Image recognition method and device, storage medium and electronic device

Abstract

The application discloses an image recognition method and device, a storage medium and an electronic device. The method comprises the following steps: obtaining a target feature map; performing a first convolution operation in a channel dimension on the target feature map to obtain a first feature map in a first channel dimension; performing a second convolution operation in a space dimension on the first feature map to obtain a second feature map; performing a third convolution operation in the space dimension on the second feature map to obtain a third feature map; performing a fourth convolution operation in the channel dimension on the third feature map to obtain a fourth feature map in a second channel dimension; and performing image recognition on a target image by using the fourth feature map. The method can be applied to intelligent traffic, auxiliary driving and the like, and can involve image recognition, artificial intelligence and the like. The application solves the technical problem that related technologies lack an image recognition mode considering both precision and calculation amount.

Description

Technical Field

[0001] This application relates to the field of computers, and more specifically, to an image recognition method, apparatus, storage medium, and electronic device. Background Technology

[0002] With the development of convolutional neural networks, image recognition applications such as semantic segmentation, depth estimation, counting, and pose estimation have also developed rapidly. When these applications are implemented in products, they require high accuracy to ensure a good user experience. On the other hand, when these image recognition applications are used in low-computing-power platforms such as mobile phones, in-vehicle systems, drones, and robots, the algorithms often need to have low computational load to ensure smooth operation. However, there is a conflict between accuracy and computational load; higher accuracy requires greater computation. Related technologies have not yet provided an image recognition method that balances accuracy and computational load. Therefore, related technologies suffer from the problem of lacking an image recognition method that balances accuracy and computational load.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This application provides an image recognition method, apparatus, storage medium, and electronic device to at least address the technical problem of related technologies lacking an image recognition method that balances accuracy and computational load.

[0005] According to one aspect of the embodiments of this application, an image recognition method is provided, comprising: acquiring a target feature map, wherein the target feature map is used to represent features corresponding to a target image to be recognized; performing a first convolution operation in the channel dimension on the target feature map to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; performing a second convolution operation in the spatial dimension on the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is larger than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is larger than the initial resolution of the target feature map; performing a third convolution operation in the spatial dimension on the second feature map to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; performing a fourth convolution operation in the channel dimension on the third feature map to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; and performing image recognition on the target image using the fourth feature map.

[0006] According to another aspect of the embodiments of this application, an image recognition apparatus is also provided, comprising: an acquisition unit, configured to acquire a target feature map, wherein the target feature map represents features corresponding to a target image to be recognized; a first execution unit, configured to perform a first convolution operation in the channel dimension on the target feature map to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; and a second execution unit, configured to perform a second convolution operation in the spatial dimension on the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is larger than the initial spatial dimension of the target feature map. The spatial resolution of the second feature map is greater than the initial resolution of the target feature map; the third execution unit is used to perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; the fourth execution unit is used to perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; the recognition unit is used to perform image recognition on the target image using the fourth feature map.

[0007] As an optional embodiment, the second execution unit includes: a first execution device, configured to perform up-row convolution operations on the first feature map in both high and wide dimensions to obtain a first output feature map, wherein the high dimension of the first output feature map is greater than the initial high dimension of the target feature map, the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map, and the spatial resolution of the first output feature map is greater than the initial resolution; and a first acquisition device, configured to acquire the second feature map based on the first output feature map. The third execution unit includes: a second execution device, configured to perform down-row convolution operations on the second feature map in both high and wide dimensions to obtain a second output feature map, wherein the high dimension of the second output feature map is equal to the initial high dimension, the wide dimension of the second output feature map is equal to the initial wide dimension, and the spatial resolution of the second output feature map is equal to the spatial resolution of the first output feature map; and a second acquisition device, configured to acquire the third feature map based on the second output feature map.

[0008] As an optional embodiment, the first execution device includes: a first execution sub-device, configured to perform the high-dimensional first target convolution operation on the first feature map to obtain a first target feature map; a second execution sub-device, configured to perform the wide-dimensional second target convolution operation on the first target feature map to obtain a second target feature map; or, a third execution sub-device, configured to perform the wide-dimensional second target convolution operation on the first feature map to obtain a third target feature map; and a fourth execution sub-device, configured to perform the high-dimensional first target convolution operation on the third target feature map to obtain a fourth target feature map; the first acquisition device includes: a first acquisition sub-device, configured to determine the second target feature map as the second feature map when the first output feature map includes the second target feature map; or, a second acquisition sub-device, configured to determine the fourth target feature map as the second feature map when the first output feature map includes the fourth target feature map.

[0009] As an optional solution, the first execution device includes: a first execution sub-device, configured to perform the high-dimensional first target convolution operation on the first feature map to obtain a first target feature map; and a third execution sub-device, configured to perform the wide-dimensional second target convolution operation on the first feature map to obtain a third target feature map; the first acquisition device includes: a third acquisition sub-device, configured to determine the first target feature map and the third target feature map as the second feature map; or, a fourth acquisition sub-device, configured to integrate the first target feature map and the third target feature map to obtain the second feature map.

[0010] As an optional embodiment, the second execution device includes: a fifth execution sub-device, configured to perform the high-dimensional first downlink convolution operation on the first target feature map after determining the first target feature map and the third target feature map as the second feature map, to obtain a first output sub-feature map; and a sixth execution sub-device, configured to perform the wide-dimensional second downlink convolution operation on the third target feature map after determining the first target feature map and the third target feature map as the second feature map, to obtain a second output sub-feature map; and a fifth acquisition sub-device, configured to integrate the first output sub-feature map and the second output sub-feature map after determining the first target feature map and the third target feature map as the second feature map, to obtain the second output feature map.

[0011] As an optional solution, the above apparatus further includes: a third acquisition device, used to acquire the target image to be identified before acquiring the target feature map; a fourth acquisition device, used to input the target image into a downsampling module to obtain the low-level features corresponding to the target image before acquiring the target feature map; a fifth acquisition device, used to input the low-level features into a residual convolution module to obtain the global high-level semantic features corresponding to the target image before acquiring the target feature map; and a sixth acquisition device, used to acquire the target feature map based on the global high-level semantic features before acquiring the target feature map.

[0012] As an optional solution, the above-mentioned apparatus further includes: a seventh acquisition device, used to acquire multiple consecutive images to be identified collected by the target terminal before acquiring the target feature map, wherein the computing power value of the target terminal is less than or equal to a preset threshold; and an eighth acquisition device, used to determine the target image from the multiple consecutive images to be identified in response to a task execution request triggered by the target terminal before acquiring the target feature map, wherein the task execution request is used to request the target terminal to perform a dense estimation visual perception task, the dense estimation visual perception task being a task of performing image recognition on the target image to obtain a dense estimation result.

[0013] According to another aspect of the embodiments of this application, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the image recognition method described above.

[0014] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the image recognition method described above through the computer program.

[0015] In this embodiment, a target feature map is obtained, wherein the target feature map represents the features corresponding to the target image to be identified; a first convolution operation in the channel dimension is performed on the target feature map to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; a second convolution operation in the spatial dimension is performed on the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is larger than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is larger than the initial resolution of the target feature map; a third convolution operation in the spatial dimension is performed on the second feature map to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; a fourth convolution operation in the channel dimension is performed on the third feature map to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; and image recognition is performed on the target image using the fourth feature map.

[0016] Using the above image recognition method, firstly, by performing a convolution operation on the channel dimension of the target feature map, a first feature map with a lower channel dimension is obtained, thereby reducing channel redundancy. Since the network computation is lower for feature maps with lower channel dimensions, the technical effect of reducing network computation is achieved.

[0017] Secondly, a spatial dimension convolution operation is performed on the first feature map to obtain the second feature map. This not only increases the spatial dimension of the feature map but also improves its spatial resolution, thereby increasing the range of information received by the neurons on the feature map.

[0018] Furthermore, a spatial dimension convolution operation is performed on the second feature map. While keeping the spatial resolution of the feature map unchanged, a third feature map with the same spatial dimension as the first feature map is output. This achieves the goal of increasing the range of information received by neurons on the feature map without changing the spatial dimension of the feature map, thereby improving the technical effect of computational accuracy.

