Deep deconvolution feature learning network, generating method thereof and image classifying method

Abstract

The invention discloses a generating method of a deep deconvolution feature learning network. The generating method comprises the steps that a multi-layer deconvolution feature learning network model is pre-trained in an unsupervised mode; fine adjustment of the learning network model is conducted with object detecting information from top to bottom. The invention further provides the deep deconvolution feature learning network and an image classifying method, wherein the deep deconvolution feature learning network is generated according to the generating method. According to the generating method of the deep deconvolution feature learning network, non-negative sparsity restraints are introduced into the deep feature learning model, the recognition capacity of features is improved, and the image classification accuracy is improved; the object detection information is used as high-level guiding information from top to bottom for fine adjustment of the trained network, so that different nodes in the network have high selectivity for input image structures, especially the nodes on the highest level have different responses to different object types, in this way, obtained high-level features have obvious semantic meaning, and the image classification accuracy is improved.

Description

technical field [0001] The present invention relates to the field of machine learning, and more specifically, relates to a deep deconvolution feature learning network, a generation method and an image classification method. Background technique [0002] In computer vision, how to construct good image features has always been a key and extremely challenging problem. The quality of features directly determines the performance of the entire computer vision system, such as image recognition, image retrieval, and pedestrian detection. Some artificially designed image descriptors (such as SIFT and HOG) have achieved great success. Although these artificially designed features can make good use of human understanding and prior knowledge of images, their performance depends on specific task and cannot characterize the middle and high-level structures of complex images. [0003] In recent years, many research works have attempted to construct deep networks for image feature learning...

AI Technical Summary

Problems solved by technology

These deep networks embody the hierarchy of the human visual system, and automatically learn to extract image features from image data, but these models lack good discriminative selectivity and some high-level information guidance, and the learned features are used in natural image recognition tasks. Difficult to get good performance in

Images