JP7428075B2 - Object detection device, object detection method and terminal equipment - Google Patents

Description

The present invention relates to the field of information technology.

In recent years, research in the field of computer vision has made great progress thanks to deep learning. Deep learning refers to a collection of algorithms that apply various machine learning algorithms to hierarchical neural networks to solve various problems such as images and text. Feature learning, as the core of deep learning, aims to solve the traditional problem of having to manually design features by acquiring hierarchical feature information through a hierarchical neural network. .

Currently, there are several popular deep learning methods, such as YOLO network, which is a promising deep learning method for object recognition and detection. For example, the YOLO-Darknet53 network, which has darknet53 as its backbone network structure, has multi-scale object detection and an excellent classifier, so it has faster processing speed and higher recognition accuracy than a single stage. Here, darknet53 structure is used for feature extraction.

It should be noted that the above description of the technical background is provided to provide a clear and complete understanding of the technical solution of the present invention, and is provided to provide a clear and complete understanding to those skilled in the art. These technical solutions are merely explained as a background technical part of the present invention, and are not well known by those skilled in the art.

However, neural networks with high recognition accuracy, such as the YOLO-Darknet53 network, have wide and deep layers, and therefore have high memory and processing speed requirements for the processor. For example, in the YOLO-Darknet53 network, the number of operations per second (FLOPs) is 739.8M, and the processing speed of the central processing unit (CPU) is 1. This is 1375.8 ms per frame, and the processing speed of a graphics processing unit (GPU) is required to be 37.0 ms per frame. For example, terminal equipment such as in-vehicle devices can only support approximately 100M FLOPs. For this reason, current neural networks with high recognition accuracy may not be applicable to mobile devices.

Embodiments of the present invention provide an object detection device, an object detection method, and a terminal device. Since the number of input and output channels of all convolutional layers in the shuffle part for feature extraction is the same, there is no need to perform feature expansion and compression, reducing memory and performance requirements on the processor and processing speed. Speed can be improved. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, so it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

In a first aspect of the embodiment of the present invention, the object detection device includes a feature extraction unit that extracts features in an input image, and detects an object in the input image based on the features extracted by the feature extraction unit. a detection unit that performs a detection unit, the feature extraction unit includes at least one shuffling unit, the shuffling unit includes a plurality of convolutional layers, and the number of input channels and the output of each convolutional layer of the plurality of convolutional layers. The number of channels is the same, and the plurality of convolutional layers includes at least one depthwise separable convolutional layer.

In a second aspect of an embodiment of the invention there is provided a terminal device comprising the apparatus according to the first aspect of an embodiment of the invention.

In a third aspect of the embodiment of the present invention, there is provided an object detection method, wherein the feature extraction section extracts features in the input image, and the detection section extracts features from the input image based on the features extracted by the feature extraction section. detecting an object in the plurality of convolutional layers, wherein the feature extraction unit includes at least one shuffling unit, the shuffling unit includes a plurality of convolutional layers, and the feature extraction unit includes at least one shuffling unit, and the shuffling unit includes a plurality of convolutional layers, and the feature extraction unit includes at least one shuffling unit, and the shuffler unit includes a plurality of convolutional layers, and the feature extraction unit includes at least one shuffling unit, and the shuffler unit includes a plurality of convolutional layers, and the feature extraction unit includes at least one shuffling unit, and the shuffler unit includes a plurality of convolutional layers. and the number of output channels are the same, and the plurality of convolutional layers includes at least one depth-separable convolutional layer.

The advantageous effects of the present invention are as follows. Since the number of input and output channels of all convolutional layers in the shuffle part for feature extraction is the same, there is no need to perform feature expansion and compression, reducing memory and performance requirements on the processor and processing speed. Speed can be improved. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, so it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

Certain embodiments of the invention are disclosed in detail and illustrate the manner in which the principles of the invention may be employed, as set forth in the following description and drawings. Note that the embodiments of the present invention are not limited in scope. Embodiments of the present invention include various alterations, modifications, and equivalents within the spirit and content of the appended claims.

Features described and/or illustrated in one embodiment may be used in one or more other embodiments in the same or similar manner, and may be combined with features in other embodiments; Features in other embodiments may be substituted.

