An Image Classification Method Including Online Few-Sample Incentives

Abstract

The invention belongs to the field of image processing, and discloses an image classification method including online small sample excitation. The purpose of the present invention is to enable the neural network used for image classification to have the ability to "learn to learn" in a constantly changing environment, and to stimulate the accuracy of classification according to the reference pictures received online. Based on the convolutional neural network, the invention adds neural adjustment parameters to the frame of the plastic neural network to make the whole network closer to the biological neural network, and adopts the principle of online and real-time in the selection of reference pictures in the prediction stage. The invention has the advantages of high flexibility and high accuracy in the task of image classification, and can adjust the output result according to the environment in real time.

Description

technical field [0001] The invention belongs to the field of image processing, and relates to a method for image classification based on an artificial neural network trained with small samples, in particular to an image classification method including online small sample excitation. Background technique [0002] Image classification refers to the process of analyzing an unfamiliar test image through a trained computer and finding out what category the content belongs to. It has a wide range of application requirements in the fields of computer vision and machine learning. Deep neural network is a very popular and effective image classification method. Through a multi-layer convolutional neural network, the original features of the image are abstracted into feature vectors that can be linearly divided, and then the feature vectors are processed by a fully connected layer. Linear combination, and finally get the probability that this picture belongs to each category. [0003]...

AI Technical Summary

Problems solved by technology

However, this model will be difficult to produce results in the face of the following more flexible and intelligent tasks: (1) The entire image classification task is divided into several stages (life cycle), and the categories of images in each stage may belong to different categories , for example, the first life cycle picture to be classified may belong to one of the three categories of apple, cat, and boat, while the second life cycle picture to be classified may belong to one of the three categories of tree, truck, and spider, and so on ; (2) The categories in the task of the entire image classification are fixed, but as the classification task progresses, the true distribution of the images to be classified belonging to each category changes; (3) The categories in the entire image classification task are fixed , but in the process of classification tasks, new training pictures with classification labels that did not appear in the original training set can be continuously obtained, and if the features in these new pictures are learned, it will be useful for the next short-term future The classification ability of classifying pictures will be greatly improved

For the network trained by the traditional image classification method, after the entire network is trained (this training often requires more than one million training data and more than ten hours to several weeks of training time), the network structure, parameters, including the output probability vector The category of each representative in has been fixed, and in a situation similar to the one mentioned above, it cannot adapt to the classification task according to the change of the environment

Images