Image feature extraction method, image recognition method, device, medium and vehicle

Abstract

The present invention relates to the field of image feature extraction technology, and specifically provides an image feature extraction method, an image recognition method, a device, a medium, and a vehicle, which are intended to solve the problem of timely extracting image features using a neural network deployed on a mobile device. To this end, the method of the present invention includes constructing an initial image feature extraction network deployed on a mobile device; replacing the deep convolution in the initial image feature extraction network with a normal convolution; performing quantization processing on the image feature extraction network after replacing the deep convolution to obtain a final image feature extraction network; and using the final image feature extraction network to extract features from the image. Through the above method, the mobile device can extract the image features of the image more timely and effectively, while not consuming too much computing resources and electrical power, nor taking up too much memory, thereby reducing the storage pressure of the mobile device.

Description

Technical Field

[0001] The present invention relates to the technical field of image feature extraction, and specifically provides an image feature extraction method, an image recognition method, a device, a medium and a vehicle. Background Art

[0002] In scenarios such as image recognition, deep neural networks can be used to construct image feature extraction models. These models are then used to extract features from images, enabling better image recognition based on the extracted image features. For example, in lane recognition for autonomous driving, an image feature extraction model can be used to extract features from lane images. The extracted image features can then be used to detect the location of lane markings within the lane, and the driving trajectory can be planned based on the location of the lane markings. However, with the continuous development of deep neural networks, their network structures are becoming increasingly complex, consuming increasing amounts of computing resources and power. Deploying complex deep neural networks on mobile devices, such as vehicle-mounted computers, can place significant strain on the computing resources and power consumption of these devices. Currently, conventional neural network lightweighting methods primarily quantize floating-point parameters in the neural network into fixed-point parameters. However, depthwise convolution in deep neural networks is a special type of grouped convolution. Even quantizing the parameters of the depthwise convolution into fixed-point numbers does not improve the processing speed of the depthwise convolution, and consequently, the processing speed of the entire deep neural network. This results in the mobile device being unable to timely acquire image features and perform image recognition based on these features. In addition, the reduced processing speed of deep neural networks will inevitably consume more electricity and increase the power consumption pressure of mobile devices.

[0003] Accordingly, this field requires a new technical solution to solve the above problems. Summary of the Invention

[0004] In order to overcome the above-mentioned defects, the present invention is proposed to provide an image feature extraction method, apparatus, device and medium for mobile devices that solve or at least partially solve the technical problem of how to use a neural network deployed on a mobile device to timely extract image features to better perform image recognition.

[0005] In a first aspect, the present invention provides an image feature extraction method for a mobile device, the method comprising:

[0006] Build an initial image feature extraction network deployed on mobile devices;

[0007] Replacing the depth convolution in the initial image feature extraction network with ordinary convolution;

[0008] Quantizing the image feature extraction network after replacing the depth convolution to obtain a final image feature extraction network;

[0009] The final image feature extraction network is used to extract features from the image.

[0010] In one technical solution of the above-mentioned method for extracting image features on a mobile device, the step of "constructing an initial image feature extraction network deployed on the mobile device" specifically includes:

[0011] The ConvNeXt network is used to build the initial image feature extraction network deployed on mobile devices.

[0012] In one technical solution of the above-mentioned method for extracting image features of a mobile device, the step of “replacing the deep convolution in the initial image feature extraction network with a normal convolution” specifically includes:

[0013] The depthwise convolution is replaced with a small-kernel ordinary convolution and the small-kernel ordinary convolution is subjected to a dilation process to increase the receptive field of the small-kernel ordinary convolution.

[0014] In one technical solution of the above-mentioned method for extracting image features of a mobile device, before the step of “performing quantization processing on the image feature extraction network after replacing the depthwise convolution to obtain a final image feature extraction network”, the method further includes:

[0015] The special operators in the image feature extraction network after replacing the depth convolution are replaced with non-special operators.

[0016] In one technical solution of the above-mentioned method for extracting image features of a mobile device, the step of “performing quantization processing on the image feature extraction network after replacing the depthwise convolution to obtain a final image feature extraction network” specifically includes:

[0017] The floating-point parameters of the image feature extraction network after replacing the depth convolution are quantized into fixed-point parameters to obtain a final image feature extraction network.