[0019] Furthermore, a convolution operation is performed again on the third feature map along the channel dimension, outputting a fourth feature map that has the same channel and spatial dimensions as the target feature map but with higher spatial resolution. The fourth feature map is then used to perform image recognition on the target image. This achieves the goal of reducing network computation and improving spatial resolution without changing the original feature dimensions of the target feature map. Thus, it achieves the effect of balancing accuracy and computation in the image recognition process, thereby solving the technical problem of the lack of image recognition methods that balance accuracy and computation in related technologies. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0021] Figure 1 This is a schematic diagram of an application environment for an optional image recognition method according to an embodiment of this application;

[0022] Figure 2 This is a schematic diagram of the flow of an optional image recognition method according to an embodiment of this application;

[0023] Figure 3 This is a schematic diagram of an optional image recognition method according to an embodiment of this application;

[0024] Figure 4 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0025] Figure 5 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0026] Figure 6 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0027] Figure 7 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0028] Figure 8 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0029] Figure 9 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0030] Figure 10 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0031] Figure 11 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0032] Figure 12 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0033] Figure 13 This is a schematic diagram of another optional image recognition method according to an embodiment of this application;

[0034] Figure 14This is a schematic diagram of an optional image recognition device according to an embodiment of this application;

[0035] Figure 15 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation

[0036] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0037] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0038] To facilitate understanding, the following explanations are provided for the terms:

[0039] Intelligent Traffic Systems (ITS), also known as Intelligent Transportation Systems, effectively integrate advanced science and technology (information technology, computer technology, data communication technology, sensor technology, electronic control technology, automatic control theory, operations research, artificial intelligence, etc.) into transportation, service control, and vehicle manufacturing. This strengthens the connection between vehicles, roads, and users, thereby forming a comprehensive transportation system that ensures safety, improves efficiency, enhances the environment, and saves energy.

[0040] Intelligent Vehicle Infrastructure Cooperative Systems (IVICS) are a development direction of Intelligent Transportation Systems (ITS). IVICS utilizes advanced wireless communication and next-generation Internet technologies to implement comprehensive, real-time dynamic information exchange between vehicles and infrastructure. Based on the collection and fusion of dynamic traffic information across all times and spaces, it conducts active vehicle safety control and cooperative road management, fully realizing effective collaboration between people, vehicles, and roads. This ensures traffic safety, improves traffic efficiency, and ultimately forms a safe, efficient, and environmentally friendly road traffic system.

[0041] Convolution: used to describe the influence of past actions on the present, that is, the superposition of a spatiotemporal response; for example, when an earthquake occurs and seismic waves propagate outward, convolution is needed to calculate the signal received at any point in space.

[0042] Upsampling: Enlarging an image (also known as upsampling or image interpolation) primarily aims to enlarge the original image so that it can be displayed on a higher resolution display device.

[0043] Downsampling: Reducing the size of an image (also called subsampled or downsampled) is mainly for the purpose of making the image fit the size of the display area and generating a thumbnail of the corresponding image.

[0044] Multidirectional convolution: Multidirectional convolution is relative to traditional convolution. In traditional convolution, the convolution kernel is fixed in the channel dimension and then slides along the height and width dimensions, making convolution "unidirectional". In multidirectional convolution, the convolution kernel is not only fixed in the channel dimension, but is fixed in the channel, height and width dimensions in sequence, and then slides along the other two dimensions.

[0045] Receptive field: The range of information a neuron in a feature map receives in an image. The larger the receptive field, the more contextual information it can acquire.

[0046] According to one aspect of the embodiments of this application, an image method is provided. Optionally, as an optional implementation, the above-described image recognition method may be applied to, but is not limited to, [examples of image recognition methods]. Figure 1 The environment shown may include, but is not limited to, user equipment 102 and server 112. User equipment 102 may include, but is not limited to, mobile phones, computers, smart voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc., and may include, but is not limited to, display 104, processor 106 and memory 108. Server 112 includes database 114 and processing engine 116.

[0047] The specific process can be summarized in the following steps:

[0048] Step S102: User equipment 102 obtains the target feature map corresponding to the target image to be identified based on the data collected by UAV 118 during the execution of a certain dense estimation task. The dense estimation task can be, but is not limited to, computer vision tasks that require high-resolution representation, such as semantic segmentation, depth estimation, counting, and pose estimation.

[0049] Steps S104-S106: Send the image recognition request to the server 112 via network 110;

[0050] In step S108, server 112 responds to the above image recognition request and performs a first convolution operation on the target feature map in the channel dimension through the processing engine to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map.

[0051] Step S110: Perform a second convolution operation on the spatial dimension of the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map.

[0052] Step S112: Perform a third convolution operation on the spatial dimension of the second feature map to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map.

[0053] Step S114: Perform a fourth convolution operation on the third feature map of the channel dimension to obtain a fourth feature map of the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension.

[0054] Step S116: Using the fourth feature map described above, perform image recognition on the target image.

[0055] In steps S118-S120, the image recognition result is sent to the user equipment 102 via the network 110. The user equipment 102 uses the information from the image recognition result through the processor 106 to determine the completion result of the dense estimation task, displays the completion result of the dense estimation task on the display 108, and stores the image recognition result used to determine the completion status of the dense estimation task in the memory 104.

[0056] remove Figure 1Beyond the examples shown, the above steps can be completed independently by the client or server, or collaboratively by both, such as by user equipment 102 performing the first, second, third, and fourth convolution operations, thereby reducing the processing load on server 112. User equipment 102 includes, but is not limited to, handheld devices (such as mobile phones), computers, intelligent voice interaction devices, smart home appliances, vehicle terminals, aircraft, etc. This application does not limit the specific implementation of user equipment 102.

[0057] Alternatively, as an alternative implementation method, such as Figure 2 As shown, image recognition methods include:

[0058] Step S202: Obtain the target feature map, wherein the target feature map is used to represent the features corresponding to the target image to be identified;

[0059] Step S204: Perform a first convolution operation on the target feature map in the channel dimension to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map;

[0060] Step S206: Perform a second convolution operation on the spatial dimension of the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map.

[0061] Step S208: Perform a third convolution operation on the second feature map in terms of spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map.

[0062] Step S210: Perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, where the second channel dimension is equal to the initial channel dimension.

[0063] Step S212: Use the fourth feature map to perform image recognition on the target image.

[0064] Optionally, in this embodiment, the above image recognition method can be applied, but is not limited to, in scenarios where there is a high dependence on high-resolution representation and low computing power, such as using convolutional neural networks (CNNs) to achieve computer vision tasks such as semantic segmentation, depth estimation, counting, and pose estimation.

[0065] On the one hand, when using convolutional neural networks (CNNs) to perform computer vision tasks, since convolution is a local operation, each convolution can only capture features within the size of the previous layer's kernel. To allow the network to acquire a larger receptive field, existing CNNs typically employ downsampling of feature maps, so that each convolution on a smaller feature map corresponds to a larger receptive field in the original image. Simultaneously, reducing the feature map resolution decreases the computational cost of the network. To capture higher-dimensional features, existing CNNs typically increase the number of network channels after downsampling.

[0066] On the other hand, as the number of network channels increases, the resolution of the feature maps becomes lower and lower, which results in the loss of a large amount of spatial information and greatly reduces the effectiveness of dense estimation tasks.

[0067] To address the aforementioned issues, existing technologies employ two approaches. On one hand, there are methods using decoders to recover spatial details from downsampled feature maps. While these methods can recover detailed features to some extent, they struggle to recover all high-resolution information effectively. On the other hand, there are methods using dilated convolutions to obtain a larger receptive field without reducing feature map resolution. Although this approach can achieve a larger receptive field without lowering resolution, it is based on traditional CNN operations, resulting in a high computational cost. Furthermore, the acquired contextual information remains localized, and using contextual information modules would further increase the computational burden.

