Location-sensitive image segmentation method and apparatus

Abstract

The present disclosure provides a position-sensitive image segmentation method and device. The method comprises: constructing a position-sensitive network, constructing a down-sampling network and an up-sampling network using a plurality of residual units and the position-sensitive network, constructing a feature processing network using a plurality of residual units, constructing an image segmentation model using a plurality of down-sampling networks, the feature processing network, and a plurality of up-sampling networks; obtaining a target image to be segmented, inputting the target image into the image segmentation model; processing the target image through the plurality of down-sampling networks to obtain a target down-sampled feature map; processing the target down-sampled feature map through the feature processing network to obtain a target processed feature map; processing the target processed feature map through the plurality of up-sampling networks to obtain a target segmentation feature map; and determining a segmentation image corresponding to a target object in the target image according to the segmentation feature map. The above technical means solves the problem of poor image segmentation effect in the prior art.

Description

Technical Field

[0001] This disclosure relates to the field of target detection technology, and in particular to a location-sensitive image segmentation method and apparatus. Background Technology

[0002] The purpose of image segmentation algorithms is to segment target objects in the original image at the pixel level, that is, to extract the image portion corresponding to the target object from the original image. Existing image segmentation algorithms often use convolutional neural networks to segment images. Convolutional neural networks have a strong ability to capture local features, but because their convolution operation is a local sliding operation, it is difficult to establish global dependencies, resulting in poor image segmentation results. Summary of the Invention

[0003] In view of this, the present disclosure provides a location-sensitive image segmentation method, apparatus, electronic device, and computer-readable storage medium to solve the problem of poor image segmentation performance in the prior art.

[0004] A first aspect of this disclosure provides a position-sensitive image segmentation method, comprising: constructing a position-sensitive network; constructing a downsampling network and an upsampling network using multiple residual units and the position-sensitive network; constructing a feature processing network using multiple residual units; and constructing an image segmentation model using multiple downsampling networks, the feature processing network, and multiple upsampling networks; acquiring a target image to be segmented; inputting the target image into the image segmentation model; processing the target image through multiple downsampling networks to obtain a target downsampling feature map; processing the target downsampling feature map through the feature processing network to obtain a target processed feature map; processing the target processed feature map through multiple upsampling networks to obtain a target segmentation feature map; and determining the segmentation image corresponding to the target object in the target image based on the segmentation feature map.

[0005] A second aspect of this disclosure provides a position-sensitive image segmentation apparatus, comprising: a construction module configured to construct a position-sensitive network, construct a downsampling network and an upsampling network using multiple residual units and the position-sensitive network, construct a feature processing network using multiple residual units, and construct an image segmentation model using multiple downsampling networks, feature processing networks, and multiple upsampling networks; an acquisition module configured to acquire a target image to be segmented and input the target image into the image segmentation model; an upsampling module configured to process the target image through multiple downsampling networks to obtain a target downsampling feature map; a processing module configured to process the target downsampling feature map through the feature processing network to obtain a target processed feature map; a downsampling module configured to process the target processed feature map through multiple upsampling networks to obtain a target segmentation feature map; and a determination module configured to determine the segmentation image corresponding to a target object in the target image based on the segmentation feature map.

[0006] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0007] A fourth aspect of this disclosure provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0008] The beneficial effects of this disclosed embodiment compared to the prior art are as follows: A position-sensitive network is constructed; multiple residual units and the position-sensitive network are used to construct downsampling and upsampling networks; multiple residual units are used to construct a feature processing network; and multiple downsampling networks, feature processing networks, and upsampling networks are used to construct an image segmentation model. The target image to be segmented is obtained and input into the image segmentation model. The target image is processed by multiple downsampling networks to obtain a target downsampling feature map; the target downsampling feature map is processed by the feature processing network to obtain a target processed feature map; the target processed feature map is processed by multiple upsampling networks to obtain a target segmentation feature map; and the segmentation image corresponding to the target object in the target image is determined based on the segmentation feature map. By employing the above technical means, the problem of poor image segmentation performance in the prior art can be solved, thereby improving the image segmentation effect. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic flowchart of a position-sensitive image segmentation method provided in an embodiment of this disclosure; Figure 2 This is a schematic flowchart of a training method for an image segmentation model provided in an embodiment of this disclosure; Figure 3 This is a schematic diagram of the structure of a position-sensitive image segmentation device provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of this disclosure. However, those skilled in the art will understand that this disclosure may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this disclosure with unnecessary detail.

