Method and system for enabling CNNs with fully connected layers to accept input of indeterminate shape

Abstract

The present invention discloses a method and system for enabling CNN with a fully connected layer to accept indeterminate shape input. The present invention adds a new feature included in the convolution pooling of the existing CNN network with a fully connected layer. After the coordinate channel is added to the graph, it is processed to output a variable-step pooling layer of a new feature map with a fixed size and a coordinate prediction network composed of multiple fully connected layers and a softmax function layer. The coordinate prediction network determines the features that contain effective pixel areas. The probability distribution vector of the coordinates (X, Y) of the center point of the block is processed differently according to the horizontal dimension W and vertical dimension H of the feature map, so that the network can finally accept indeterminate shape input. The present invention can effectively solve the defect that indefinite shape input cannot be accepted in a CNN with a fully connected layer, and can enable a given CNN with a fully connected layer that can only accept fixed shape input to accept an indeterminate shape input, and other requirements remain unchanged , has the advantage of good compatibility.

Description

technical field [0001] The invention belongs to the field of model design for deep learning, and in particular relates to a method and system for enabling a given CNN with a fully connected layer that can only accept input of a fixed shape to accept an input of an indeterminate shape. Background technique [0002] In real life, the pictures we can come into contact with are various and varied, some are long strips, some are square, some have high resolution and some have low resolution. These large numbers and various types of pictures have promoted the development of computer vision technology based on convolutional neural network (CNN for short). CNN-based computer vision technology has shown great potential in image classification, object detection and many other recognition tasks, even non-recognition tasks, so the development of basic operators in CNN is more necessary and meaningful. [0003] Many current CNN structures consist of two parts, the convolutional part and...

AI Technical Summary

Problems solved by technology

[0010] The technical problem to be solved by the present invention: Aiming at the above-mentioned problems of the prior art, a method and a system for enabling CNN with a fully connected layer to accept indeterminate shape input are provided. The defect of accepting indeterminate shape input can make a given CNN with a fully connected layer that can only accept fixed shape input can accept indeterminate shape input, and other requirements remain unchanged, which has the advantage of good compatibility

Images