[0068] Using the aforementioned image recognition method, on the one hand, convolution operations are used based on the initial channel dimension of the feature map to obtain a feature map with a lower channel dimension, thereby reducing channel redundancy and decreasing the computational load of the network. On the other hand, firstly, convolution operations are used based on the initial spatial dimension of the feature map to obtain a feature map with a higher spatial dimension. During this convolution operation, as the spatial dimension increases, the contextual information that neurons on the feature map can acquire increases accordingly, and the spatial resolution also increases accordingly. Secondly, convolution operations are used to obtain a feature map with the initial spatial dimension. Since the contextual information acquired by neurons on the feature map remains unchanged during this convolution operation, the spatial resolution is still the spatial resolution corresponding to the feature map with a higher spatial dimension. That is, using multi-directional convolution in the spatial dimension, all information in each dimension is acquired and then fused into global information, solving the problem of insufficient receptive field. In addition, multi-directional convolution exists in the spatial dimension, so increasing the convolution kernel can increase the feature map resolution, thereby enhancing the high-resolution representation of the network.

[0069] To further illustrate, such as Figure 3As shown, in a scenario where a drone 302 is used to perform image recognition tasks, for example, in the energy sector, the drone 302 is used to acquire thermal images of solar panels and detect, locate, and classify defects in the solar panels; in the agricultural sector, the drone 302 is used to acquire spectral images of plant leaves and detect and determine pathogens or diseases present in the plants; or in the public safety sector, the drone 302 is used to acquire images of sea areas and detect sharks in the water.

[0070] It should be noted that, on the one hand, UAV 302 often relies heavily on high-resolution representation in the process of recognizing, detecting, and judging images in various fields; on the other hand, aircraft 302 often do not have high-configuration computing power.

[0071] It should be noted that the aircraft 302 can determine the target feature map corresponding to the target image based on the target image, but is not limited to this application. This application does not limit the specific acquisition method of the target image and the target feature map corresponding to the target image.

[0072] Using the image recognition method described above, after acquiring the target feature map, the aircraft 302 performs a first convolution operation on the target feature map in the channel dimension to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; performs a second convolution operation on the first feature map in the spatial dimension to obtain a second feature map, wherein the first spatial dimension of the second feature map is larger than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is larger than the initial resolution of the target feature map; performs a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; performs a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; and uses the fourth feature map to perform image recognition on the target image.

[0073] It should be noted that, by using the above image recognition method, the aircraft 302, in the process of performing image recognition tasks, achieves the purpose of reducing network computation and improving spatial resolution by performing convolution operations on the channel dimension and spatial dimension without changing the original feature dimension of the target feature map. This achieves the effect of simultaneously balancing accuracy and computation in the image recognition process, and thus solves the technical problem that related technologies lack image recognition methods that balance accuracy and computation.

[0074] Optionally, in this embodiment, the above-described image recognition method can be, but is not limited to, applied in the image recognition system 402, such as... Figure 4As shown, the image recognition system 402 may include, but is not limited to, a first convolution module, a second convolution module, a third convolution module, a fourth convolution module, and an image recognition module.

[0075] It should be noted that the first convolution module may, but is not limited to, perform a first convolution operation on the channel dimension on the input target feature map to output a first feature map of the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map.

[0076] It should be noted that the second convolution module may, but is not limited to, perform a second convolution operation on the input first feature map in a spatial dimension to output a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map.

[0077] It should be noted that the third convolution module can, but is not limited to, perform a third convolution operation on the input second feature map in terms of spatial dimensions to output a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map.

[0078] It should be noted that the fourth convolution module can, but is not limited to, perform a fourth convolution operation on the input third feature map in the channel dimension to output a fourth feature map in the second channel dimension, where the second channel dimension is equal to the initial channel dimension.

[0079] It should be noted that the image recognition module can be used, but is not limited to, to perform image recognition on the target image using the fourth feature map.

[0080] Optionally, in this embodiment, the target feature map may be used to represent the features corresponding to the target image. The features corresponding to the target image may include, but are not limited to, color features, shape features, texture features, spatial relationship features, etc. The target feature map may be obtained based on the target image and implemented using methods such as histogram of oriented gradients, local binary pattern features, and Haar features. This application does not limit the specific acquisition method of the target image and the target feature map corresponding to the target image.

[0081] Optionally, in this embodiment, the feature map including the target feature map may include, but is not limited to, multiple feature map dimensions. The multiple feature map dimensions may include, but are not limited to, channel dimensions and spatial dimensions. The channel dimension may include, but is not limited to, the feature dimension used to indicate the image data output. The spatial dimension may include, but is not limited to, the feature dimension used to indicate the image spatial resolution. The spatial dimension may include, but is not limited to, the high dimension and the wide dimension.

[0082] Optionally, in this embodiment, the first convolution operation may be used, but is not limited to, to reduce the channel dimension of the target feature map to obtain a first feature map corresponding to the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map.

[0083] It should be noted that by using the first convolution operation described above to obtain a first feature map with a lower channel dimension than the target feature map, the purpose of reducing channel redundancy is achieved, thus realizing the technical effect of reducing the amount of network computation.

[0084] Optionally, in this embodiment, the channel dimension may include, but is not limited to, the dimension corresponding to a single input channel, the dimension corresponding to a single output channel, the dimension corresponding to multiple input channels, or the dimension corresponding to multiple output channels. The spatial dimension may include, but is not limited to, the height dimension and the width dimension. The spatial resolution may be used to indicate the resolution of the image within the current spatial dimension, and the spatial resolution may be used to indicate the pixel precision of the image within the current spatial dimension. The higher the spatial resolution, the higher the pixel precision of the image, that is, the clearer the image.

[0085] Optionally, in this embodiment, the second convolution operation may be used, but is not limited to, to improve the spatial dimension and spatial resolution of the first feature map, wherein the spatial resolution of the first feature map increases as the spatial dimension of the first feature map increases.

[0086] It should be noted that by using the above-mentioned second convolution operation, a second feature map with a larger spatial dimension than the first feature map can be obtained. In the process of increasing the spatial dimension of the feature map, the range of information received by the neurons on the feature map is improved, and the contextual information that can be obtained is increased (i.e., receptive field information). As a result, the second feature map has a higher spatial resolution than the first feature map, thereby achieving the goal of improving the image pixel accuracy and realizing the technical effect of improving the image recognition accuracy.

[0087] For example, taking a spatial dimension that includes both high and wide dimensions, when the spatial dimension of the first feature map is h1*w1, the corresponding first spatial resolution is k1; performing the second convolution operation on the spatial dimension yields a second feature map with a spatial dimension of 2h1*2w1, and the corresponding second spatial resolution is k2, where k2 is greater than k1.

[0088] Optionally, in this embodiment, the third convolution operation may be used, but is not limited to, to reduce the spatial dimension of the second feature map to the initial spatial dimension of the target feature map to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the aforementioned initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map.

[0089] It should be noted that after performing the third convolution operation, the spatial dimension of the resulting third feature map is restored to the initial spatial dimension of the target feature map before performing the second convolution operation, but the spatial resolution of the third feature map remains unchanged and is still equal to the spatial resolution of the second feature map. This achieves the goal of improving the spatial resolution of the target feature map while maintaining its initial spatial dimension as input, thereby improving the technical effect of image recognition accuracy.

[0090] It should be noted that during the process of restoring the spatial dimension of the second feature map to the initial spatial dimension of the target feature map, since the contextual information acquired by the neurons on the stored feature map remains unchanged, that is, the receptive field information does not change, the spatial resolution of the third feature map also does not change and is equal to the spatial resolution of the second feature map.

[0091] Optionally, in this embodiment, the fourth convolution operation may be used, but is not limited to, to increase the channel dimension of the third feature image to the initial channel dimension of the target feature map to obtain a fourth feature map, wherein the second channel dimension of the fourth feature map is equal to the aforementioned initial channel dimension.

[0092] It should be noted that after performing the fourth convolution operation, the channel dimension of the resulting fourth feature map is restored to the initial channel dimension of the target feature map before performing the first convolution operation, but the spatial resolution of the fourth feature map remains unchanged and is still equal to the spatial resolution of the second feature map. This achieves the goal of improving the spatial resolution of the target feature map while maintaining its initial feature dimension as input, thereby improving the technical effect of image recognition accuracy.

[0093] It should be noted that the fourth convolution operation is a convolution operation on the channel dimension of the third feature map. Therefore, the spatial dimension information of the third feature map does not change. Thus, the spatial resolution of the fourth feature map also does not change and is equal to the spatial resolution of the third feature map.

