DEEP LEARNING MODEL USED FOR IMAGE RECOGNITION AND TRAINING APPARATUS OF THE MODEL AND METHOD THEREOF

• US 20200134385A1
• Filed: 10/16/2019
• Published: 04/30/2020
• Est. Priority Date: 10/29/2018
• Status: Active Grant

• Current Assignee: Fujitsu Limited
• Original Assignee: Fujitsu Limited
• Inventors: YIN, Rui, TAN, Zhiming
• Granted Patent: US 11,361,190 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/214 Generating training patterns; Bootstrap methods, e.g. bagging or boosting
• CPC Class Codes: G06F 18/217 Validation; Performance evaluation; Active pattern learning techniques
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• 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 10/776 Validation; Performance evaluation
• CPC Class Codes: G06V 10/82 using neural networks

Abstract

Embodiments of this disclosure provide a deep learning model used for image recognition and apparatus and method thereof. The model includes a determination layer configured to determine whether features in feature maps are features of positions where objects of attention are located, and different weights are granted for the positions where the objects of attention are located and other features in performing weight and composition processing on the features. Hence, the model may be guided to be focused on attention features and make correct determination, thereby improving performance and precision of the model.

Claims

1. A deep learning model used for image recognition, the model comprising:

a plurality of convolutional layers configured to extract features from an input image in turn and output a plurality of feature maps of identical sizes;

a determination layer configured to, according to positions where objects of attention in the input image are located, determine whether features related to positions contained in the feature maps are features of the positions where the objects of attention are located;

a compositing layer configured to, according to an output result of the determination layer, perform weight and composition processing on the features in the plurality of feature maps outputted by the plurality of convolutional layers, weights of the features of the positions where the objects of attention are located being different from weights of other features; and

a fully-connected layer configured to output a recognition result according to the plurality of feature maps after being weight and composition processed by the compositing layer.

2. The deep learning model according to claim 1, wherein,

the compositing layer multiplies the plurality of feature maps by the weights of the features related to positions contained in the feature maps to obtain the plurality of feature maps after being weight and composition processed.

3. The deep learning model according to claim 1, wherein the deep learning model further comprises:

a long short-term memory layer provided between the compositing layer and the fully-connected layer;

and the input image comprises a temporally consecutive frame sequence.

4. A training apparatus of the deep learning model according to claim 1, the apparatus comprising:

an inputting unit configured to input a training image into the plurality of convolutional layers of the deep learning model;

a first calculating unit configured to calculate an attention loss according to the output result of the determination layer of the deep learning model and real values of positions where the preset objects of attention are located;

a second calculating unit configured to calculate a classification loss according to the output result of the fully-connected layer of the deep learning model and a preset real value of classification; and

an adjusting unit configured to perform back propagation according to the attention loss and the classification loss to adjust parameters of the plurality of convolutional layers and the determination layer of the deep learning model.

5. The apparatus according to claim 4, wherein,

the first calculating unit calculates an accumulative value of differences between probabilities that the positions where the features outputted by the determination layer are located are the objects of attention and real values of the positions being the objects of attention to obtain the attention loss.

6. The apparatus according to claim 4, wherein,

the adjusting unit performs back propagation according to a weighted sum of the attention loss and the classification loss, to adjust the parameters of the plurality of convolutional layers and the determination layer of the deep learning model.

7. The apparatus according to claim 6, wherein the apparatus further comprises:

a determining unit configured to determine respective weights of the attention loss and the classification loss.

8. A training method of the deep learning model according to claim 1, the method comprising:

inputting a training image into the plurality of convolutional layers of the deep learning model;

calculating an attention loss according to the output result of the determination layer of the deep learning model and real values of positions where the preset objects of attention are located;

calculating a classification loss according to the output result of the fully-connected layer of the deep learning model and a preset real value of classification; and

performing back propagation according to the attention loss and the classification loss, to adjust parameters of the plurality of convolutional layers and the determination layer of the deep learning model.

9. The method according to claim 8, wherein,

the calculating an attention loss according to the output result of the determination layer of the deep learning model and real values of positions where the preset objects of attention are located comprises:

calculating an accumulative value of differences between probabilities that the positions where the features output by the determination layer are located are the objects of attention and real values of the positions being the objects of attention to obtain the attention loss.

10. The method according to claim 8, wherein,

the performing back propagation according to the attention loss and the classification loss, to adjust parameters of the plurality of convolutional layers and the determination layer of the deep learning model, comprises:

performing back propagation according to a weighted sum of the attention loss and the classification loss, to adjust the parameters of the plurality of convolutional layers and the determination layer of the deep learning model.