Note that the term "comprising/comprising", when used in the main text, means the presence of a feature, element, step, or component, and does not include the presence or absence of one or more other features, elements, steps, or components. This does not exclude additions.

The drawings included herein are for the purpose of providing an understanding of embodiments of the invention, constitute a part of this specification, and are intended to illustrate embodiments of the invention, and together with the written description. The principle of the present invention will now be explained. Note that the drawings described here are merely for explaining embodiments of the present invention, and those skilled in the art can easily obtain other drawings based on these drawings.
1 is a diagram showing an object detection device according to a first embodiment of the present invention. FIG. 3 is a diagram showing detection results of an input image by the object detection device 10 according to Example 1 of the present invention. FIG. 2 is a diagram showing a feature extraction unit 100 according to Example 1 of the present invention. FIG. 3 is a diagram showing a first shuffle section 103 according to Example 1 of the present invention. It is a figure showing the second shuffle part 104 concerning Example 1 of the present invention. It is a figure showing the terminal equipment concerning Example 2 of the present invention. FIG. 2 is a block diagram showing a system configuration of a terminal device according to a second embodiment of the present invention. FIG. 7 is a diagram showing an object detection method according to Example 3 of the present invention.

These and other features of the invention will become clear from the drawings and the following description. The specification and drawings disclose certain embodiments of the invention, ie, some embodiments in accordance with the principles of the invention. Note that the present invention is not limited to the described embodiments, and the present invention includes all modifications, variations, and equivalents within the scope of the claims.

<Example 1>
Embodiments of the present invention provide an object detection device. FIG. 1 is a diagram showing an object detection device according to a first embodiment of the present invention.

As shown in FIG. 1, the object detection device 10 includes a feature extraction section 100 and a detection section 200.

The feature extraction unit 100 extracts features in the input image.

The detection unit 200 detects an object in the input image based on the features extracted by the feature extraction unit 100.

Here, the feature extraction unit 100 includes at least one shuffle unit, the shuffle unit includes a plurality of convolutional layers, and the number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers. The number is the same, and the plurality of convolutional layers includes at least one depthwise separable convolutional layer.

FIG. 2 is a diagram showing a detection result of an input image by the object detection device 10 according to the first embodiment of the present invention. As shown in FIG. 2, the object detection device 10 can accurately detect each object in an image.

According to this embodiment, since the number of input channels and output channels of all convolutional layers in the shuffle unit for feature extraction is the same, there is no need to expand and compress features, and the memory and performance of the processor are reduced. requirements and increase processing speed. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, so it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

In this embodiment, the input image may be an image acquired in real time or an image acquired in advance. For example, the input images are video images taken by an in-vehicle device, and each input image corresponds to one frame of the video image.

In this embodiment, the feature extraction unit 100 extracts features in the input image. The feature extraction unit 100 includes at least one shuffle unit, the shuffle unit includes a plurality of convolutional layers, and the number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers are the same, The plurality of convolutional layers includes at least one depth-wise separable convolutional layer.

In this embodiment, the at least one shuffle section may include at least one first shuffle section and/or at least one second shuffle section. Here, the first shuffle part is a shuffle part with a stride of 1, and the second shuffle part is a shuffle part with a stride of 2.

The configuration of the feature extraction unit 100 of this embodiment will be exemplified below.

FIG. 3 is a diagram showing the feature extraction unit 100 according to the first embodiment of the present invention. As shown in FIG. 3, the feature extraction section 100 includes a first convolution layer 101, a pooling layer 102, a plurality of first shuffle sections 103, and a plurality of second shuffle sections 104.

The first convolutional layer 101 processes the input image.

The pooling layer 102 performs pooling processing on the features output by the first convolutional layer 101.

In this embodiment, the first convolution device and pooling layer 102 may use a conventional structure.

In this embodiment, the number of the first shuffle section 103 and the second shuffle section 104 and the order of rearrangement may be set according to actual demand. In other words, a predetermined rule may be determined according to actual demand, and the number and rearrangement order of the first shuffle sections 103 and second shuffle sections 104 may be determined according to the predetermined rule.