[0094] Optionally, in this embodiment, the fourth feature map is used to perform image recognition on the target image to obtain the image recognition result.

[0095] It should be noted that the fourth feature map used for image recognition of the target image is a feature map with the same feature dimensions as the target feature map but with higher spatial resolution. That is, the image recognition method described above uses multi-directional convolution to enhance the spatial information redundancy of the convolutional features, thus realizing a convolutional neural network with powerful high-resolution representation. Furthermore, before performing the second and third convolution operations in the spatial dimension, the image method reduces channel redundancy through the first convolution operation, enabling the network to also have low computational complexity.

[0096] To further illustrate, such as Figure 5 As shown, assume that the feature map dimension of the target feature map 502 is c*h*w, where c is the channel dimension, w is the width dimension in the spatial dimension, and h is the height dimension in the spatial dimension.

[0097] Step S502: Perform a first convolution operation on the channel dimension of the target feature map 502 to reduce the channel dimension of the target feature map 502 and obtain a first feature map 504 with the first channel dimension. The target feature map 502 is used to represent the features corresponding to the target image to be identified. The feature map dimension of the first feature map 504 is (c / n)*h*w, where n is a real number greater than 1.

[0098] Step S504: Perform a second convolution operation on the spatial dimension of the first feature map 504 to increase the spatial dimension of the first feature map 504, and obtain a second feature map 506 with the first spatial dimension. The spatial dimension of the second feature map 506 is (c / n)*2h*2w, and the spatial resolution of the second feature map 506 is greater than the initial resolution of the target feature map 502.

[0099] Step S506: Perform a third convolution operation on the second feature map 506 in the spatial dimension to restore the spatial dimension of the second feature map 506 to the spatial dimension of the first feature map 504, and obtain a third feature map 508 in the second spatial dimension, wherein the spatial dimension of the third feature map 508 is (c / n)*h*w, and the spatial resolution of the third feature map 508 is equal to the spatial resolution of the second feature map 506.

[0100] Step S508: Perform a fourth convolution operation on the channel dimension of the third feature map 508 to restore the channel dimension of the third feature map 508 to the initial channel of the target feature map 502, and obtain the fourth feature map 510 with the second channel dimension. The feature map dimension of the fourth feature map 510 is c*h*w. The fourth feature map 510 is used for image recognition of the target image.

[0101] It should be noted that the use of the second convolution operation to double the height and width dimensions mentioned in step S504 is just an example, and increasing them to any multiple should be within the scope of protection of this invention.

[0102] The embodiments provided in this application involve obtaining a target feature map, wherein the target feature map represents the features corresponding to a target image to be identified; performing a first convolution operation in the channel dimension on the target feature map to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; performing a second convolution operation in the spatial dimension on the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is larger than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is larger than the initial resolution of the target feature map; performing a third convolution operation in the spatial dimension on the second feature map to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; performing a fourth convolution operation in the channel dimension on the third feature map to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; and using the fourth feature map, image recognition is performed on the target image.

[0103] Using the above image recognition method, firstly, by performing a convolution operation on the channel dimension of the target feature map, a first feature map with a lower channel dimension is obtained, thereby reducing channel redundancy. Since the network computation is lower for feature maps with lower channel dimensions, the technical effect of reducing network computation is achieved.

[0104] Secondly, a spatial dimension convolution operation is performed on the first feature map to obtain the second feature map. This not only increases the spatial dimension of the feature map but also improves its spatial resolution, thereby increasing the range of information received by the neurons on the feature map.

[0105] Furthermore, a spatial dimension convolution operation is performed on the second feature map. While keeping the spatial resolution of the feature map unchanged, a third feature map with the same spatial dimension as the first feature map is output. This achieves the goal of increasing the range of information received by neurons on the feature map without changing the spatial dimension of the feature map, thereby improving the technical effect of computational accuracy.

[0106] Furthermore, a convolution operation is performed again on the third feature map along the channel dimension, outputting a fourth feature map that has the same channel and spatial dimensions as the target feature map but with higher spatial resolution. The fourth feature map is then used to perform image recognition on the target image. This achieves the goal of reducing network computation and improving spatial resolution without changing the original feature dimensions of the target feature map. Thus, it achieves the effect of balancing accuracy and computation in the image recognition process, thereby solving the technical problem of the lack of image recognition methods that balance accuracy and computation in related technologies.

[0107] As an optional approach, a second convolution operation in the spatial dimension is performed on the first feature map to obtain a second feature map, including:

[0108] S1, perform up-convolution operations on the first feature map in both high and wide dimensions to obtain the first output feature map, wherein the high dimension of the first output feature map is greater than the initial high dimension of the target feature map, the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map, and the spatial resolution of the first output feature map is greater than the initial resolution.

[0109] S2, based on the first output feature map, obtain the second feature map.

[0110] Optionally, in this embodiment, the spatial dimension may include, but is not limited to, a high dimension and a wide dimension. The second convolution operation may include, but is not limited to, performing an up-convolution operation on the first feature map with high and wide dimensions. The up-convolution operation may be used to increase the high and wide dimensions of the first feature map to obtain a first output feature map. The high dimension of the first output feature map is greater than the initial high dimension of the target feature map, and the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map.

[0111] Optionally, in this embodiment, a second feature map is obtained based on the first output feature map.

[0112] The embodiments provided in this application perform uplink convolution operations on a first feature map in both high and wide dimensions to obtain a first output feature map. The high dimension of the first output feature map is greater than the initial high dimension of the target feature map, the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map, and the spatial resolution of the first output feature map is greater than the initial resolution. Based on the first output feature map, a second feature map is obtained. By performing uplink convolution operations on the first feature map in both high and wide dimensions, the high and wide dimensions of the first feature map, as well as the corresponding spatial resolution, are improved, thereby achieving the technical effect of improving image recognition accuracy.

[0113] As an optional approach, a third convolution operation in the spatial dimension is performed on the second feature map to obtain a third feature map, including:

[0114] S1, Perform down-convolution operations on the high-dimensional and wide-dimensional second feature map to obtain the second output feature map, wherein the high-dimensionality of the second output feature map is equal to the initial high-dimensionality, the wide-dimensionality of the second output feature map is equal to the initial wide-dimensionality, and the spatial resolution of the second output feature map is equal to the spatial resolution of the first output feature map.

[0115] S2, based on the second output feature map, obtain the third feature map.

[0116] Optionally, in this embodiment, the spatial dimension may include, but is not limited to, a high dimension and a wide dimension. The third convolution operation may include, but is not limited to, performing a down-convolution operation on the second feature map with both high and wide dimensions. The down-convolution operation may be used, but is not limited to, to reduce the high and wide dimensions of the second feature map to obtain a second output feature map. The high dimension of the second output feature map is equal to the initial high dimension of the target feature map, the wide dimension of the first output feature map is equal to the initial wide dimension of the target feature map, and the spatial resolution of the second output feature map is equal to the spatial resolution of the first output feature map.

[0117] Optionally, in this embodiment, a third feature map is obtained based on the second output feature map.

[0118] It should be noted that by using the above-mentioned upward convolution operation, a second feature map with a larger spatial dimension than the first feature map can be obtained. In the process of increasing the spatial dimension of the feature map, the range of information received by the neurons on the feature map is improved, and the contextual information that can be obtained is increased (i.e., receptive field information). As a result, the second feature map has a higher spatial resolution than the first feature map, thereby achieving the goal of improving image pixel accuracy and realizing the technical effect of improving image recognition accuracy.

[0119] It should be noted that during the execution of the aforementioned downlink convolution operation, as the spatial dimension of the second feature map is restored to the initial spatial dimension of the target feature map, the contextual information acquired by the neurons on the stored feature map remains unchanged, i.e., the receptive field information does not change. Therefore, the spatial resolution of the third feature map also remains unchanged, equal to the spatial resolution of the second feature map. This achieves the goal of improving the spatial resolution of the target feature map while maintaining its initial spatial dimension as input, thereby enhancing image recognition accuracy.

