# US 2020/0265254 A1 · Object recognition method, apparatus and network

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2020/0265254 A1
- Document ID: US20200265254A1
- Status: Published application
- Kind: A1
- Application number: 16/777,314
- Filed: 2020-01-30
- Published: 2020-08-20
- Earliest claimed: 2019-02-19
- Assignees: Fujitsu Limited
- Applicants: Fujitsu Limited
- Inventors: Lu SHI; Qi WANG
- Web page: https://patentgrep.com/patent/US20200265254A1

## CPC

- G06V 10/764: using classification, e.g. of video objects
- G06V 10/82: using neural networks
- G06V 10/95: {structured as a network, e.g. client-server architectures}
- G06V 20/10: Terrestrial scenes (scenes under surveillance with static cameras G06V20/52; scenes perceived from the exterior of a vehicle G06V20/56; scenes perceived from the interior of a vehicle G06V20/59)
- G06V: IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING

## Abstract

Embodiments of this disclosure provide an object recognition method, apparatus and network, in which by replacing a part of convolutional layers in a conventional YOLO network with depth-wise separable convolutional layers, an amount of computation may be reduced; and by replacing residual structures in the conventional YOLO network with dense connection structures, convergence may be sped up, complexity of the network may be lowered, and the original accuracy of object recognition may be maintained.

## Claims

### Claim 1

An object recognition network, comprising:
multiple first convolutional layer groups connected to each other to extract features of an input image, at least three first convolutional layer groups among the multiple first convolutional layer groups being respectively connected to a second convolutional layer group among second convolutional layer groups,
each of the multiple first convolutional layer groups including multiple first convolutional layers, multiple first convolutional layers included in at least one first convolutional layer group being connected by dense connection structures, and one of first convolutional layers among the multiple first convolutional layers connected by the dense connection structures using depth-wise separable convolution;
multiple second convolutional layers, each second convolutional layer being located between two first convolutional layer groups among the first convolutional layer groups or being located before a first one of the first convolutional layer groups; and
three second convolutional layer groups among the second convolutional layer groups being connected to each other;
wherein, a first one of the three second convolutional layer groups is connected to a last first convolutional layer group and output thereof is taken as input of a processor, a sum of output of a second one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, a sum of output of a third one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, and each of the second convolutional layer groups includes multiple third convolutional layers.

### Claim 2 (depends on 1)

The object recognition network according to claim 1, wherein the second convolutional layers are pooling layers, or convolutional layers of a stride of 2.

### Claim 3 (depends on 1)

The object recognition network according to claim 1, wherein the second convolutional layers use an Atrous algorithm.

### Claim 4 (depends on 1)

The object recognition network according to claim 1, wherein multiple first convolutional layers contained in each of the at least three of the multiple first convolutional layer groups are connected by dense connection structures.

### Claim 5 (depends on 1)

The object recognition network according to claim 1, wherein multiple first convolutional layers contained in at least one first convolutional layer group are connected by residual structures.

### Claim 6 (depends on 1)

The object recognition network according to claim 1, wherein multiple third convolutional layers in each of the second convolutional layer groups are connected by dense connection structures.

### Claim 7

An object recognition apparatus, comprising:
a memory; and
a processor coupled to the memory and configured to:
extract features of an input image by using an object recognition network; and
perform object recognition according to the features extracted,
wherein, the object recognition network comprises:
multiple first convolutional layer groups connected to each other to extract features of an input image, at least three first convolutional layer groups among the multiple first convolutional layer groups being respectively connected to a second convolutional layer group among second convolutional layer groups,
each of the multiple first convolutional layer group including multiple first convolutional layers, multiple first convolutional layers included in at least one first convolutional layer group being connected by dense connection structures, and one of first convolutional layers among the multiple first convolutional layers connected by the dense connection structures using depth-wise separable convolution;
multiple second convolutional layers, each second convolutional layer being located between two first convolutional layer groups among the first convolutional layer groups or being located before a first one of the first convolutional layer groups; and
three second convolutional layer groups among the second convolutional layer groups being connected to each other;
wherein, a first one of the three second convolutional layer groups is connected to a last first convolutional layer group and output thereof is taken as input of a processor, a sum of output of a second one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, a sum of output of a third one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, and each of the second convolutional layer groups includes multiple third convolutional layers.

### Claim 8

An object recognition method, comprising:
extracting features of an input image by using an object recognition network; and
performing object recognition according to the extracted features,
the object recognition network includes:
multiple first convolutional layer groups connected to each other to extract features of an input image, and at least three first convolutional layer groups among the multiple first convolutional layer groups being respectively connected to a second convolutional layer group among second convolutional layer groups,
each first convolutional layer group including multiple first convolutional layers, multiple first convolutional layers included in at least one first convolutional layer group being connected by dense connection structures, and one of first convolutional layers among the multiple first convolutional layers connected by the dense connection structures using depth-wise separable convolution;
multiple second convolutional layers, each second convolutional layer being located between two first convolutional layer groups among the first convolutional layer groups or being located before a first one of the first convolutional layer groups; and
three second convolutional layer groups among the second convolutional layer groups being connected to each other;
wherein, a first one of the three second convolutional layer groups is connected to a last first convolutional layer group and output thereof is taken as input of a processor, a sum of output of a second one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, a sum of output of a third one of the three second convolutional layer groups and output of one first convolutional layer group is taken as input of the processor, and each of the second convolutional layer groups includes multiple third convolutional layers.