[0012] A location-sensitive image segmentation method and apparatus according to embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings.

[0013] Figure 1 This is a schematic flowchart of a position-sensitive image segmentation method provided in an embodiment of this disclosure. Figure 1 Location-sensitive image segmentation methods can be executed by a computer or server, or by software on a computer or server. For example... Figure 1 As shown, this location-sensitive image segmentation method includes: S101, construct a position-sensitive network, construct a downsampling network and an upsampling network using multiple residual units and the position-sensitive network, construct a feature processing network using multiple residual units, and construct an image segmentation model using multiple downsampling networks, feature processing networks and multiple upsampling networks; S102, Obtain the target image to be segmented and input the target image into the image segmentation model; S103, the target image is processed through multiple downsampling networks to obtain the target downsampling feature map; S104, The target downsampled feature map is processed by the feature processing network to obtain the target processed feature map; S105, the target processing feature map is processed through multiple upsampling networks to obtain the target segmentation feature map; S106, Determine the segmented image corresponding to the target object in the target image based on the segmentation feature map.

[0014] Position-sensitive networks segment feature maps according to different dimensions, then process, concatenate, and fuse the segmentation results to obtain new feature maps. Downsampling networks perform downsampling operations during feature map processing, hence the name. Upsampling networks perform upsampling operations during feature map processing, hence the name. Feature processing networks only process feature maps without changing their dimensions. Connecting multiple downsampling networks, feature processing networks, and multiple upsampling networks sequentially yields an image segmentation model.

[0015] In the image segmentation model: multiple downsampling networks process the target image to obtain a target downsampling feature map; a feature processing network processes the target downsampling feature map to obtain a target processed feature map; and multiple upsampling networks process the target processed feature map to obtain a target segmentation feature map. In the target segmentation feature map, each element indicates whether the corresponding position belongs to a body part of the target object. If the position belongs, the element is 1; otherwise, it is 0. Determining the segmentation image corresponding to the target object in the target image based on the segmentation feature map involves segmenting the positions with elements of 1 from the target image. The segmented image (segmentation image) is entirely about the target object; the segments that are not about the target object are separated. For example, if the target image contains one target object and five other objects, this embodiment uses an image segmentation model to obtain a segmentation feature map, and then segments the portion of the target image related to the target object based on the segmentation feature map to obtain the segmentation image.

[0016] According to the technical solution provided in the embodiments of this application, a position-sensitive network is constructed. Multiple residual units and the position-sensitive network are used to construct downsampling and upsampling networks. Multiple residual units are used to construct a feature processing network. Multiple downsampling networks, feature processing networks, and multiple upsampling networks are used to construct an image segmentation model. The target image to be segmented is obtained and input into the image segmentation model. The target image is processed by multiple downsampling networks to obtain a target downsampling feature map. The target downsampling feature map is processed by the feature processing network to obtain a target processed feature map. The target processed feature map is processed by multiple upsampling networks to obtain a target segmentation feature map. Based on the segmentation feature map, the segmentation image corresponding to the target object in the target image is determined. Using the above technical means can solve the problem of poor image segmentation effect in the prior art, thereby improving the image segmentation effect.

[0017] Furthermore, a location-sensitive network is constructed, comprising: sequentially connecting a height grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a height processing branch; sequentially connecting a width grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a width processing branch; sequentially connecting a channel grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a channel processing branch; and connecting the input side of the feature fusion layer to the output side of the parallelized height processing branch, width processing branch, and channel processing branch to obtain the location-sensitive network.