As shown in Figure 3, "type" represents the type of each layer of the feature extraction section, "channel parameters" (filters) represents the size of the channel, and "size" represents the parameters processed by each layer. ``stride'' represents the stride of each layer, and ``output'' represents the size of the output feature.

In this embodiment, the channel parameters, size, stride and output feature size of each layer may be determined according to actual demand.

Further, as shown in FIG. 3, a combination of numbers and "x" represents the number of repeated arrangements of the layer; for example, "7x" represents that seven corresponding layers are repeatedly arranged; 3×” indicates that three corresponding layers are repeatedly arranged.

As shown in FIG. 3, the first convolutional layer 101 extracts features from the input image, the extracted features are input to the pooling layer 102, where pooling processing is performed, and the pooled features are processed in order. The rearranged features are input to a plurality of first shuffle units 103 and a plurality of second shuffle units 104 for shuffling processing, and the extracted features are output to a detection unit 200 for detection.

The configurations of the first shuffle section 103 and the second shuffle section 104 will be exemplified below.

FIG. 4 is a diagram showing the first shuffle section 103 according to the first embodiment of the present invention. As shown in FIG. 4, the first shuffle unit 103 includes a first channel division module 401, a second convolutional layer 402, a first depthwise separable convolutional layer 403, a third convolutional layer 404, and a first merging module 405. , and a first shuffle module 406 .

The first channel division module 401 divides the feature input to the first shuffle unit 103 into a first partial feature and a second partial feature.

The second convolutional layer 402 processes the second partial features.

The first depthwise separable convolutional layer 403 processes the second partial features processed by the second convolutional layer 402 .

The third convolutional layer 404 processes the second partial features processed by the first depthwise separable convolutional layer 403 .

The first merging module 405 merges the first partial feature and the second partial feature processed by the third convolutional layer 404 .

The first shuffle module 406 performs shuffling processing on the merged first partial features and second partial features.

As shown in FIG. 4, the input features are divided into two parts, namely a first partial feature and a second partial feature, by a first channel splitting module 401. The first partial feature is input to the first merging module 405 via the left branch without any processing. The second partial feature takes the right branch and is first input into the 1×1 second convolutional layer 402 . The features output by the second convolutional layer 402 are input to a 3×3 first depthwise separable convolutional layer 403 after being subjected to normalization and activation processing. The features acquired by the first depthwise separable convolutional layer 403 are input to a 1×1 third convolutional layer 404 after normalization. The features output by the third convolutional layer 404 are output to the first merging module 405 after normalization and activation processing. The first merging module 405 merges the first partial feature and the second partial feature and inputs the merged feature to the first shuffle module 406 . The first shuffle module 406 performs shuffle processing on the merged first partial feature and second partial feature and outputs the result.

FIG. 5 is a diagram showing the second shuffle section 104 according to the first embodiment of the present invention. As shown in FIG. 5, the input features include a third partial feature and a fourth partial feature, and the second shuffle unit 104 includes a second depthwise separable convolutional layer 501, a fourth convolutional layer 502, a fifth It includes a convolutional layer 503 , a third depthwise separable convolutional layer 504 , a sixth convolutional layer 505 , a second merging module 506 , and a second shuffling module 507 .

A second depthwise separable convolutional layer 501 processes the third partial feature.

The fourth convolutional layer 502 processes the third partial feature processed by the second depthwise separable convolutional layer 501.

The fifth convolutional layer 503 processes the fourth partial feature.

The third depthwise separable convolutional layer 504 processes the fourth partial feature processed by the fifth convolutional layer 503.

The sixth convolutional layer 505 processes the fourth partial feature processed by the third depthwise separable convolutional layer 504 .

A second merging module 506 merges the third partial feature processed by the fourth convolutional layer 502 and the fourth partial feature processed by the sixth convolutional layer 505.

The second shuffle module 507 performs shuffling processing on the merged third and fourth partial features.

