IMAGE RECOGNITION METHOD

• US 20170083796A1
• Filed: 09/12/2016
• Published: 03/23/2017
• Est. Priority Date: 09/18/2015
• Status: Active Grant

• Current Assignee: Panasonic Intellectual Property Corporation of America (Panasonic Holdings Corporation)
• Original Assignee: Panasonic Intellectual Property Corporation of America (Panasonic Holdings Corporation)
• Inventors: Kim, Min Young, Rigazio, Luca, FUJIMURA, RYOTA, Tsukizawa, Sotaro, KOZUKA, Kazuki
• Granted Patent: US 10,496,901 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/2414 Smoothing the distance, e.g. radial basis function networks [RBFN]
• 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 20/58 Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
• CPC Class Codes: G06V 40/103 Static body considered as a whole, e.g. static pedestrian or occupant recognition

Abstract

In an image recognition method executed by a computer of an image recognizer using a convolutional neural network, the convolutional neural network is a first convolutional neural network in which a fully connected layer is changed to a convolutional layer, and the method includes controlling a first convolutional neural network to acquire an input image, controlling the first convolutional neural network to estimate a center area of a recognition target in the acquired input image and to output a value indicating the estimated center area as a location of the recognition target in the input image.

Claims

1. An image recognition method executed by a computer of an image recognizer using a convolutional neural network, the convolutional neural network being a first convolutional neural network in which a fully connected layer is changed to a convolutional layer, the method comprising:

controlling the first convolutional neural network to acquire an input image; and

controlling the first convolutional neural network to estimate a center area of a recognition target in the acquired input image and to output a value indicating the estimated center area as a location of the recognition target in the input image.

2. The image recognition method according to claim 1, further comprising:

before performing the acquisition of the input image by the first convolutional neural network, controlling the first convolutional neural network to learn a weight of the first convolutional neural network by using learning data including a plurality of sets of a training image and a correct answer image, the training image including a recognition target, the correct answer image including the training image and a value indicating a center area of the recognition target.

3. The image recognition method according to claim 2, wherein the learning including

controlling a second convolutional neural network to output a value indicating a location of the recognition target in the training image included in the learning data, the second convolutional neural network having the same structure as that of the first convolutional neural network except that a fully connected layer in the convolutional neural network is replaced with a convolutional layer,

calculating an error, according to an error function, between the value indicating the center area of the recognition target indicated by the correct answer image corresponding to the training image included in the learning data and the value indicating the recognition target output in the learning,

updating the weight of the second convolutional neural network based on the calculated error,

recursively performing the calculation by again making the second convolutional neural network with the updated weight perform the outputting in the learning, and

in a case where the calculated error achieves a minim value in the recursive calculation, ending the recursive calculation and determining the weight of the second convolutional neural network assigned when the error achieves the minimum value as the weight of the first convolutional neural network.

4. The image recognition method according to claim 3, wherein the error function is an error function using cross entropy.

5. The image recognition method according to claim 1, wherein the center area is a center area extending in a vertical direction, of center areas one of which extends in the vertical direction and the other one of which extends in a horizontal direction.

10. The image recognition method according to claim 1, wherein the recognition target is a person in the input image.

6. An image recognition method executed by a computer of an image recognizer using a convolutional neural network, the convolutional neural network being a first convolutional neural network in which a fully connected layer is changed to a convolutional layer, the method comprising:

controlling the first convolutional neural network to acquire an input image; and

controlling the first convolutional neural network to estimate coordinates of two vertices on a diagonal of a rectangle surrounding a recognition target in the acquired input image and to output the estimated coordinates of the two vertices as a location of the recognition target in the input image.

7. The image recognition method according to claim 6, further comprising

before performing the acquisition of the input image by the first convolutional neural network, controlling the first convolutional neural network to learn a weight of the first convolutional neural network by using learning data including a plurality of sets of a training image and a correct answer image, the training image including a recognition target, the correct answer image including the training image and coordinates of two vertices on a diagonal of a rectangle surrounding the recognition target.

8. The image recognition method according to claim 7, wherein the learning includes

controlling a second convolutional neural network to output a value indicating a location of the recognition target in the training image included in the learning data, the second convolutional neural network having the same structure as that of the first convolutional neural network except that a fully connected layer in the convolutional neural network is replaced with a convolutional layer,

calculating an error, according to an error function, between the output value indicating the location of the recognition target and coordinates of two vertices on a diagonal of a rectangle surrounding the recognition target indicated by the correct answer image corresponding to the training image included in the learning data,

updating the weight of the second convolutional neural network based on the calculated error,

recursively performing the calculation by controlling the second convolutional neural network with the updated weight to again perform the outputting in the learning, and

in a case where the calculated error achieves a minim value in the recursive calculation, ending the recursive calculation and determining the weight of the second convolutional neural network assigned when the error achieves the minimum value as the weight of the first convolutional neural network.

9. The image recognition method according to claim 8, wherein the error function is an error function using a Euclidean distance.