Riemannian manifold deep convolutional network image classification method

Abstract

Aiming at the defects of ResNet series models, the invention provides a Riemannian manifold deep convolutional neural network capable of improving image classification performance, which is called RM-CNN. An RM-CNN image classification model is designed on a ResNet series network, feature correlation on each level is fully utilized, a Riemannian manifold network is adopted to make up for information loss, and meanwhile, a self-attention model is utilized to improve performance; on the whole, the width of the ResNet network is increased, the gradient disappearance problem is solved through the width, and the purpose of improving the image classification performance is achieved.

Description

technical field [0001] The invention relates to the field of image classification, in particular to an image classification method utilizing a Riemannian manifold model fused into a deep convolutional network. Background technique [0002] Image classification technology is widely used in education, aerospace, military and other fields. At present, most images use deep convolutional neural models (such as ResNet, DenseNet) to extract image features, and then identify image content. However, the performance of mainstream networks such as ResNet has reached a bottleneck, the room for improvement is limited, and it still does not meet the needs of practical application development. [0003] The mainstream ResNet adopts the method of adding a residual module in the network, which solves the problem that the performance of the deep network stack does not increase but decreases at a certain level: that is, after the network depth reaches a certain level, the gradient disappears s...

AI Technical Summary

Problems solved by technology

However, the performance of mainstream networks such as ResNet has reached a bottleneck, the room for improvement is limited, and it still does not meet the needs of practical application development.

[0003] The mainstream ResNet adopts the method of adding a residual module in the network, which solves the problem that the performance of the deep network stack does not increase but decreases at a certain level: that is, after the network depth reaches a certain level, the gradient disappears seriously and the error increases. The effect of recognition and classification becomes worse, and the gradient cannot be fed back to the front layer network during backward propagation, resulting in the inability to update the front network parameters

The residual module can increase the depth of the network, but when the depth of the network increases to a certain level (the recognition performance still does not meet the requirements), there will still be a problem of performance degradation

At the same time, another disadvantage of the Riemannian manifold is that it will generate a large amount of information loss. This information loss occurs at multiple levels. The loss mainly comes from convolution operations and downsampling operations.

Images