[0018] In one technical solution of the above-mentioned image feature extraction method for mobile devices, the step of “quantizing the floating-point parameters of the image feature extraction network after replacing the depthwise convolution into fixed-point parameters to obtain a final image feature extraction network” specifically includes:

[0019] The 32-bit floating-point number parameters of the image feature extraction network after replacing the depth convolution are quantized into 8-bit fixed-point number parameters to obtain the final image feature extraction network.

[0020] In a second aspect, a method for image recognition for a mobile device is provided, the method comprising:

[0021] Acquire an image captured by the mobile device;

[0022] Extracting image features of the image using the image feature extraction method for mobile devices described in the above technical solution;

[0023] Image recognition is performed according to the image features.

[0024] In a third aspect, a computer device is provided, comprising a processor and a storage device, wherein the storage device is suitable for storing multiple program codes, and the program codes are suitable for being loaded and run by the processor to execute the image feature extraction method for a mobile device described in the above technical solution or the image recognition method for a mobile device described in the above technical solution.

[0025] In a fourth aspect, a computer-readable storage medium is provided, in which multiple program codes are stored, and the program codes are suitable for being loaded and run by a processor to execute the image feature extraction method for a mobile device described in the above technical solution or the image recognition method for a mobile device described in the above technical solution.

[0026] In a fifth aspect, a vehicle is provided, comprising the computer device described in the above technical solution.

[0027] The above one or more technical solutions of the present invention have at least one or more of the following beneficial effects:

[0028] In implementing the technical solution of the present invention, an initial image feature extraction network deployed on a mobile device can be first constructed, and then the deep convolution in the initial image feature extraction network can be replaced with ordinary convolution. Finally, the image feature extraction network after replacing the deep convolution is quantized to obtain a final image feature extraction network, and the final image feature extraction network is used to extract features from the image. Replacing the deep convolution in the image feature extraction network with ordinary convolution can avoid the problem that after quantizing the deep convolution, the processing speed of the quantized convolution cannot be increased, thereby reducing the extraction speed of the quantized image feature extraction network when extracting image features. In addition, by quantizing the image feature extraction network, the memory occupied by the image feature extraction network can be reduced, thereby reducing the storage pressure of the mobile device.

[0029] Furthermore, in some technical solutions for implementing the present invention, a ConvNeXt network can be used to construct an initial image feature extraction network deployed on a mobile device. Based on the advantages of high precision and fast processing speed of the ConvNeXt network, using the ConvNeXt network to construct an image feature extraction network can simultaneously improve the accuracy and extraction speed of image features. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The disclosure of the present invention will become more easily understood with reference to the accompanying drawings. Those skilled in the art will readily appreciate that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Among them:

[0031] Figure 1 This is a flow chart of main steps of an image feature extraction method for a mobile device according to an embodiment of the present invention;

[0032] Figure 2 1 is a schematic diagram of the network structure of ConvNeXt and an initial image feature extraction network according to an embodiment of the present invention;

[0033] Figure 3 2 is a schematic diagram comparing the running time of depthwise convolution and ordinary convolution according to an embodiment of the present invention;

[0034] Figure 4 is a flowchart of main steps of an image recognition method for a mobile device according to an embodiment of the present invention;

[0035] Figure 5 2 is a schematic diagram of a network structure of an image feature extraction network for performing image recognition of lane lines according to an embodiment of the present invention;

[0036] Figure 6 2 is a schematic diagram of a network structure of an image feature extraction network for performing image recognition on dynamic objects according to an embodiment of the present invention;

[0037] Figure 7 2 is a schematic diagram of the network structure of an image feature extraction network for performing image recognition of parking space corner points according to an embodiment of the present invention. DETAILED DESCRIPTION

[0038] Some embodiments of the present invention are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0039] In the description of the present invention, "processor" may include hardware, software, or a combination of the two. The processor may be a central processing unit, a microprocessor, an image processor, a digital signal processor, or any other suitable processor. The processor has data and / or signal processing functions. The processor may be implemented in software, hardware, or a combination of the two. Computer-readable storage media include any suitable medium that can store program code, such as a magnetic disk, a hard disk, an optical disk, a flash memory, a read-only memory, a random access memory, etc. The term "A and / or B" represents all possible combinations of A and B, such as only A, only B, or A and B.