[0018] The height grouping layer divides the feature map into smaller feature maps along the image height direction (which is the same as the feature map height direction), i.e., horizontally. The width grouping layer divides the feature map into smaller feature maps along the image width direction (which is the same as the feature map width direction), i.e., vertically. The channel grouping layer divides the feature map into smaller feature maps along the feature map channel direction. The sliding window layer divides the smaller feature maps into even smaller feature maps by sliding the window multiple times. The feature dimensionality reduction layer transforms the smaller feature maps divided by the sliding window layer into one-dimensional vectors. The feature dimensionality increase layer transforms the one-dimensional vector processed by the fully connected layer into a multi-dimensional feature map. The feature map output by the feature dimensionality increase layer and the feature map output by the sliding window layer have the same dimension.

[0019] Furthermore, a downsampling network and an upsampling network are constructed using multiple residual units and a position-sensitive network. A feature processing network is also constructed using multiple residual units. Finally, an image segmentation model is constructed using multiple downsampling networks, a feature processing network, and multiple upsampling networks. This includes: sequentially connecting multiple residual units and a position-sensitive network to obtain a downsampling network, where the downsampling network downsamples the feature map by adjusting the step size of multiple residual units; sequentially connecting a position-sensitive network and multiple residual units to obtain an upsampling network, where the upsampling network upsamples the feature map by adjusting the step size of multiple residual units; using multiple serially connected residual units as a feature processing network; and sequentially connecting multiple downsampling networks, a feature processing network, and multiple upsampling networks to obtain the image segmentation model.

[0020] The residual unit is an existing network structure that consists of multiple convolutional layers and a shortcut connection. The shortcut connection directly passes the input of the residual unit to the last layer of the residual unit, allowing the input to be added to the output of the convolutional layers, which is then used as the output of the residual unit.

[0021] Furthermore, a downsampling network and an upsampling network are constructed using multiple residual units and a position-sensitive network. A feature processing network is also constructed using multiple residual units. Finally, an image segmentation model is constructed using multiple downsampling networks, a feature processing network, and multiple upsampling networks. This includes: sequentially connecting two residual units and a position-sensitive network to obtain a downsampling network, wherein the step size of the first residual unit in the downsampling network is a first preset step size, the step size of the second residual unit is a second preset step size, and the second preset step size is a preset multiple of the first preset step size; sequentially connecting a position-sensitive network and two residual units to obtain an upsampling network, wherein the step size of both residual units in the upsampling network is the first preset step size; using three serially connected residual units as a feature processing network; and sequentially connecting two downsampling networks, a feature processing network, and two upsampling networks to obtain the image segmentation model.

[0022] Assuming the first preset stride is 1, then the second preset stride is 2. In the downsampling network, the stride of the first residual unit is 1, and the stride of the second residual unit is 2. Therefore, the downsampling network can achieve a 2x downsampling during feature map processing. The convolutional kernels used by the two residual units can be chosen according to standard settings, for example, a 1x1 kernel for the first residual unit and a 3x3 kernel for the second. The upsampling network uses two residual units, each with a stride of 1. The upsampling network can achieve a 2x upsampling during feature map processing. The convolutional kernels used by the two residual units are 1x1 and 3x3. Using three serially connected residual units as a feature processing network, the feature processing network only processes the feature map without changing its size.

[0023] Furthermore, the input and output of the position-sensitive network are denoted as the first feature map and the second feature map, respectively. This includes: within the position-sensitive network: processing the first feature map through a height processing branch, a width processing branch, and a channel processing branch to obtain a height processing feature map, a width processing feature map, and a channel processing feature map, respectively; and fusing the height processing feature map, the width processing feature map, and the channel processing feature map through a feature fusion layer to obtain the second feature map.

[0024] Both downsampling and upsampling networks include position-sensitive networks, and the role of the position-sensitive network is the same in both. Let the input of any position-sensitive network be denoted as the first feature map. The first feature map is processed through height processing branches, width processing branches, and channel processing branches to obtain height-processed feature maps, width-processed feature maps, and channel-processed feature maps. The feature fusion layer fuses the height-processed feature maps, width-processed feature maps, and channel-processed feature maps to obtain the second feature map output by the position-sensitive network.