[0120] The embodiments provided in this application perform downlink convolution operations on the second feature map in both high and wide dimensions to obtain a second output feature map. The high dimension of the second output feature map is equal to the initial high dimension, the wide dimension is equal to the initial wide dimension, and the spatial resolution is equal to the spatial resolution of the first output feature map. A third feature map is then obtained based on the second output feature map. By performing downlink convolution operations on the second feature map in both high and wide dimensions, the high and wide dimensions of the second feature map are restored to the initial high and wide dimensions of the target feature map, while maintaining the same spatial resolution as the first output feature map. This achieves the technical effect of improving the spatial resolution of the target feature map without changing its original feature dimensions.

[0121] As an optional approach, an uplink convolution operation with high and wide dimensions is performed on the first feature map to obtain a first output feature map, including:

[0122] S1, Perform a high-dimensional first target convolution operation on the first feature map to obtain the first target feature map; perform a wide-dimensional second target convolution operation on the first target feature map to obtain the second target feature map; or,

[0123] S2, perform a wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map; perform a high-dimensional first target convolution operation on the third target feature map to obtain the fourth target feature map;

[0124] S3, based on the first output feature map, obtain the second feature map, including:

[0125] S4, if the first output feature map includes the second target feature map, then the second target feature map is determined as the second feature map; or,

[0126] S5, if the first output feature map includes the fourth target feature map, the fourth target feature map is determined as the second feature map.

[0127] Optionally, in this embodiment, the uplink convolution operation may be used, but is not limited to, to instruct the first target feature map to be subjected to a high-dimensional first target convolution operation to obtain a first target feature map, and then the first target feature map to be subjected to a wide-dimensional second target convolution operation to obtain a second target feature map.

[0128] Optionally, in this embodiment, the uplink convolution operation may be used, but is not limited to, to instruct the first target feature map to be subjected to a wide-dimensional second target convolution operation to obtain a third target feature map, and then to be subjected to a high-dimensional first target convolution operation on the third target feature map to obtain a fourth target feature map.

[0129] Optionally, in this embodiment, if the first output feature map includes the second target feature map, the second target feature map is determined as the second feature map.

[0130] Optionally, in this embodiment, if the first output feature map includes the fourth target feature map, the fourth target feature map is determined as the second feature map.

[0131] It should be noted that, in this embodiment, the first target convolution operation and the second target convolution operation are serial convolution operations. The up-row convolution operation is to further perform the second target convolution operation on the result obtained from the first target convolution operation, or to further perform the first target convolution operation on the result obtained from the second target convolution operation.

[0132] To further illustrate, such as Figure 6As shown, assume that the feature map dimension of the first feature map 602 is (c / n)*h*w, where (c / n) is the channel dimension, w is the width dimension in the spatial dimension, and h is the height dimension in the spatial dimension.

[0133] Step S602: Perform a high-dimensional uplink convolution operation on the first feature map 602 to obtain the first output feature map 604, wherein the feature map dimension of the first output feature map 604 is (c / n)*2h*w;

[0134] Step S604: Step S606 performs a wide-dimensional second target convolution operation on the first target feature map 604 to obtain the second target feature map 606, wherein the feature map dimension of the second target feature map 606 is (c / n)*2h*2w;

[0135] Step S606: Based on the second target feature map 606, determine the second feature map 608, wherein the feature map dimension of the second target feature map 608 is (c / n)*2h*2w.

[0136] It should be noted that the example of doubling the height and width dimensions in the uplink convolution operation is just that—an example. Any increase in these dimensions should be within the scope of protection of this invention.

[0137] Through the embodiments provided in this application, a high-dimensional first target convolution operation is performed on a first feature map to obtain a first target feature map; a wide-dimensional second target convolution operation is performed on the first target feature map to obtain a second target feature map; or, a wide-dimensional second target convolution operation is performed on the first feature map to obtain a third target feature map; a high-dimensional first target convolution operation is performed on the third target feature map to obtain a fourth target feature map; and a second feature map is obtained based on a first output feature map, including: if the first output feature map includes a second target feature map, determining the second target feature map as the second feature map; or, if the first output feature map includes a fourth target feature map, determining the fourth target feature map as the second feature map.

[0138] As an optional approach, an uplink convolution operation with high and wide dimensions is performed on the first feature map to obtain a first output feature map, including:

[0139] S1, Perform a high-dimensional first-target convolution operation on the first feature map to obtain the first target feature map; and

[0140] S2, perform a wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map;

[0141] S3, based on the first output feature map, obtain the second feature map, including:

[0142] S4, determine the first target feature map and the third target feature map as the second feature map; or,

[0143] S5. Integrate the first target feature map and the third target feature map to obtain the second feature map.

[0144] Optionally, in this embodiment, the uplink convolution operation may include, but is not limited to, performing a high-dimensional first target convolution operation on the first feature map to obtain a first target feature map; and performing a wide-dimensional second target convolution operation on the first feature map to obtain a third target feature map.

[0145] Optionally, in this embodiment, the first target feature map and the third target feature map are determined as the second feature map; or, the first target feature map and the third target feature map are integrated to obtain the second feature map.

[0146] It should be noted that in this embodiment, the first target convolution operation and the second target convolution operation are parallel convolution operations, the up-row convolution operation is a parallel operation of the first target convolution operation and the second target convolution operation, and the second feature map can be the first target feature map and the third target feature map, or it can be the feature map after integrating the first target feature map and the third target feature map.

[0147] To further illustrate, such as Figure 7 As shown, assume that the feature map dimension of the first feature map 702 is (c / n)*h*w, where (c / n) is the channel dimension, w is the width dimension in the spatial dimension, and h is the height dimension in the spatial dimension.

[0148] Step S702-1: Perform a high-dimensional first target convolution operation on the first feature map 702 to obtain the first target feature map 704, wherein the feature map dimension of the first target feature map 704 is (c / n)*h*2w; and,

[0149] Step S702-2: Perform a wide-dimensional second target convolution operation on the first feature map 702 to obtain the third target feature map 706, wherein the feature map dimension of the third target feature map 706 is (c / n)*2h*w;

[0150] Step S704: Determine the second feature map 708 from the first target feature map 704 and the third target feature map 706. That is, the second target feature map 708 includes the first target feature map 704 and the third target feature map 706.

[0151] It should be noted that, in addition to being determined by the first target feature map 704 and the third target feature map 706 as described above, the second feature map 708 can also be determined based on, but is not limited to, the first target feature map 704 and the third target feature map 706. In this case, the feature map dimension of the second feature map 708 is (c / n)*2h*2w, and the second feature map 708 has all the information of the first target feature map 704 and the third target feature map 706.

[0152] The embodiments provided in this application involve performing a high-dimensional first target convolution operation on a first feature map to obtain a first target feature map; and performing a wide-dimensional second target convolution operation on the first feature map to obtain a third target feature map; and obtaining a second feature map based on a first output feature map, including: determining the first target feature map and the third target feature map as the second feature map; or, integrating the first target feature map and the third target feature map to obtain the second feature map.

[0153] As an optional approach, after determining the first and third target feature maps as the second feature map, a high-dimensional and wide-dimensional downlink convolution operation is performed on the second feature map to obtain the second output feature map, including:

[0154] S1, perform a high-dimensional first down-convolution operation on the first target feature map to obtain a first output sub-feature map; and perform a wide-dimensional second down-convolution operation on the third target feature map to obtain a second output sub-feature map;

[0155] S2, integrate the first output sub-feature map and the second output sub-feature map to obtain the second output feature map.

[0156] Optionally, in this embodiment, after determining the first target feature map and the third target feature map as the second feature map, a high-dimensional first down-convolution operation is performed on the first target feature map to obtain a first output sub-feature map; and a wide-dimensional second down-convolution operation is performed on the third target feature map to obtain a second output sub-feature map; the first output sub-feature map and the second output sub-feature map are integrated to obtain a second output feature map.

[0157] It should be noted that, in this embodiment, the second feature map is the first target feature map obtained by performing a high-dimensional up-convolution operation on the first feature map and the third target feature map obtained by performing a wide-dimensional up-convolution operation on the second feature map. Correspondingly, the down-convolution operation on the second feature map is to perform a corresponding high-dimensional down-convolution operation on the first target feature map and a corresponding wide-dimensional down-convolution operation on the second target feature map, and then integrate the output sub-feature maps obtained by the high-dimensional down-convolution operation and the wide-dimensional down-convolution operation to obtain the second output feature map.