[0040] Here we first explain some terms involved in the present invention.

[0041] Convolution is a very useful computational operation in machine learning, primarily used to extract image features. Its basic properties include the kernel, stride, and channel. The kernel represents the receptive field of the convolution operation: larger kernels increase the receptive field, while smaller kernels decrease the receptive field. The stride is the number of pixels the kernel moves at each step as it traverses the feature map. The channel is the number of channels in the convolution layer. Convolution can be categorized into three types: ordinary convolution (also known as standard convolution or regular convolution), depthwise convolution, and dilated convolution. Dilated convolution is formed by dilating the kernel by the dilation rate. A major difference between ordinary convolution and depthwise convolution is that each convolution kernel of ordinary convolution performs a convolution operation on each channel of the image simultaneously, while each convolution kernel of depthwise convolution performs a convolution operation on one channel of the image separately. It should be noted that ordinary convolution, depthwise convolution, and dilated convolution are all conventional convolutions in the field of convolution technology. The embodiments of the present invention do not elaborate on the convolution operations of ordinary convolution, depthwise convolution, and dilated convolution.

[0042] An operator is an operator that performs related operations in a convolutional neural network. For example, a convolution operation is an operator. In an embodiment of the present invention, operators are divided into special operators and non-special operators. Special operators refer to operators whose operation speed cannot be improved after quantization, and non-special operators refer to operators whose operation speed can be improved after quantization. For example, ordinary convolution is a non-special operator, while depth convolution is a special operator. In addition, operators such as softmax, sigmoid, and dimshuffle are also special operators. Among them, quantization refers to quantizing floating-point parameters into fixed-point parameters.

[0043] See attached Figure 1 , Figure 1 FIG. 1 is a flow chart showing the main steps of an image feature extraction method for a mobile device according to an embodiment of the present invention. Figure 1 As shown, the image feature extraction method for a mobile device in the embodiment of the present invention mainly includes the following steps S101 to S104.

[0044] Step S101: constructing an initial image feature extraction network deployed on a mobile device.

[0045] Mobile devices include, but are not limited to, portable devices running Windows, Linux, or other operating systems. For example, mobile devices may include mobile phones and car computers. It should also be understood that in some other embodiments, the mobile device may not be a portable device, but may be a desktop computer.

[0046] An image feature extraction network is a convolutional neural network used to extract image features. After the image features are extracted by the image feature extraction network, they can be used for image recognition. For example, after the image feature extraction network extracts the image features of a lane image, lane markings can be identified based on these features to determine whether the lane image contains lane markings and their positions in the image coordinate system. The positions of the lane markings in the world coordinate system can then be determined through coordinate system conversion. Once the positions of the lane markings in the world coordinate system are determined, the vehicle's path can be better planned, such as ensuring that the vehicle stays in the center of the lane.

[0047] In embodiments of the present invention, conventional convolutional neural networks in the field of neural network technology can be used as image feature extraction networks. In some implementations, a ConvNeXt network can be used to construct an initial image feature extraction network deployed on a mobile device. ConvNeXt networks have advantages such as high accuracy and fast processing speed. Therefore, using a ConvNeXt network to construct an image feature extraction network can further improve the computational processing speed of the image feature extraction network.

[0048] Step S102: Replace the deep convolution in the initial image feature extraction network with ordinary convolution.

[0049] The initial image feature extraction network may contain multiple types of convolutions such as ordinary convolution and depth convolution. The depth convolution in the initial image feature extraction network is replaced by ordinary convolution.

