Optimization method and system of target detection model

Abstract

The invention discloses an optimization method and system of a target detection model, relates to the technical field of computer vision, and solves the technical problems that the number of output channels of a convolutional network is large, the calculation amount of a convolution process is large, and the representation capability is not strong.According to the technical scheme, an image is divided into a plurality of feature maps; the features of the feature maps are divided, the features of the different feature maps are jointly input into a shuffling channel, all the features are shuffled, the input channel and the output channel are completely related through uniform combination of the different features, and the characterization capacity is enhanced. The number of the convolution kernels is smaller than the number of the feature maps, and the number of output channels is reduced, so that the calculated amount of the convolution kernel in the convolution process is reduced.

Description

technical field [0001] The present disclosure relates to the technical field of computer vision, in particular to a method and system for optimizing a target detection model. Background technique [0002] The convolutional neural network for mobile vision applications is one of the lightweight networks. This network is mainly used in mobile and embedded devices. The biggest feature of this network is the application of deep point-by-point convolutional networks and the addition of two Hyperparameters, one is a hyperparameter that controls the number of convolution kernels in the convolutional layer, and the other is a hyperparameter that controls the size of the input image. [0003] The characteristics of the standard convolutional network are: the depth of the convolution kernel is equal to the depth of the input feature matrix, and the depth of the output feature matrix is ​​equal to the number of convolution kernels. For example, if you input a feature matrix with a dep...

AI Technical Summary

Problems solved by technology

It can be seen that the calculation amount of point-by-point convolution is proportional to the number of output channels N, and the number of output channels will continue to increase as the number of network layers deepens, which directly leads to a gradual increase in the calculation amount of point-by-point convolution operations; in addition Each output channel is not combined to generate new features, and the number of channels of features is not changed. Even if the number of calculation parameters is reduced, information between different channels cannot be circulated.

Images