[0158] To further illustrate, such as Figure 8 As shown, assume that the second feature map 802 includes a feature map with feature map dimensions of (c / n)*h*2w and a feature map with feature map dimensions of (c / n)*2h*w, where (c / n) is the channel dimension, w or 2w is the width dimension in the spatial dimension, and h or 2h is the height dimension in the spatial dimension.

[0159] Step S802-1: Perform a high-dimensional first downlink convolution operation on the (c / n)*2h*w feature map in the second feature map 802 to obtain the first output sub-feature map 804, wherein the feature map dimension of the first output sub-feature map 804 is (c / n)*h*w; and,

[0160] Step S802-2: Perform a wide-dimensional second down-row convolution operation on the (c / n)*h*2w feature map in the second feature map 802 to obtain the second output sub-feature map 806, wherein the feature map dimension of the second output sub-feature map 806 is (c / n)*h*w;

[0161] Step S804: Integrate the first output sub-feature map 804 and the second output sub-feature map 806 to obtain the second output feature map 808, wherein the feature map dimension of the second output feature map 808 is (c / n)*h*w.

[0162] The embodiments provided in this application perform a high-dimensional first down-convolution operation on a first target feature map to obtain a first output sub-feature map; and perform a wide-dimensional second down-convolution operation on a third target feature map to obtain a second output sub-feature map; and integrate the first output sub-feature map and the second output sub-feature map to obtain a second output feature map.

[0163] As an optional approach, the method also includes the following steps before obtaining the target feature map:

[0164] S1, acquire the target image to be identified;

[0165] S2, input the target image into the downsampling module to obtain the low-level features corresponding to the target image;

[0166] S3, input the low-level features into the residual convolution module to obtain the global high-level semantic features corresponding to the target image;

[0167] S4. Obtain the target feature map based on global high-level semantic features.

[0168] Optionally, in this embodiment, the downsampling module may be used, but is not limited to, to obtain the low-level features corresponding to the target image, and the downsampling module may be used, but is not limited to, to instruct the image to be reduced so that the image conforms to the size of the real area and to generate a thumbnail of the image.

[0169] Optionally, in this embodiment, the downsampling module may be, but is not limited to, a constructed stem module, which may be, but is not limited to, used to downsample the length and width of the input image to 1 / n of the initial length and width.

[0170] To further illustrate, a schematic diagram of an optional stem module structure is shown below. Figure 9 As shown, stem module 902 is used to downsample the length and width of the input image to 1 / 4 of the initial length and width, and to obtain the low-level features of the image.

[0171] It should be noted that since inputting an image into the stem module yields low-level, underlying features of the image, it is not suitable for directly obtaining the global, high-level semantic features of the image.

[0172] Optionally, in this embodiment, a schematic diagram of an optional residual convolution module structure is shown below. Figure 10 As shown, the residual convolution module 1002 can, but is not limited to, output the global high-level semantic features of the target image in response to the low-level features corresponding to the input target image.

[0173] As an optional approach, the method also includes the following steps before obtaining the target feature map:

[0174] S1, acquire multiple consecutive images to be identified collected by the target terminal, wherein the computing power value of the target terminal is less than or equal to a preset threshold.

[0175] S2, in response to a task execution request triggered by the target terminal, determines the target image from multiple consecutive images to be recognized. The task execution request is used to request the target terminal to perform a dense estimation visual perception task, which is a task of performing image recognition on the target image to obtain a dense estimation result.

[0176] Optionally, in this embodiment, the image to be identified is a series of consecutive images collected by the target terminal, wherein the computing power value of the target terminal is less than or equal to a preset threshold.

[0177] Optionally, in this embodiment, the target terminal may be, but is not limited to, a low-computing-power device used to complete various intensive estimation tasks, such as a mobile phone, vehicle-mounted vehicle, drone, robot, etc. The intensive estimation tasks may be, but are not limited to, computer vision tasks that require high-resolution representation, such as semantic segmentation, depth estimation, counting, pose estimation, etc.

[0178] As an alternative approach, the aforementioned image recognition method can be applied to a serial high-resolution representation network based on multi-directional convolution, such as... Figure 11 As shown, the specific steps are as follows:

[0179] Step S1102: Let the dimension of the input feature map be c*h*w. Apply a 1*1 convolution to it, and the output channels are reduced to 1 / n of the original. The dimension of the output feature map is (c / n)*h*w. By reducing the number of output channels, channel redundancy is reduced, thereby reducing the computational load of the network.

[0180] Step S1104: Based on the output feature map of step S1102, use a 1*1 convolution kernel to perform convolution in a high dimension, and the output height dimension is twice the input high dimension. Then the dimension of the output feature map is (c / n)*2h*w. At this time, each neuron on the output feature map has all the information in the high dimension of the current position, and the resolution in the high dimension is increased to twice the original.

[0181] Step S1106: Based on the output feature map of step S1104, use a 1*1 convolution kernel to implement width-dimensional convolution, and the output width dimension is twice the input width dimension. Then the output feature map dimension is (c / n)*2h*2w. At this time, each neuron of the output feature map has global information, and the resolution of both height and width is increased to twice the original.

[0182] Step S1108: Perform a wide-dimensional convolution on the output feature map from step S1106 using a 1*1 convolution kernel. The output width dimension is w, so the dimension of the output feature map is (c / n)*2h*w.

[0183] Step S1110: Then, perform a high-dimensional convolution on the output feature map, and output the high dimension as h. At this time, the dimension of the output feature map is (c / n)*h*w, which encodes the features represented by the higher resolution onto the feature map of the original resolution.

[0184] Step S1112: Convolve the feature map output in step S1110 using a 1*1 convolution kernel in the channel dimension, and output the feature map dimension as c*h*w. At this time, the feature map is restored to the input dimension.

[0185] As an alternative approach, the aforementioned image recognition method can be applied to a parallel high-resolution representation network based on multi-directional convolution, such as... Figure 12 As shown, the specific steps are as follows:

[0186] Step S1202: Let the dimension of the input feature map be c*h*w. Apply a 1*1 convolution to it, and the output channels are reduced to 1 / n of the original. The dimension of the output feature map is (c / n)*h*w. By reducing the number of output channels, channel redundancy is reduced, thereby reducing the computational load of the network.

[0187] Step S1204-1: Based on the output feature map from step S1202, perform convolution in the width dimension using a 1*1 convolution kernel, and the output width dimension is twice the input width dimension. Therefore, the output feature map dimension is (c / n)*2h*w. At this point, each neuron in the output feature map possesses all the information in the current width dimension, and the resolution in the width dimension is increased to twice its original value.

[0188] Step S1204-2: Based on the output feature map of step S1202, use a 1*1 convolution kernel to perform convolution in the height dimension, and the output height dimension is twice the input height dimension. Then the dimension of the output feature map is (c / n)*2h*w. At this time, each neuron on the output feature map has all the information in the height dimension of the current position, and the resolution in the high dimension is increased to twice the original.

[0189] Step S1206: Based on the output feature map of step S1204-1, use a 1*1 convolution kernel to perform convolution in the width dimension, and the output width dimension is 1 / 2 times the input width dimension, then the output feature map dimension is (c / n)*h*w;

[0190] Step S1208: Based on the output feature map of step S1204-2, use a 1*1 convolution kernel to perform convolution in the height dimension, and the output height dimension is 1 / 2 times the input height dimension, then the output feature map dimension is (c / n)*h*w;

[0191] Step S1210: Based on the output feature map of step S1206 and the output feature map of step S1208, the feature map is integrated and output. At this time, the feature map dimension of the output feature map is (c / n)*h*w, and each neuron has global information, and the resolution of height and width are increased to twice the original.

[0192] Step S1212: Convolve the feature map output in step S1210 using a 1*1 convolution kernel in the channel dimension, and output the feature map dimension as c*h*w. At this time, the feature map is restored to the input dimension.

[0193] As an alternative approach, the aforementioned image recognition method can be applied to a high-resolution representation network construction scenario, such as... Figure 13 As shown.