[0050] Deep convolution can change the size of its own convolution kernel when performing convolution operations, while the size of the convolution kernel of ordinary convolution is fixed when performing convolution operations. When deep convolution is required to change the size of the convolution kernel for convolution operations, if these deep convolutions are replaced with ordinary convolutions with the same size convolution kernel, that is, the image feature extraction network is rebuilt, it will undoubtedly greatly reduce the efficiency of image feature extraction, and thus reduce the efficiency of image recognition using image features. If these deep convolutions are not replaced with ordinary convolutions with the same size convolution kernel, it will obviously affect the accuracy of image feature extraction, and thus reduce the accuracy of image recognition using image features. In this regard, in some embodiments, when the deep convolution in the initial image feature extraction network is replaced with ordinary convolution, the deep convolution can be replaced with a small-core ordinary convolution and the small-core ordinary convolution can be hollowed out to increase the receptive field of the small-core ordinary convolution, that is, the deep convolution is replaced with a hollow convolution based on the small-core ordinary convolution.

[0051] Small-core normal convolution refers to a normal convolution in which the convolution kernel size is smaller than a preset size threshold. Those skilled in the art can flexibly set the specific value of the preset size threshold according to actual needs. For example, in this embodiment, the preset size threshold may be 7×7, and the small-core normal convolution may be a normal convolution in which the convolution kernel size is 3×3.

[0052] See attached Figure 2 , Figure 2 The left side of the figure is a schematic diagram of the network structure of the ConvNeXt network. Figure 2 The right side of the figure is the initial image feature extraction network after the depth convolution Dw conv7×7 in the ConvNeXt network is replaced by the method described in step S102 above. Figure 2 As shown, in this embodiment, the depth convolution Dw conv7×7 can be replaced by the ordinary convolution conv3×3, and the ordinary convolution conv3×3 is subjected to dilation processing, wherein the convolution expansion rate used in the dilation processing is equal to 2, that is, dilation2.

[0053] Step S103: quantize the image feature extraction network after replacing the depth convolution to obtain the final image feature extraction network.

[0054] Quantization processing refers to quantizing the floating-point parameters of the image feature extraction network into fixed-point parameters. After replacing the depth convolution, the floating-point parameters of the image feature extraction network can be quantized into fixed-point parameters, thereby obtaining the final image feature extraction network. In some embodiments, if the floating-point parameters of the initial image feature extraction network are 32-bit floating-point parameters, the 32-bit floating-point parameters can be quantized into 8-bit fixed-point parameters to obtain the final image feature extraction network. It should be noted that those skilled in the art can use conventional floating-point and fixed-point conversion methods to quantize floating-point parameters into fixed-point parameters, and the embodiment of the present invention does not specifically limit the floating-point and fixed-point conversion method.

[0055] It is understood by those skilled in the art that deep convolution has a higher processing speed than ordinary convolution. Therefore, convolutional neural networks based on deep convolution are widely used in various image recognition scenarios. However, after comparing the processing speeds of deep convolution and ordinary convolution in floating-point and fixed-point operations, it can be determined that the processing speed of deep convolution in fixed-point operations is not improved compared to that in floating-point operations. On the contrary, the processing speed of ordinary convolution in fixed-point operations is greatly improved. Figure 3 As shown in the figure, the operation processing speed of depth convolution and ordinary convolution in floating-point operations is close to 1:9, and the operation processing speed of depth convolution and ordinary convolution in fixed-point operations is close to 1:1. The operation processing speed of depth convolution in floating-point operations and fixed-point operations has almost not changed, while the operation processing speed of ordinary convolution in fixed-point operations has been greatly improved compared with fixed-point operations. It can be seen that although depth convolution has a higher operation processing speed, quantizing depth convolution cannot further improve the operation processing speed of depth convolution. On the contrary, although the operation processing speed of ordinary convolution in floating-point operations is lower than that of depth convolution, quantizing ordinary convolution can greatly improve the operation processing speed of ordinary convolution.

[0056] The processing speed of deep convolution is higher than that of ordinary convolution, and therefore there is a technical bias in using deep convolution instead of ordinary convolution to improve the processing speed of convolutional neural networks. In this embodiment of the present invention, ordinary convolution is used to replace deep convolution, so that the processing speed of the image feature extraction network can be greatly improved after quantization processing, so that the image feature extraction network can run better on mobile devices, not only without consuming too many computing resources and electric power, but also in a timely manner to obtain the image features of the image, which is conducive to the rapid completion of image recognition.

