Image recognition method and device, storage medium and equipment

Abstract

The invention discloses an image recognition method and device, a storage medium and equipment, and the method comprises the steps: firstly obtaining a to-be-recognized target image, inputting the to-be-recognized target image into a pre-constructed image recognition model, and carrying out the recognition to obtain a feature vector of the target image, wherein the image recognition model is a neural network model obtained by performing self-adaptive pruning according to the computing power of neurons and training by means of adversarial training and knowledge distillation; and then, according to the feature vector, identifying the target image to obtain an identification result of the target image. Thus, the pre-constructed image recognition model performs self-adaptive pruning by using the computing power of neurons, so that the pruning efficiency is improved, and the model can effectively inherit knowledge contained in the current mainstream neural network model through the training modes of adversarial training and knowledge distillation, and the maximized recognition performance effect is ensured to be maintained, so that the image recognition efficiency can be effectively improved under the condition.

Description

technical field [0001] The present application relates to the technical field of image processing, and in particular to an image recognition method, device, storage medium and equipment. Background technique [0002] With the continuous breakthrough of artificial intelligence technology and the increasing popularity of various smart terminal devices, the number of images to be processed has shown a geometric growth. In the field of image processing, one of the most common tasks is image recognition tasks. In image recognition datasets such as ImageNet, there are more than 1,000 image categories such as airplanes, cars, birds, cats, and deer, with a total of 15 million images. The image recognition task is essentially a classification task, and it is usually necessary to solve an effective classifier to accurately classify an image into its true category. [0003] Traditional image recognition methods usually use the current mainstream neural network for recognition, such a...

AI Technical Summary

Problems solved by technology

The current network pruning method uses L1 regularization to optimize the scale factor representing the importance of neurons, so as to directly remove neurons whose scale factors are lower than a certain threshold, so as to realize the structured pruning of neural networks. This pruning method is to use L1 regularization to advance all scale factors to 0 indiscriminately, resulting in a large number of scale factors concentrated near 0, making it difficult to choose a reasonable segmentation threshold, resulting in a serious decline in network recognition accuracy, while pruning The instability of the process will also lead to a decrease in network performance after pruning, and further fine-tuning of the pruned network is required, which will also cause a waste of computing resources

Images