Method for expanding repvg structure model, image processing method and device

Abstract

The invention discloses a method for expanding a repvg structure model, and an image processing method and device. The method comprises the following steps: S1, collecting a corresponding picture according to a problem; s2, labeling the collected pictures; s3, designing a deep convolutional neural network, and on the basis of repvg, adding k extra convolutions (conv1, conv2,..., convk) beside each conv; s4, inputting the picture and the corresponding annotation information into a designed deep convolutional neural network for training to obtain a trained model; s5, performing fusion weight processing on the trained model to obtain a final model; and S7, solving an image processing problem by using the final model. According to the method, a better weight can be learned in a larger parameter space, so that a model with a better effect is obtained, and when the method is applied to an image processing technology, the larger parameter space can be brought, the calculation amount is reduced, and the image processing efficiency is improved.

Description

technical field [0001] The present invention relates to the field of image processing, and more specifically, relates to a method for extending a repvgg structural model, an image processing method, and a device. Background technique [0002] In the deep convolutional neural network, the emergence of the resnet structure has greatly improved the effect of the model under the same amount of parameters / computation, and because of the cross-layer connection, it is convenient for gradient lossless return, which greatly accelerates the model training and begins to be able to Train very deep models. However, in the actual deployment of the model, due to the hardware characteristics and optimization settings of most of the devices that need to be deployed, resnet will be slower than the vgg structure model with the same amount of calculation, so resnet is often not suitable for practical application deployment. [0003] With the development of technology, the repvgg structure prop...

AI Technical Summary

Problems solved by technology

[0004] Repvgg mainly obtains the advantages of resnet during training by adding cross-layer connections that can be fused later to vgg: cross-layer connections are convenient for gradient return, so it is easy to model training, and the final effect may be better, but there are the following disadvantages: it does not Does not bring a larger parameter space, the parameter space is small

Therefore, the image processing technology based on the repvgg structure model has the problem of small parameter search space and low image processing efficiency

Images