As shown in FIG. 5, the input third partial feature enters the left branch path and is first input to the 3×3 second depthwise separable convolution layer 501. The features output by the second depthwise separable convolutional layer 501 are input to the 1×1 fourth convolutional layer 502 after being normalized. The features output by the fourth convolutional layer 502 are output to the second merging module 506 after being normalized and activated. The inputted fourth partial feature enters the right branch path and is first inputted to the 1×1 fifth convolutional layer 503. The features output by the fifth convolutional layer 503 are normalized and activated, and then input to a 3×3 third depthwise separable convolutional layer 504. The features output by the third depthwise separable convolutional layer 504 are input to a 1×1 sixth convolutional layer 505 after normalization. The features output by the sixth convolutional layer 505 are output to the second merging module 506 after being normalized and activated. The second merging module 506 merges the third partial feature and the fourth partial feature and inputs the merged feature to the second shuffle module 507. The second shuffle module 507 performs shuffle processing on the merged third partial feature and fourth partial feature and outputs the result.

In this embodiment, the first convolutional layer 101, the second convolutional layer 402, the third convolutional layer 404, the fourth convolutional layer 502, the fifth convolutional layer 503, and the sixth convolutional layer 505 may be ordinary convolutional layers. good. The first depthwise separable convolutional layer 403, the second depthwise separable convolutional layer 501, and the third depthwise separable convolutional layer 504 are conventional depthwise separable convolutional layers. You can.

The number of input and output channels of each convolutional layer above is the same, i.e. each convolutional layer does not need to perform feature expansion and compression, reducing memory and performance requirements on the processor and increasing processing speed. can be improved.

The configuration of the feature extraction unit 100 of this embodiment has been described above as an example.

After the feature extraction unit 100 extracts features from the input image, the detection unit 200 detects objects in the input image based on the features extracted by the feature extraction unit 100.

In this embodiment, the detection unit 200 may use a conventional network structure, for example, the detection unit 200 includes a YOLO (You Only Look Once) network. The YOLO network detects objects in the input image based on the extracted features. The principle and process of object detection by the YOLO network may refer to the prior art, and the description thereof will be omitted here.

Table 1 shows a comparison of parameters between the object detection device according to the embodiment of the present invention and a conventional network.

As shown in Table 1, the first column is the parameters of the conventional YOLO-Darknet 53, the second column is the parameters of the object detection device 10 of this embodiment, and the third column is the parameters of the object detection device 10 of this embodiment. ' is a parameter of '. Here, the object detection device 10 and the object detection device 10' have the same structure, but have different parameters. For example, the channel parameter of the object detection device 10' is smaller than that of the object detection device 10. FLOPs represents the number of operations performed per second, CPU represents the processing speed on the CPU, GPU represents the processing speed on the GPU, mAP represents the average recognition accuracy, and APperson represents the accuracy of person recognition. APbicycle represents the recognition accuracy of a bicycle, APcar represents the recognition accuracy of a car, APbus represents the recognition accuracy of a bus, APvan represents the recognition accuracy of a box-shaped truck, and APtruck represents the recognition accuracy of a flat truck. Represents the recognition accuracy of type trucks. As can be seen from Table 1, the object detection device 10 and the object detection device 10' of this embodiment maintain substantially the same recognition accuracy as the YOLO-Darknet53 network, while requiring less memory for the processor than the YOLO-Darknet53 network. and performance requirements can be significantly reduced.

According to this embodiment, since the number of input channels and output channels of all convolutional layers in the shuffle unit for feature extraction is the same, there is no need to expand and compress features, and the memory and performance of the processor are reduced. requirements and increase processing speed. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, so it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

<Example 2>
The embodiment of the present invention further provides a terminal device, and FIG. 6 is a diagram showing the terminal device according to the second embodiment of the present invention. As shown in FIG. 6, the terminal device 600 includes an object detection device 601, and the object detection device 601 is the same as that described in Example 1, and its description will be omitted here.

FIG. 7 is a block diagram showing a system configuration of a terminal device according to a second embodiment of the present invention. As shown in FIG. 7, the terminal device 700 may include a central processing unit (central control unit) 701 and a storage device 702, and the storage device 702 is connected to the central processing unit 701. The diagram is merely exemplary and other types of configurations may be used to supplement or replace the configuration to implement telecommunications or other functions.

As shown in FIG. 7, the terminal device 700 may further include an input unit 703, a display 704, and a power source 705.

