Deconvolution neural network training method

Abstract

The invention discloses a deconvolution neural network training method. With the deconvolution neural network training method adopted, image features can be effectively extracted, and the improvement of correct rate of classification can be benefitted, and convergence efficiency and convergence precision of the training of a deconvolution neural network can be improved, and training cost of the deconvolution neural network in practical application can be decreased. The deconvolution neural network training method can be applied to other convolution operation-based optimization problem solving. The deconvolution neural network training method includes a training stage and a reconstruction stage. The training stage comprises the following steps that: (1) training images are preprocessed; (2) batch setting is performed on the training images; (3) network training parameters of the training images are set; and (4) training of a first layer is started. The reconstruction stage comprises the following steps that: (5) images to be reconstructed are preprocessed; (6) network training parameters of the images to be reconstructed are set; and (7) the images to be reconstructed are inputted in a batched manner until the reconstruction of the images in all batches is completed.

Description

technical field [0001] The invention belongs to the technical field of artificial intelligence, and in particular relates to a deconvolution neural network training method. Background technique [0002] Deep learning is one of the research hotspots in the field of artificial intelligence. Based on the idea of ​​neural networks, deep learning is committed to imitating the hierarchical perception of the human brain. By constructing a multilayer perceptron (Multilayer Perception, MLP), an abstract high-level representation that can represent attribute categories or features is combined from low-level features, thus becoming One of the effective models for extracting complex image features at present. [0003] Classic deep learning models mainly include Deep Belief Networks (DBNs), multi-layer sparse auto-encoding models (Auto Encode, AE) and convolutional neural networks (Convolutional Neural Networks, CNNs). Usually, these models extract features from the input image through...

AI Technical Summary

Problems solved by technology

However, it still needs more work and improvement to extend such methods to multi-layer deconvolutional neural network models.

Images