[0057] In addition, quantization processing can also reduce the space occupied by the image feature extraction network, thereby alleviating the storage pressure of mobile devices.

[0058] Step S104: Using the final image feature extraction network to extract features from the image.

[0059] In order to extract more accurate image features through the image feature extraction network, we can first establish an image recognition model based on the image feature extraction network. After the image recognition model is established, the image recognition model is trained. Through model training, the network parameters in the image feature extraction network are optimized, so that the image features of the image can be extracted more accurately.

[0060] It should be noted that conventional model training methods in the field of machine learning can be used to train the image recognition model in the embodiments of the present invention, and the embodiments of the present invention are not specifically limited to this. For example, training samples are input into the image recognition model, the model loss value is calculated through forward propagation, the parameter gradients of the model parameters are calculated based on the loss value, and the model parameters are updated based on the parameter gradients through backpropagation until the image recognition model meets the convergence conditions and then training is terminated.

[0061] By deploying and running the image feature extraction network on a mobile device through the method described in steps S101 to S104 above, the mobile device can extract the image features of the image more timely and effectively without consuming too much computing resources and power, or taking up too much memory, thereby reducing the storage pressure of the mobile device.

[0062] In another embodiment of an image feature extraction method for a mobile device according to the present invention, the operation processing speed of the image feature extraction network after quantization processing can be further improved by replacing special operators in the image feature extraction network with non-special operators. Specifically, in addition to including steps S101 to S104 in the aforementioned method embodiment, the image feature extraction method in this embodiment of the present invention also includes a step of replacing special operators after executing step S102 and before executing step S103. Specifically, after replacing the deep convolution in the initial image feature extraction network with ordinary convolution, the special operator is further replaced with a non-special operator.

[0063] Special operators refer to operators whose operation speed cannot be improved after quantization, while non-special operators refer to operators whose operation speed can be improved after quantization. For example, ordinary convolution is a non-special operator, while depthwise convolution is a special operator. In addition, operators such as softmax, sigmoid, and dimshuffle are also special operators. Since depthwise convolution is replaced in step S102, the special operators replaced in this step are special operators existing in the image feature extraction network in addition to depthwise convolution.

[0064] In the embodiment of the present invention, a comparison test of the operation speed before and after quantization can be conducted for different special operators. Based on the results of the comparison test, it can be determined which operators are special operators and which are non-special operators. For example, operators whose speed increase ratio after quantization is less than a preset ratio threshold can be considered special operators, while operators whose speed increase ratio after quantization is greater than or equal to the preset ratio threshold can be considered non-special operators.

[0065] The above is a description of an embodiment of an image feature extraction method for a mobile device. The following is a description of an embodiment of an image recognition method for a mobile device.

[0066] See attached Figure 4 In an embodiment of an image recognition method for a mobile device according to the present invention, the image recognition method for a mobile device may include the following steps S201 to S203.

[0067] Step S201: Acquire an image captured by a mobile device.

[0068] The mobile device may be provided with an image acquisition device, through which images are acquired. Alternatively, the mobile device may not be provided with an image acquisition device, but may communicate with an image acquisition device provided elsewhere to acquire images acquired by the image acquisition device.

[0069] Step S202: extracting image features of the image using an image feature extraction method for mobile devices.

[0070] The image feature extraction method for a mobile device refers to the image feature extraction method for a mobile device described in the aforementioned method embodiment, which will not be described in detail here.

[0071] It should be noted that, for different image recognition requirements, when constructing the initial image feature extraction network through the method described in step S101 in the aforementioned method embodiment, the network structure will be slightly different. Those skilled in the art can flexibly change the network structure according to actual needs. However, no matter which network structure is adopted, the image feature extraction network is subjected to deep convolution replacement and quantization processing and other operations through the methods described in steps S102 to S103 in the aforementioned method embodiment, thereby obtaining the image feature extraction network finally used to extract image features.

[0072] See attached Figure 5 To the attached Figure 7 , Figure 5 The network structure of the image feature extraction network for lane line image recognition is shown as an example. Figure 6 The network structure of the image feature extraction network for image recognition of dynamic objects is shown as an example. Figure 7 The network structure of the image feature extraction network for image recognition of parking space corner points is exemplified.