In one aspect, the functionality of the object detection device of Example 1 may be integrated into central processing unit 701. Here, the central processing unit 701 is configured to have a feature extraction unit extract features in the input image, and a detection unit to detect an object in the input image based on the features extracted by the feature extraction unit. Good too. Here, the feature extraction unit includes at least one shuffling unit, the shuffling unit includes a plurality of convolutional layers, and the number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers are and the plurality of convolutional layers includes at least one depthwise separable convolutional layer.

For example, the at least one shuffle section includes at least one first shuffle section and/or at least one second shuffle section, the first shuffle section is a shuffle section with a stride of 1, and the second shuffle section 10. The method of claim 9, wherein the section is a shuffle section with a stride of 2.

In another aspect, the object detection devices described in Example 1 may be configured with the central processing unit 701, for example, the object detection devices are chips connected to the central processing unit 701, and the central processing unit The functions of the object detection device may be realized by controlling the object detection device 701.

The terminal device 700 in this embodiment does not need to include all the components shown in FIG. 7.

As shown in FIG. 7, a central processing unit 701, also referred to as a controller or operating control unit, may include a microprocessor or other processing and/or logic device, and the central processing unit 701 receives input and Controls the operation of each part of the device 700.

Storage device 702 may be, for example, one or more of a buffer, flash memory, hard disk, removable media, volatile memory, nonvolatile memory, or other suitable device. Further, the central processing unit 701 may execute a program stored in the storage device 702 to realize storage or processing of information. Since other members are similar to the prior art, their description will be omitted here. Each part of the terminal device 700 may be implemented by specific hardware, firmware, software, or a combination thereof without departing from the scope of the present invention.

According to this embodiment, since the number of input channels and output channels of all convolutional layers in the shuffle unit for feature extraction is the same, there is no need to expand and compress features, and the memory and performance of the processor are reduced. requirements and increase processing speed. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, so it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

<Example 3>
The embodiment of the present invention further provides an object detection method, which corresponds to the object detection device described in the first embodiment. FIG. 8 is a diagram showing an object detection method according to Example 3 of the present invention. As shown in FIG. 8, the method includes the following steps.

Step 801: The feature extractor extracts features in the input image.

Step 802: The detection unit detects an object in the input image based on the features extracted by the feature extraction unit.

Here, the feature extraction unit includes at least one shuffle unit, the shuffle unit includes a plurality of convolutional layers, and the number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers are the same. , the plurality of convolutional layers includes at least one depthwise separable convolutional layer.

In this embodiment, the specific implementation method of each of the above steps is the same as that described in the first embodiment, and the explanation thereof will be omitted here.

According to this embodiment, since the number of input channels and output channels of all convolutional layers in the shuffle unit for feature extraction is the same, there is no need to expand or compress features, and the memory and performance of the processor is reduced. requirements and increase processing speed. In addition, since the shuffle section has at least one depthwise separable convolutional layer, it significantly reduces the memory and performance requirements for processors such as FLOPs, and compared to networks such as YOLO-Darknet53, the recognition Processor requirements can be significantly reduced while accuracy is substantially maintained. Therefore, since it is possible to provide a detection method that is lightweight, has a high processing speed, and has high recognition accuracy, it can be applied to terminal devices with limited memory and performance, and excellent recognition effects can be obtained.

An embodiment of the present invention is a computer-readable computer-readable device that causes a computer to execute the object detection method described in the third embodiment in the object detection device or terminal device when the program is executed in the object detection device or terminal device. Offer more programs.

The embodiment of the present invention further provides a storage medium that stores a computer-readable program for causing a computer to execute the object detection method described in the third embodiment in an object detection device or a terminal device.

The object detection method executed in the object detection device or terminal device described with reference to the embodiments of the present invention may be implemented in hardware, a software module executed by a processor, or a combination of both. For example, one or more of the functional block diagrams shown in FIG. 1, or one or more combinations of functional block diagrams, may correspond to each software module of a computer program flow, or each hardware module may You may respond. These software modules may correspond to each step shown in FIG. 8, respectively. These hardware modules may be realized by converting these software modules into hardware using, for example, a field programmable gate array (FPGA).

