Convolutional neural network construction method and device, equipment and medium

Abstract

The embodiment of the invention provides a convolutional neural network construction method and device, equipment and a medium. The method comprises the steps: determining a to-be-replaced convolutionmodule from an original convolutional neural network, the to-be-replaced convolution module comprises a plurality of convolution layers, and there is a direct connection branch between the input endand the output end of the to-be-replaced convolution module; replacing the convolution module to be replaced with a receptive field adaptive module to obtain a target convolutional neural network, thereceptive field adaptive module being used for respectively generating corresponding weight values for a plurality of receptive fields so as to process the feature maps of the plurality of receptivefields, and outputting the processed feature maps; wherein the sum of the output of the receptive field adaptive module and the output of the direct connection branch is the input of the next convolution module to be replaced.

Description

technical field [0001] The present invention relates to the technical field of deep learning, in particular to a method, device, equipment and medium for constructing a convolutional neural network. Background technique [0002] Image recognition is a basic task in the field of computational vision, which can identify or verify the identity, attributes or categories of target subjects in images. The existing image recognition methods are mainly learnable feature methods represented by neural networks. Neural networks are widely used in image recognition tasks due to their powerful, self-adaptive feature expression capabilities that do not require artificial fine design. [0003] In practice, in order to improve the recognition efficiency, convolutional neural networks are generally used for image recognition. Convolutional neural networks include a variety of framework models, such as AlexNet, Resnet, Resnext, Mobilenet, Shufflenet, and VGG. These models can expand the rec...

AI Technical Summary

Problems solved by technology

These models can expand the receptive field of the model by accumulating layer by layer, saving computing power and storage resources, but the global information is lost, and there is no more efficient attention mechanism, so the information extracted during the recognition process is not effective key information

[0004] In response to this problem, model structures such as GoogLeNe, Res2net and 3FPN have been proposed in related technologies to alleviate the problems of information loss and single receptive field

However, these models still have the following shortcomings: the model design is still not intelligent enough, and the accuracy of feature extraction for feature maps of different receptive fields is not high

Images