Image pattern recognition device and recording medium

• US 10,664,727 B2
• Filed: 12/21/2016
• Issued: 05/26/2020
• Est. Priority Date: 04/01/2016
• Status: Active Grant

• Current Assignee: Fujifilm Business Innovation Corp. (Fujifilm Holdings Corporation)
• Original Assignee: Fuji Xerox Company Limited (Fujifilm Holdings Corporation)
• Inventors: Shigenaka, Ryosuke, Tsuboshita, Yukihiro, Kato, Noriji
• Primary Examiner(s): Seth, Manav
• Application Number:
  US15/386,473
• Publication Number:
  US 20170286810A1
• Time in Patent Office: 1,252 Days
• CPC Class Codes: G06F 18/2431 Multiple classes
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 20/46 Extracting features or characteristics from the video content, e.g. video fingerprints, representative shots or key frames

Abstract

An image pattern recognition device includes: a data reception unit that receives data; a supervision reception unit that receives supervision; and an artificial neural network processing unit that performs artificial neural network processing. The artificial neural network processing unit includes a first sub-network including one or more layers that process a main task, a second sub-network including one or more layers that process a sub-task, and a third sub-network including one or more layers that do not belong to any of the first sub-network and the second sub-network. The third sub-network includes a branch processing unit that outputs a value same as an input feature amount to a plurality of layers. The first sub-network includes a coupling processing unit that couples inputs from the plurality of layers and outputs a result.

Claims

1. A non-transitory computer readable recording medium storing a program that causes a computer to execute:

receiving data;

receiving supervision; and

performing artificial neural network processing to determine a gender of a poster of an image from image content of the image, the image being read from the received data,

wherein performing the artificial neural network processing includes

processing a main task with a first sub-network including one or more layers, the main task corresponding to gender estimation, the first sub-network configured to extract a first feature amount which is a feature amount of the gender,

processing a sub-task with a second sub-network including one or more layers, the sub-task corresponding to content recognition, the second sub-network configured to extract a second feature amount which is a feature amount of the image content, and

performing processing not included in any of the main task and the sub-task with a third sub-network including one or more layers, the third sub-network not sharing a layer with any of the first sub-network and the second sub-network, the third sub-network configured to extract a third feature amount that is an output result from the image, wherein

performing the processing not included in any of the main task and the sub-task includes outputting a value same as an input feature amount to a plurality of layers of the first sub-network and the second sub-network, and

processing the main task includes coupling inputs from the plurality of layers of the first sub-network and the second sub-network and outputting a result, wherein the coupling includes coupling the feature amount of the gender and the feature amount of the image content by performing processing of concatenating the feature amount of the gender and the feature amount of the image content, or processing of calculating with respect to the plurality of layers of the first sub-network and the second sub-network for each element of the received feature amount of the gender and the received feature amount of the image content so as to integrate the feature amount of the gender and the feature amount of the image content, and outputting the result.

2. The non-transitory computer readable recording medium according to claim 1,

wherein a non-linear calculation is not applied in a final layer in the first sub-network and a final layer in the second sub-network, and a layer other than the final layer in the first sub-network and other than the final layer in the second sub-network is formed from any one of a total coupling layer that applies a linear calculation and the non-linear calculation, and a convolution layer that performs a convolution processing, the non-linear calculation, and a pooling processing.

3. The non-transitory computer readable recording medium according to claim 1,

wherein the program further causes the computer to execute

outputting a determination result of each task,

outputting an error between two types of inputs, and

executing an error back-propagation method based on a stochastic gradient descent.

4. The non-transitory computer readable recording medium according to claim 1, wherein

processing the main task further includes branching input errors in accordance with the processing of concatenating the inputs or the processing of calculating and outputting the result to the plurality of layers, and

performing the processing not included in any of the main task and the sub-task further includes integrating errors input from the plurality of layers.

5. The non-transitory computer readable recording medium according to claim 1,

wherein the program further causes the computer to execute a normalization processing.