[0025] Furthermore, the first feature map is processed through a height processing branch to obtain a height-processed feature map, including: dividing the first feature map into multiple third feature maps in the height dimension through a height grouping layer; dividing each third feature map into multiple fourth feature maps by sliding a window layer according to a preset size; transforming each fourth feature map into a one-dimensional feature vector through a feature dimensionality reduction layer; processing each feature vector through a fully connected layer to obtain a fully connected vector corresponding to each feature vector; transforming each fully connected vector into a fifth feature map through a feature dimensionality increase layer, wherein each fifth feature map has the same dimension as each fourth feature map; and concatenating each fifth feature map through a feature concatenation layer to obtain the height-processed feature map.

[0026] The process of processing the first feature map through the width processing branch to obtain the width-processed feature map, and the process of processing the first feature map through the channel processing branch to obtain the channel-processed feature map, is similar to the process of processing the first feature map through the height processing branch to obtain the height-processed feature map. The process of processing the first feature map through the width processing branch and the channel processing branch will not be described again.

[0027] Figure 2 This is a flowchart illustrating a training method for an image segmentation model provided in an embodiment of this disclosure, as shown below. Figure 2 As shown, the method includes: S201, Obtain the training image and input it into the image segmentation model: S202, the training image is processed through multiple downsampling networks to obtain downsampling feature maps; S203, the downsampled feature map is processed by a feature processing network to obtain the processed feature map; S204, the feature map is processed through multiple upsampling networks to obtain the segmentation feature map; S205 uses the divergence loss function to calculate the loss between the segmentation feature map and the labels of the training image, and optimizes the image segmentation model based on the loss to complete the training of the image segmentation model.

[0028] This application's embodiment pertains to the training phase of an image segmentation model. The model is trained using training images. Within the model: multiple downsampling networks process the training images to obtain downsampling feature maps; a feature processing network processes the downsampling feature maps to obtain processed feature maps; and multiple upsampling networks process the processed feature maps to obtain segmentation feature maps. A divergence loss function is used to calculate the loss between the segmentation feature maps and the labels on the training images, and the image segmentation model is optimized based on this loss. The labels on the training images are pre-annotated feature maps that represent the segmentation features.

[0029] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0030] The following are embodiments of the apparatus disclosed herein, which can be used to execute embodiments of the method disclosed herein. For details not disclosed in the apparatus embodiments of this disclosure, please refer to the embodiments of the method disclosed herein.

[0031] Figure 3 This is a schematic diagram of a position-sensitive image segmentation device provided in an embodiment of this disclosure. Figure 3 As shown, the location-sensitive image segmentation device includes: The construction module 301 is configured to construct a position-sensitive network, construct a downsampling network and an upsampling network using multiple residual units and the position-sensitive network, construct a feature processing network using multiple residual units, and construct an image segmentation model using multiple downsampling networks, feature processing networks and multiple upsampling networks; The acquisition module 302 is configured to acquire the target image to be segmented and input the target image into the image segmentation model. The downsampling module 303 is configured to process the target image through multiple downsampling networks to obtain a target downsampling feature map; Processing module 304 is configured to process the target downsampled feature map through a feature processing network to obtain the target processed feature map; The upsampling module 305 is configured to process the target processing feature map through multiple upsampling networks to obtain the target segmentation feature map; The determination module 306 is configured to determine the segmented image corresponding to the target object in the target image based on the segmentation feature map.

[0032] According to the technical solution provided in the embodiments of this application, a position-sensitive network is constructed. Multiple residual units and the position-sensitive network are used to construct downsampling and upsampling networks. Multiple residual units are used to construct a feature processing network. Multiple downsampling networks, feature processing networks, and multiple upsampling networks are used to construct an image segmentation model. The target image to be segmented is obtained and input into the image segmentation model. The target image is processed by multiple downsampling networks to obtain a target downsampling feature map. The target downsampling feature map is processed by the feature processing network to obtain a target processed feature map. The target processed feature map is processed by multiple upsampling networks to obtain a target segmentation feature map. Based on the segmentation feature map, the segmentation image corresponding to the target object in the target image is determined. Using the above technical means can solve the problem of poor image segmentation effect in the prior art, thereby improving the image segmentation effect.