[0194] Since inputting an image into the stem module yields low-level, low-level features, it is not suitable for directly obtaining global high-level semantic features. Therefore, firstly, Layer 1302 is constructed using the aforementioned residual convolution module, and the output of the stem module is used as the input to Layer 1302. Secondly, Layers 1304, 1306, and 1308 are constructed using high-resolution multi-directional convolution modules. Further, the output of Layer 1302 is concatenated and input into Layers 1304, 1306, and 1308 to obtain the output. If the number of channels in the output feature map of the previous layer is not equal to the number of channels in the feature map of the current layer, a 1*1 convolution is used to obtain the same number of channels.

[0195] It should be noted that, by constructing a high-resolution representation network using the aforementioned image recognition method, on the one hand, the network obtains a global receptive field on the high-resolution feature map through multi-directional convolution, while increasing spatial information redundancy and possessing powerful high-resolution representation capabilities. On the other hand, the high-resolution multi-directional convolution module in the network only uses 1*1 convolutions with extremely low computational cost, while also reducing channel redundancy, making the network lightweight.

[0196] It is understood that in the specific embodiments of this application, data such as user information are involved. When the above embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0197] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0198] According to another aspect of the embodiments of this application, an image recognition apparatus for implementing the above-described image recognition method is also provided. For example... Figure 14 As shown, the device includes:

[0199] The acquisition unit 1402 is used to acquire a target feature map, wherein the target feature map is used to represent the features corresponding to the target image to be identified;

[0200] The first execution unit 1404 is used to perform a first convolution operation on the target feature map in the channel dimension to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map.

[0201] The second execution unit 1406 is used to perform a second convolution operation on the first feature map in a spatial dimension to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map.

[0202] The third execution unit 1408 is used to perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map.

[0203] The fourth execution unit 1410 is used to perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension.

[0204] The recognition unit 1412 is used to perform image recognition on the target image using the fourth feature map.

[0205] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0206] As an optional solution, the second execution unit 1406 mentioned above includes:

[0207] A first execution device is configured to perform up-convolution operations on the first feature map in both high and wide dimensions to obtain a first output feature map, wherein the high dimension of the first output feature map is greater than the initial high dimension of the target feature map, the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map, and the spatial resolution of the first output feature map is greater than the initial resolution.

[0208] The first acquisition device is used to acquire the second feature map based on the first output feature map.

[0209] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0210] As an optional solution, the third execution unit 1408 mentioned above includes:

[0211] The second execution device is used to perform the downlink convolution operation of the high dimension and the wide dimension on the second feature map to obtain a second output feature map, wherein the high dimension of the second output feature map is equal to the initial high dimension, the wide dimension of the second output feature map is equal to the initial wide dimension, and the spatial resolution of the second output feature map is equal to the spatial resolution of the first output feature map.

[0212] The second acquisition device is used to acquire the third feature map based on the second output feature map.

[0213] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0214] As an optional embodiment, the first actuator includes:

[0215] The first execution sub-device is used to perform the high-dimensional first target convolution operation on the first feature map to obtain the first target feature map;

[0216] The second execution sub-device is configured to perform the wide-dimensional second target convolution operation on the first target feature map to obtain a second target feature map; or,

[0217] The third execution sub-device is used to perform the wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map;

[0218] The fourth execution sub-device is used to perform the aforementioned high-dimensional first target convolution operation on the aforementioned third target feature map to obtain the fourth target feature map.

[0219] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0220] As an optional solution, the first acquisition device mentioned above includes:

[0221] The first acquisition sub-device is configured to determine the second target feature map as the second feature map when the first output feature map includes the second target feature map; or,

[0222] The second acquisition sub-device is used to determine the fourth target feature map as the second feature map when the first output feature map includes the fourth target feature map.

[0223] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0224] As an optional embodiment, the first actuator includes:

[0225] A first execution sub-device is configured to perform the aforementioned high-dimensional first target convolution operation on the aforementioned first feature map to obtain a first target feature map; and

[0226] The third execution sub-device is used to perform the wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map;

[0227] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0228] As an optional solution, the first acquisition device mentioned above includes:

[0229] The third acquisition sub-device is used to determine the first target feature map and the third target feature map as the second feature map; or,

[0230] The fourth acquisition sub-device is used to integrate the first target feature map and the third target feature map to obtain the second feature map.

[0231] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0232] As an optional embodiment, the second actuator includes:

[0233] The fifth execution sub-device is configured to, after determining the first target feature map and the third target feature map as the second feature map, perform the high-dimensional first downlink convolution operation on the first target feature map to obtain a first output sub-feature map; and,

[0234] The sixth execution sub-device is used to perform the wide-dimensional second downlink convolution operation on the third target feature map after determining the first target feature map and the third target feature map as the second feature map, to obtain the second output sub-feature map;

[0235] The fifth acquisition sub-device is used to integrate the first output sub-feature map and the second output sub-feature map after determining the first target feature map and the third target feature map as the second feature map to obtain the second output feature map.

[0236] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0237] As an optional solution, the above-mentioned device further includes:

[0238] The third acquisition device is used to acquire the target image to be identified before acquiring the target feature map;

[0239] The fourth acquisition device is used to input the target image into the downsampling module before acquiring the target feature map to obtain the underlying features corresponding to the target image;

[0240] The fifth acquisition device is used to input the low-level features into the residual convolution module before acquiring the target feature map to obtain the global high-level semantic features corresponding to the target image.

[0241] The sixth acquisition device is used to acquire the target feature map based on the global high-level semantic features before acquiring the target feature map.

[0242] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0243] As an optional solution, the above-mentioned device further includes:

[0244] The seventh acquisition device is used to acquire multiple consecutive images to be identified collected by the target terminal before acquiring the target feature map, wherein the computing power value of the target terminal is less than or equal to a preset threshold.

[0245] The eighth acquisition device is configured to determine the target image from the plurality of consecutive images to be identified in response to a task execution request triggered by the target terminal before acquiring the target feature map, wherein the task execution request is used to request the target terminal to perform a dense estimation visual perception task, the dense estimation visual perception task being a task of performing image recognition on the target image to obtain a dense estimation result.

[0246] For specific implementation examples, please refer to the examples shown in the image recognition method above. These examples will not be repeated here.

[0247] According to another aspect of the embodiments of this application, an electronic device for implementing the above-described image recognition method is also provided, such as... Figure 15 As shown, the electronic device includes a memory 1502 and a processor 1504. The memory 1502 stores a computer program, and the processor 1504 is configured to execute the steps of any of the above method embodiments via the computer program.

[0248] Optionally, in this embodiment, the aforementioned electronic device may be located in at least one of a plurality of network devices in a computer network.

[0249] Optionally, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0250] S1, Obtain the target feature map, where the target feature map is used to represent the features corresponding to the target image to be identified;

[0251] S2, perform the first convolution operation of the channel dimension on the target feature map to obtain the first feature map of the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map;

[0252] S3, perform a second convolution operation on the first feature map in the spatial dimension to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map;

[0253] S4, perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, where the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map;

[0254] S5, perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, where the second channel dimension is equal to the initial channel dimension;

[0255] S6. Using the fourth feature map, perform image recognition on the target image.

[0256] Alternatively, as those skilled in the art will understand, Figure 15 The structure shown is for illustrative purposes only. Electronic devices can also be smartphones (such as Android phones, iOS phones, etc.), tablets, PDAs, mobile internet devices (MIDs), PADs, and other terminal devices. Figure 15 This does not limit the structure of the aforementioned electronic devices. For example, the electronic device may also include components that are more... Figure 15 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 15 The different configurations shown.

[0257] The memory 1502 can be used to store software programs and modules, such as the program instructions / modules corresponding to the image recognition method and apparatus in this embodiment. The processor 1504 executes various functional applications and data processing by running the software programs and modules stored in the memory 1502, thereby realizing the aforementioned image recognition method. The memory 1502 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 1502 may further include memory remotely located relative to the processor 1504, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 1502 may be used, but is not limited to, to store channel dimension information, spatial dimension information, spatial resolution, and other information. As an example, such as... Figure 15 As shown, the memory 1502 may include, but is not limited to, the acquisition unit 1402, the first execution unit 1404, the second execution unit 1406, the third execution unit 1408, the fourth execution unit 1410, and the recognition unit 1412 in the image recognition device. Furthermore, it may include, but is not limited to, other module units in the image recognition device, which will not be elaborated upon in this example.