The software modules may be located in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disks, mobile hard disks, CD-ROMs or any other form of storage medium known to those skilled in the art. The storage medium may be coupled to the processor such that the processor reads information from, and writes information to, the storage medium or may be a component of the processor. The processor and storage medium are located in an ASIC. The software module may be stored in the memory of the mobile terminal or on a memory card inserted into the mobile terminal. For example, if the terminal equipment uses a relatively large capacity MEGA-SIM card or a large capacity flash memory device, the software module may be stored on the MEGA-SIM card or large capacity flash memory device.

One or more functional blocks and/or one or more combinations of functional blocks in the functional block diagram described in FIG. 1 may include a general purpose processor, a digital signal processor ( DSP), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or any suitable combination thereof. may be done. One or more functional blocks and/or one or more combinations of functional blocks in the functional block diagram depicted in FIG. It may be implemented in a combination of processors, one or more microprocessors in combination with DSP communications, or any other configuration.

Although the present invention has been described above with reference to specific embodiments, the above description is merely illustrative and does not limit the scope of protection of the present invention. Various modifications and changes may be made to the present invention without departing from the spirit and principles of the present invention, and these modifications and changes also fall within the scope of the present invention.

Claims (8)

An object detection device,
a feature extraction unit that extracts features in the input image;
a detection unit that detects an object in the input image based on the features extracted by the feature extraction unit,
The feature extraction unit includes at least one shuffle unit,
The shuffle section includes a plurality of convolutional layers,
The number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers are the same,
The plurality of convolutional layers includes at least one depthwise separable convolutional layer,
The at least one shuffle section includes at least one first shuffle section and/or at least one second shuffle section,
The first shuffle part is a shuffle part with a stride of 1,
The second shuffle section is a shuffle section with a stride of 2 . The feature extraction unit is
a first convolutional layer that processes the input image;
2. A pooling layer that performs a pooling process on the features output by the first convolutional layer and inputs the pooled features to the first shuffle unit or the second shuffle unit. equipment. The first shuffle section is
a first channel division module that divides the feature input into the first shuffle unit into a first partial feature and a second partial feature;
a second convolutional layer processing the second partial feature;
a first depthwise separable convolutional layer that processes a second partial feature processed by the second convolutional layer;
a third convolutional layer that processes the second partial feature processed by the first depthwise separable convolutional layer;
a first merging module that merges the first partial feature and the second partial feature processed by the third convolutional layer;
3. The apparatus according to claim 1 , further comprising a first shuffle module that performs a shuffling process on the merged first partial feature and the second partial feature. The features input to the second shuffle unit include a third partial feature and a fourth partial feature,
The second shuffle section is
a second depthwise separable convolutional layer that processes the third partial feature;
a fourth convolutional layer that processes the third partial feature processed by the second depthwise separable convolutional layer;
a fifth convolutional layer that processes the fourth partial feature;
a third depthwise separable convolutional layer that processes the fourth partial feature processed by the fifth convolutional layer;
a sixth convolutional layer that processes the fourth partial feature processed by the third depthwise separable convolutional layer;
a second merging module that merges a third partial feature processed by the fourth convolutional layer and a fourth partial feature processed by the sixth convolutional layer;
4. The apparatus according to claim 1, further comprising a second shuffle module that performs shuffling processing on the merged third partial feature and the fourth partial feature. The at least one shuffle section includes a plurality of the first shuffle sections and a plurality of the second shuffle sections,
5. The apparatus according to claim 1, wherein the plurality of first shuffle sections and the plurality of second shuffle sections are rearranged according to a predetermined rule. The apparatus according to any one of claims 1 to 5 , wherein the detection unit includes a YOLO network. A terminal device comprising the device according to any one of claims 1 to 6 . An object detection method, comprising:
a step in which the feature extraction unit extracts features in the input image;
a detection unit detecting an object in the input image based on the features extracted by the feature extraction unit,
The feature extraction unit includes at least one shuffle unit,
The shuffle section includes a plurality of convolutional layers,
The number of input channels and the number of output channels of each convolutional layer of the plurality of convolutional layers are the same,
The plurality of convolutional layers includes at least one depthwise separable convolutional layer,
The at least one shuffle section includes at least one first shuffle section and/or at least one second shuffle section,
The first shuffle part is a shuffle part with a stride of 1,
The method , wherein the second shuffle section is a shuffle section with a stride of 2 .

Images