[0033] In some embodiments, the construction module 301 is further configured to sequentially connect a height grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a height processing branch; sequentially connect a width grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a width processing branch; sequentially connect a channel grouping layer, a window sliding layer, a feature dimensionality reduction layer, a fully connected layer, a feature dimensionality increase layer, and a feature concatenation layer to obtain a channel processing branch; and after connecting the input side of the feature fusion layer to the output side of the parallel height processing branch, the width processing branch, and the channel processing branch, a position-sensitive network is obtained.

[0034] In some embodiments, the construction module 301 is further configured to sequentially connect multiple residual units and a position-sensitive network to obtain a downsampling network, wherein the downsampling network downsamples the feature map by adjusting the step size of multiple residual units; sequentially connect the position-sensitive network and multiple residual units to obtain an upsampling network, wherein the upsampling network upsamples the feature map by adjusting the step size of multiple residual units; use multiple serially connected residual units as a feature processing network; and sequentially connect multiple downsampling networks, feature processing networks, and multiple upsampling networks to obtain an image segmentation model.

[0035] In some embodiments, the construction module 301 is further configured to sequentially connect two residual units and a position-sensitive network to obtain a downsampling network, wherein the step size of the first residual unit in the downsampling network is a first preset step size, the step size of the second residual unit is a second preset step size, and the second preset step size is a preset multiple of the first preset step size; sequentially connect the position-sensitive network and two residual units to obtain an upsampling network, wherein the step size of the two residual units in the upsampling network is the first preset step size; use the three serially connected residual units as a feature processing network; and sequentially connect the two downsampling networks, the feature processing network, and the two upsampling networks to obtain an image segmentation model.

[0036] In some embodiments, the downsampling module 303 and the upsampling module 305 are further configured to process the first feature map through the height processing branch, the width processing branch and the channel processing branch respectively within the position-sensitive network to obtain the height processing feature map, the width processing feature map and the channel processing feature map; and to fuse the height processing feature map, the width processing feature map and the channel processing feature map through the feature fusion layer to obtain the second feature map.

[0037] In some embodiments, the downsampling module 303 and the upsampling module 305 are further configured to: divide the first feature map into multiple third feature maps in the height dimension of the feature map through a height grouping layer; divide each third feature map into multiple fourth feature maps by sliding a window sliding layer according to a preset size; transform each fourth feature map into a one-dimensional feature vector through a feature dimensionality reduction layer; process each feature vector through a fully connected layer to obtain a fully connected vector corresponding to each feature vector; transform each fully connected vector into a fifth feature map through a feature dimensionality increase layer, wherein each fifth feature map has the same dimension as each fourth feature map; and concatenate each fifth feature map through a feature concatenation layer to obtain a height-processed feature map.

[0038] In some embodiments, the acquisition module 302 is further configured to acquire training images and input the training images into the image segmentation model: process the training images through multiple downsampling networks to obtain downsampling feature maps; process the downsampling feature maps through a feature processing network to obtain processed feature maps; process the processed feature maps through multiple upsampling networks to obtain segmentation feature maps; calculate the loss between the segmentation feature maps and the labels of the training images using a divergence loss function, and optimize the image segmentation model based on the loss to complete the training of the image segmentation model.

[0039] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this disclosure.

[0040] Figure 4 This is a schematic diagram of the electronic device 4 provided in an embodiment of this disclosure. Figure 4 As shown, the electronic device 4 of this embodiment includes: a processor 401, a memory 402, and a computer program 403 stored in the memory 402 and executable on the processor 401. When the processor 401 executes the computer program 403, it implements the steps in the various method embodiments described above. Alternatively, when the processor 401 executes the computer program 403, it implements the functions of each module / unit in the various device embodiments described above.

[0041] Electronic device 4 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 4 may include, but is not limited to, processor 401 and memory 402. Those skilled in the art will understand that... Figure 4 This is merely an example of electronic device 4 and does not constitute a limitation on electronic device 4. It may include more or fewer components than shown, or different components.

[0042] The processor 401 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0043] The memory 402 can be an internal storage unit of the electronic device 4, such as a hard disk or RAM of the electronic device 4. The memory 402 can also be an external storage device of the electronic device 4, such as a plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, etc., equipped on the electronic device 4. The memory 402 can also include both internal and external storage units of the electronic device 4. The memory 402 is used to store computer programs and other programs and data required by the electronic device.

