the stereo dataset.The target dataset provides color images in order to provide multiple target objects in the campus and help improve the performance of image classification algorithm through training.At the same time,common image classification networks are used for training and confusion matrix is obtained to prove that the target dataset can be used for image classification algorithm training.A three-dimensional dataset provides left and right views to assist in the development of various practical applications on campus.This article compares the depth images obtained independently by the real camera with those obtained by three binocular stereo matching algorithms,proving that the stereo dataset can provide depth information of the image.(2)Design image classification algorithms for small imbalanced datasets.The BNDVGG-19 proposed in this paper selects VGG-19 as the backbone network.At the same time,batch regularization layer,random drop layer,dynamic learning rate update,and early stop algorithm are added.This can not only reduce the fluctuation of the number of features of the model as much as possible to alleviate the problem of overfitting of the model,accelerate the convergence speed of the model,but also improve the generalization ability and flexibility of the network.To verify the performance of the algorithm,two typical small uneven datasets were used for training and validation,and accuracy rates of 0.93 and 0.95 were obtained,respectively.