[0258] Optionally, the transmission device 1506 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 1506 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 1506 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0259] In addition, the aforementioned electronic device also includes: a display 1508 for displaying the aforementioned channel dimension information, spatial dimension information, spatial resolution, and other information; and a connection bus 1510 for connecting the various module components in the aforementioned electronic device.

[0260] In other embodiments, the aforementioned terminal device or server can be a node in a distributed system, wherein the distributed system can be a blockchain system, which is a distributed system formed by connecting multiple nodes through network communication. The nodes can form a peer-to-peer (P2P) network, and any form of computing device, such as a server, terminal, or other electronic device, can become a node in the blockchain system by joining this peer-to-peer network.

[0261] According to one aspect of this application, a computer program product is provided, comprising a computer program / instructions containing program code for performing the methods shown in the flowchart. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit, it performs various functions provided in embodiments of this application.

[0262] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0263] It should be noted that the computer system of the electronic device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0264] A computer system includes a Central Processing Unit (CPU), which performs various appropriate actions and processes based on programs stored in Read-Only Memory (ROM) or loaded from RAM. ROM also stores various programs and data required for system operation. The CPU, ROM, and RAM are interconnected via a bus. Input / output interfaces (I / O interfaces) are also connected to the bus.

[0265] The following components are connected to the input / output interface: input sections including keyboards, mice, etc.; output sections including cathode ray tubes (CRTs), liquid crystal displays (LCDs), and speakers; storage sections including hard drives; and communication sections including network interface cards such as LAN cards and modems. The communication section performs communication processing via a network such as the Internet. Drives are also connected to the input / output interface as needed. Removable media, such as disks, optical discs, magneto-optical discs, semiconductor memories, etc., are installed on the drive as needed so that computer programs read from them can be installed into the storage section as required.

[0266] Specifically, according to embodiments of this application, the processes described in the various method flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by a central processing unit, it performs various functions defined in the system of this application.

[0267] According to one aspect of this application, a computer-readable storage medium is provided, wherein a processor of a computer device reads computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various alternative implementations described above.

[0268] Optionally, in this embodiment, the computer-readable storage medium described above may be configured to store a computer program for performing the following steps:

[0269] S1, Obtain the target feature map, where the target feature map is used to represent the features corresponding to the target image to be identified;

[0270] S2, perform the first convolution operation of the channel dimension on the target feature map to obtain the first feature map of the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map;

[0271] S3, perform a second convolution operation on the first feature map in the spatial dimension to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map;

[0272] S4, perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, where the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map;

[0273] S5, perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, where the second channel dimension is equal to the initial channel dimension;

[0274] S6. Using the fourth feature map, perform image recognition on the target image.

[0275] Optionally, in this embodiment, those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0276] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0277] If the integrated units in the above embodiments are implemented as software functional units and sold or used as independent products, they can be stored in the aforementioned computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause one or more computer devices (which may be personal computers, servers, or network devices, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0278] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0279] In the several embodiments provided in this application, it should be understood that the disclosed client can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, indirect coupling or communication connection between units or modules, and may be electrical or other forms.

[0280] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0281] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0282] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. An image recognition method, characterized in that, include: Obtain a target feature map, wherein the target feature map is used to represent the features corresponding to the target image to be identified; Perform a first convolution operation on the target feature map along the channel dimension to obtain a first feature map along the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; Perform a second convolution operation on the first feature map in spatial dimension to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map; Perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map; Perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; The target image is then used for image recognition.

2. The method according to claim 1, characterized in that, The step of performing a second convolution operation on the first feature map in the spatial dimension to obtain a second feature map includes: Perform uplink convolution operations on the first feature map in both high and wide dimensions to obtain a first output feature map, wherein the high dimension of the first output feature map is greater than the initial high dimension of the target feature map, the wide dimension of the first output feature map is greater than the initial wide dimension of the target feature map, and the spatial resolution of the first output feature map is greater than the initial resolution; based on the first output feature map, obtain the second feature map; The step of performing a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map includes: Perform the downlink convolution operation of the high dimension and the wide dimension on the second feature map to obtain the second output feature map, wherein the high dimension of the second output feature map is equal to the initial high dimension, the wide dimension of the second output feature map is equal to the initial wide dimension, and the spatial resolution of the second output feature map is equal to the spatial resolution of the first output feature map; based on the second output feature map, obtain the third feature map.

3. The method according to claim 2, characterized in that, The step of performing high-dimensional and wide-dimensional uplink convolution operations on the first feature map to obtain the first output feature map includes: Perform the high-dimensional first target convolution operation on the first feature map to obtain a first target feature map; perform the wide-dimensional second target convolution operation on the first target feature map to obtain a second target feature map; or, Perform the wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map; perform the high-dimensional first target convolution operation on the third target feature map to obtain the fourth target feature map; The step of obtaining the second feature map based on the first output feature map includes: If the first output feature map includes the second target feature map, then the second target feature map is determined as the second feature map; or, If the first output feature map includes the fourth target feature map, the fourth target feature map is determined as the second feature map.

4. The method according to claim 2, characterized in that, The step of performing high-dimensional and wide-dimensional uplink convolution operations on the first feature map to obtain the first output feature map includes: Perform the high-dimensional first target convolution operation on the first feature map to obtain the first target feature map; and, Perform the wide-dimensional second target convolution operation on the first feature map to obtain the third target feature map; The step of obtaining the second feature map based on the first output feature map includes: The first target feature map and the third target feature map are determined as the second feature map; or, The second feature map is obtained by integrating the first target feature map and the third target feature map.

5. The method according to claim 4, characterized in that, After determining the first target feature map and the third target feature map as the second feature map, the next step of performing the downlink convolution operation on the second feature map in both the high-dimensional and wide-dimensional dimensions to obtain the second output feature map includes: Perform the high-dimensional first down-convolution operation on the first target feature map to obtain a first output sub-feature map; and perform the wide-dimensional second down-convolution operation on the third target feature map to obtain a second output sub-feature map; The first output sub-feature map and the second output sub-feature map are integrated to obtain the second output feature map.

6. The method according to any one of claims 1 to 5, characterized in that, Prior to acquiring the target feature map, the method further includes: Obtain the target image to be identified; The target image is input into the downsampling module to obtain the low-level features corresponding to the target image; The low-level features are input into the residual convolution module to obtain the global high-level semantic features corresponding to the target image; Based on the global high-level semantic features, the target feature map is obtained.

7. The method according to any one of claims 1 to 5, characterized in that, Prior to acquiring the target feature map, the method further includes: Acquire multiple consecutive images to be identified from a target terminal, wherein the computing power value of the target terminal is less than or equal to a preset threshold; In response to a task execution request triggered by the target terminal, the target image is determined from the multiple consecutive images to be identified, wherein the task execution request is used to request the target terminal to perform a dense estimation visual perception task, the dense estimation visual perception task being a task of performing image recognition on the target image to obtain a dense estimation result.

8. An image recognition device, characterized in that, include: An acquisition unit is used to acquire a target feature map, wherein the target feature map is used to represent the features corresponding to the target image to be identified; The first execution unit is configured to perform a first convolution operation on the target feature map in the channel dimension to obtain a first feature map in the first channel dimension, wherein the first channel dimension is smaller than the initial channel dimension of the target feature map; The second execution unit is used to perform a second convolution operation in the spatial dimension on the first feature map to obtain a second feature map, wherein the first spatial dimension of the second feature map is greater than the initial spatial dimension of the target feature map, and the spatial resolution of the second feature map is greater than the initial resolution of the target feature map; The third execution unit is used to perform a third convolution operation on the second feature map in the spatial dimension to obtain a third feature map, wherein the second spatial dimension of the third feature map is equal to the initial spatial dimension, and the spatial resolution of the third feature map is equal to the spatial resolution of the second feature map. The fourth execution unit is used to perform a fourth convolution operation on the third feature map in the channel dimension to obtain a fourth feature map in the second channel dimension, wherein the second channel dimension is equal to the initial channel dimension; The recognition unit is used to perform image recognition on the target image using the fourth feature map.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program can be executed by a terminal device or computer at runtime as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1 to 7.

11. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method described in any one of claims 1 to 7 through the computer program.

Images