A Convolutional Neural Network Vertical Segmentation Method for Image Processing

Abstract

The invention discloses an image processing-oriented convolutional neural network vertical segmentation method, which belongs to the field of deep learning and distributed computing. This method first divides the input feature map of the last layer of the continuous convolutional layer into continuous sub-feature maps, and then calculates the corresponding sub-feature maps of the previous layer layer by layer according to the sub-feature maps and convolution calculation operations. Up to the first layer, refer to the sub-feature map of the first layer, segment the input feature map of the first layer, and distribute the sub-feature map after the first layer segmentation to multiple computing nodes. Finally, according to the parameters and hyperparameters of the single-link continuous convolution layer, the distributed collaborative inference without loss of precision is implemented, and after all the inference results are generated, the inference results are summarized to generate the final output feature map. Compared with the previous methods, the method of the present invention has the characteristics of greatly reducing the reasoning delay of the convolutional neural network without loss of precision.

Description

technical field [0001] The invention relates to the field of deep learning and distributed computing, in particular to an image processing-oriented convolutional neural network vertical segmentation method. Background technique [0002] With the development of computer hardware and the surge in the amount of application data, the capabilities of deep learning models have been gradually released. The results obtained by the deep learning model after processing the data are highly accurate, so it is widely used in various data processing programs. Among these data processing programs worth mentioning are image processing programs. The convolutional neural network used in it greatly improves the accuracy of image processing, making high-precision image recognition, video analysis and other programs possible. However, the convolution operation required by the convolutional layer in the convolutional neural network requires a large amount of computing power. In some computing n...

AI Technical Summary

Problems solved by technology

[0005] 1. The current channel segmentation method has a result splicing process after the convolution operation is performed on each sub-channel, and the spliced ​​result is used as the input feature map of the next convolutional layer. This method has a large amount of data transmission redundancy, and The multiple stitching process of the result caused unnecessary computational overhead

[0006] 2. The existing space segmentation method does not fully consider the filling process of the input feature map, and there is a loss of precision, resulting in inaccurate model reasoning results

[0007] 3. The existing space segmentation method does not correctly consider the reverse derivation process of the convolution operation, resulting in inaccurate correspondence between the sub-feature maps of the continuous convolution layer on a single link, resulting in loss of precision, resulting in inaccurate model reasoning results

Images