[0073] like Figure 5 As shown in the figure, the image feature extraction network consists of two branches, the left branch consisting of a sequentially connected convolution layer (conv3×3), convolution layer (conv1×3), convolution layer (conv3×3), and convolution layer (conv3×1). The right branch's input is connected to the output of the first convolution layer (conv3×3), and its output is summed with the output of convolution layer (conv3×1). This means the right branch is a shortcut layer. The network structure of "convolution layer (conv3×3), convolution layer (conv1×3), convolution layer (conv3×3), and convolution layer (conv3×1)" in the left branch can better extract both horizontal and vertical features of the image.

[0074] like Figure 6 As shown in the figure, the image feature extraction network consists of two branches, the left branch consisting of a sequentially connected convolution conv1×3, a convolution conv3×3, and a convolution conv3×1. The right branch consists of a single convolution conv3×3, whose input is connected to the output of the convolution conv1×3 and whose output is summed with the output of the convolution conv3×1. To downsample the image, the convolution conv3×3 in both the left and right branches can be used simultaneously. For example, the stride of the convolution conv3×3 can be set from 1 to 2.

[0075] like Figure 7 As shown in the figure, after the image data (c, h, w) of the input image is split, the split image data (c / 2, h, w) is input to the identity module and the block module respectively. The output sides of the identity module and the block module are connected to the concat module at the same time. The output data of the identity module and the block module are spliced ​​through the concat module, and then the spliced ​​image data (c, h, w) is input to the convolution conv1×1. Among them, the identity module is a shortcut layer, and the block module includes the convolution conv1×3, convolution conv3×3, and convolution conv3×1 connected in sequence.

[0076] Step S203: performing image recognition based on image features.

[0077] After determining the image feature extraction network in step S202, a feature recognition network can be constructed based on the neural network. The output of the image feature extraction network is connected to the input of the feature recognition network to form an image recognition model. The extracted image features are input into the feature recognition network, which then performs image recognition on the image features to obtain recognition results. For example, in lane line recognition, the feature recognition network can identify whether the image contains lane lines based on the image features.

[0078] It should be noted that for different image recognition requirements, those skilled in the art can flexibly set up different feature recognition networks. As long as the image features can be recognized through the feature recognition network and the corresponding image recognition requirements are met, the embodiments of the present invention do not specifically limit the network structure of the feature recognition network.

[0079] Through the method described in steps S201 to S203 above, the mobile device can extract the image features of the input image more timely and effectively, thereby quickly completing the image recognition task.

[0080] It should be pointed out that although the various steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effects of the present invention, different steps do not have to be performed in such an order. They can be performed simultaneously (in parallel) or in other orders. These changes are within the scope of protection of the present invention.

[0081] Those skilled in the art will appreciate that all or part of the processes in the method for implementing the above-mentioned embodiment of the present invention may also be accomplished by instructing the relevant hardware through a computer program. The computer program may be stored in a computer-readable storage medium. When the computer program is executed by a processor, it may implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which may be in source code form, object code form, executable file, or some intermediate form. The computer-readable storage medium may include: any entity or device, medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electric carrier signal, telecommunication signal, and software distribution medium capable of carrying the computer program code. It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable storage media do not include electric carrier signals and telecommunication signals.

[0082] Furthermore, the present invention also provides a computer device. In one embodiment of a computer device according to the present invention, the computer device includes a processor and a storage device. The storage device can be configured to store a program for executing the image feature extraction method for a mobile device or the image recognition method for a mobile device according to the above method embodiment. The processor can be configured to execute the program in the storage device, which includes but is not limited to a program for executing the image feature extraction method for a mobile device or the image recognition method for a mobile device according to the above method embodiment. For ease of explanation, only the parts related to the embodiment of the present invention are shown. For specific technical details not disclosed, please refer to the method part of the embodiment of the present invention. The computer device can be a device formed by various electronic devices.

