A Domain Adaptive Image Classification Method Based on Hybrid Pooling

Abstract

The invention discloses a domain adaptive image classification method based on hybrid pooling, which sends the target domain image to be classified into the trained image classification prediction model to output an n×1-dimensional feature vector, and then uses one-hot encoding Obtain the category of the target domain image. The image classification prediction model includes several convolutional layers connected in sequence. The convolutional layer is connected to the maximum pooling layer, and then a layer of average pooling layer is cascaded. The average pooling layer is connected with a softmax activation function. The fully connected layer, the target domain image is extracted through several convolutional layers to extract image features, then down-sampled through the maximum pooling layer to obtain the first descriptor feature, and then the local information in the image feature is extracted through the average pooling layer to obtain the second descriptor feature, and finally the feature vector is obtained by the fully connected layer. The method of the invention can tolerate small changes in the input, reduce overfitting, improve the fault tolerance of the model, and optimize the migration effect.

Description

technical field [0001] The invention relates to an image classification method, in particular to a hybrid pooling-based domain adaptive image classification method. Background technique [0002] Domain adaptation is a subclass of transfer learning. How to use a small amount of labeled data and data in other related fields to build a reliable model to predict target fields with different data distributions is the research content of transfer learning. The main goal of domain adaptation is to find common features as much as possible, so as to minimize the difference in data distribution between the two domains, so as to realize knowledge transfer. The target domain is a data set with few or no labeled samples, which is the domain to be learned. The source domain is a dataset that has a different distribution than the target domain data but contains a large number of similarly labeled samples. [0003] Image is now a very important information carrier. Image classification ca...

AI Technical Summary

Problems solved by technology

This is because in different fields, the lighting conditions are different, and the viewing angles of objects are also different, which makes it difficult to obtain the common features of the two fields.

In addition, traditional domain adaptive methods can only extract the underlying features of objects, and cannot effectively identify and classify objects in the target domain.

[0007] 2. The utilization rate of local information is not high

At the same time, it is also easy to lose a lot of important information, deepen the degree of network overfitting, lead to insufficient abstraction of extracted features, and cannot produce greater tolerance to small changes in input

Images