Kaleidoscope convolution network construction method

Abstract

The invention discloses a kaleidoscope convolution network construction method and belongs to a deep learning algorithm network. According to the prior art, a void filter is used for expanding the receptive field of a convolutional network, and when the void ratio is set to be too large, a boundary effect influences feature capture. The invention aims to solve the above problem in the prior art. The method comprises the following steps that: a kaleidoscope filter is constructed, the outer contour of the kaleidoscope filter is made to be a square, the square is equally divided into four segmentation units, and nine primary sampling points are determined according to the four corners of the four segmentation units; three secondary sampling points are determined in each segmentation unit according to the intersection point positions of kaleidoscope type petals; a kaleidoscope cascade spatial pyramid pooling module is constructed based on kaleidoscope filters of different sizes; and finally, Xception is adopted as a backbone network, and the kaleidoscope cascaded spatial pyramid pooling module is assembled on a decoder layer, so that a kaleidoscope convolutional network can be formed.With the method of the invention adopted, the shape features of an object can be better captured.

Description

technical field [0001] The invention relates to a construction method of a kaleidoscope convolutional network, which belongs to a deep learning algorithm network. Background technique [0002] Deep feature representations based on convolutional neural networks (CNNs) have greatly improved semantic segmentation. At present, more successful semantic segmentation methods based on CNNs mainly rely on high-resolution representations and rich feature representations. [0003] It is generally believed that the larger the convolution size, the larger the receptive field, and the more accurate the features captured. These features include texture features, shape features, and contextual features, which are crucial for object recognition. But larger convolution filters lead to higher computational cost. [0004] In order to obtain a large receptive field with low computational cost, many traditional convolutional networks use a large number of 3×3 filters for serial convolution, an...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the problem that the existing hole filter is used to expand the receptive field of the convolutional network, and the boundary effect generated when the hole rate is set too large affects the feature capture, the present invention provides a method for constructing a kaleidoscope convolutional network

Images