[0083] Furthermore, the present invention also provides a computer-readable storage medium. In a computer-readable storage medium embodiment according to the present invention, the computer-readable storage medium can be configured to store a program for executing the image feature extraction method for a mobile device or the image recognition method for a mobile device according to the above-mentioned method embodiment. The program can be loaded and run by a processor to implement the above-mentioned image feature extraction method for a mobile device or the image recognition method for a mobile device. For ease of explanation, only the parts related to the embodiment of the present invention are shown. For specific technical details not disclosed, please refer to the method part of the embodiment of the present invention. The computer-readable storage medium can be a storage device formed by various electronic devices. Optionally, the computer-readable storage medium in the embodiment of the present invention is a non-temporary computer-readable storage medium.

[0084] Furthermore, the present invention provides a vehicle. In one embodiment of a vehicle according to the present invention, the vehicle may include the computer device described in the aforementioned computer device embodiment. In this embodiment, the vehicle may be an autonomous vehicle, an unmanned vehicle, or other similar vehicle. Furthermore, based on the type of power source, the vehicle in this embodiment may be a fuel vehicle, an electric vehicle, a hybrid vehicle using a combination of electric and fuel, or a vehicle using other new energy sources.

[0085] Thus far, the technical solutions of the present invention have been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art may make equivalent changes or substitutions to the relevant technical features, and the technical solutions after such changes or substitutions will fall within the scope of protection of the present invention.

Claims

1. A method for extracting image features for a mobile device, characterized in that: The method comprises: Build an initial image feature extraction network deployed on mobile devices; Replacing the depth convolution in the initial image feature extraction network with ordinary convolution; Quantizing the image feature extraction network after replacing the depth convolution to obtain a final image feature extraction network; Using the final image feature extraction network to perform feature extraction on the image; in, The step of “replacing the deep convolution in the initial image feature extraction network with ordinary convolution” specifically includes: replacing the deep convolution with a small-core ordinary convolution and performing a dilation process on the small-core ordinary convolution to increase the receptive field of the small-core ordinary convolution, wherein the small-core ordinary convolution is an ordinary convolution with a convolution kernel size smaller than a preset size threshold; Before the step of "performing quantization processing on the image feature extraction network after replacing the deep convolution to obtain a final image feature extraction network", the method also includes: replacing the special operators in the image feature extraction network after replacing the deep convolution with non-special operators, wherein the special operators refer to operators whose operation speed cannot be improved after quantization processing, and the non-special operators refer to operators whose operation speed can be improved after quantization processing.

2. The image feature extraction method for mobile devices according to claim 1, characterized in that: The steps of "building the initial image feature extraction network deployed on mobile devices" specifically include: The ConvNeXt network is used to build the initial image feature extraction network deployed on mobile devices.

3. The image feature extraction method for mobile devices according to claim 1, characterized in that: The step of “performing quantization processing on the image feature extraction network after replacing the deep convolution to obtain a final image feature extraction network” specifically includes: The floating-point parameters of the image feature extraction network after replacing the depth convolution are quantized into fixed-point parameters to obtain a final image feature extraction network.

4. The image feature extraction method for mobile devices according to claim 3, characterized in that: The step of “quantizing the floating-point parameters of the image feature extraction network after replacing the depthwise convolution into fixed-point parameters to obtain a final image feature extraction network” specifically includes: The 32-bit floating-point number parameters of the image feature extraction network after replacing the depth convolution are quantized into 8-bit fixed-point number parameters to obtain the final image feature extraction network.

5. An image recognition method for a mobile device, characterized in that: The method comprises: Acquire an image captured by the mobile device; Extracting image features of the image using the image feature extraction method for mobile devices according to any one of claims 1 to 4; Image recognition is performed according to the image features.

6. A computer device comprising a processor and a storage device, wherein the storage device is suitable for storing a plurality of program codes, characterized in that: The program code is suitable for being loaded and run by the processor to execute the image feature extraction method for a mobile device according to any one of claims 1 to 4 or the image recognition method for a mobile device according to claim 5.

7. A computer-readable storage medium storing a plurality of program codes, characterized in that: The program code is suitable for being loaded and run by a processor to execute the image feature extraction method for a mobile device according to any one of claims 1 to 4 or the image recognition method for a mobile device according to claim 5.

8. A vehicle, characterized in that: The vehicle includes the computer device of claim 6.

Images