[0044] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments 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.

[0045] If an integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable medium may include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in a computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in a jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0046] The above embodiments are only used to illustrate the technical solutions of this disclosure, and are not intended to limit it. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be included within the protection scope of this disclosure.

Claims

1. A location-sensitive image segmentation method, characterized in that, include: A position-sensitive network is constructed, and a downsampling network and an upsampling network are constructed using multiple residual units and the position-sensitive network. A feature processing network is constructed using multiple residual units, and an image segmentation model is constructed using multiple downsampling networks, the feature processing network, and multiple upsampling networks. Obtain the target image to be segmented, and input the target image into the image segmentation model: The target image is processed by multiple downsampling networks to obtain a target downsampling feature map; The target downsampled feature map is processed by the feature processing network to obtain the target processed feature map; The target processing feature map is processed by multiple upsampling networks to obtain the target segmentation feature map; The segmented image corresponding to the target object in the target image is determined based on the segmentation feature map; The construction of location-sensitive networks includes: The height processing branch is obtained by sequentially connecting the height grouping layer, the first window sliding layer, the first feature dimensionality reduction layer, the first fully connected layer, the first feature dimensionality increase layer, and the first feature splicing layer. By sequentially connecting the width grouping layer, the second window sliding layer, the second feature dimensionality reduction layer, the second fully connected layer, the second feature dimensionality increase layer, and the second feature splicing layer, the width processing branch is obtained. By sequentially connecting the channel grouping layer, the third window sliding layer, the third feature dimensionality reduction layer, the third fully connected layer, the third feature dimensionality increase layer, and the third feature splicing layer, the channel processing branch is obtained. The location-sensitive network is obtained by connecting the input side of the feature fusion layer to the output side of the parallel height processing branch, width processing branch, and channel processing branch.

2. The method according to claim 1, characterized in that, A downsampling network and an upsampling network are constructed using multiple residual units and the position-sensitive network. A feature processing network is constructed using multiple residual units. An image segmentation model is constructed using multiple downsampling networks, the feature processing network, and multiple upsampling networks, including: By sequentially connecting multiple residual units and the position-sensitive network, the downsampling network is obtained, wherein the downsampling network achieves downsampling of the feature map by adjusting the step size of multiple residual units; The position-sensitive network and multiple residual units are connected sequentially to obtain the upsampling network, wherein the upsampling network achieves upsampling of the feature map by adjusting the step size of multiple residual units; Multiple serially connected residual units are used as the feature processing network; The image segmentation model is obtained by sequentially connecting multiple downsampling networks, the feature processing network, and multiple upsampling networks.

3. The method according to claim 1, characterized in that, A downsampling network and an upsampling network are constructed using multiple residual units and the position-sensitive network. A feature processing network is constructed using multiple residual units. An image segmentation model is constructed using multiple downsampling networks, the feature processing network, and multiple upsampling networks, including: By sequentially connecting two residual units and the position-sensitive network, the downsampling network is obtained, wherein the step size of the first residual unit in the downsampling network is a first preset step size, the step size of the second residual unit is a second preset step size, and the second preset step size is a preset multiple of the first preset step size. The position-sensitive network and two residual units are connected sequentially to obtain the upsampling network, wherein the step size of the two residual units in the upsampling network is a first preset step size; The feature processing network consists of three serially connected residual units. The image segmentation model is obtained by sequentially connecting two downsampling networks, the feature processing network, and two upsampling networks.

4. The method according to claim 1, characterized in that, The input and output of the location-sensitive network are denoted as the first feature map and the second feature map, respectively, including: Inside the location-sensitive network: The first feature map is processed through the height processing branch, the width processing branch, and the channel processing branch respectively to obtain a height processing feature map, a width processing feature map, and a channel processing feature map; The second feature map is obtained by fusing the height processing feature map, the width processing feature map, and the channel processing feature map through the feature fusion layer.

5. The method according to claim 4, characterized in that, The first feature map is processed through the height processing branch to obtain a height-processed feature map, including: The height grouping layer divides the first feature map into multiple third feature maps along the height dimension of the feature map. The window sliding layer slides across each third feature map according to a preset size to divide each third feature map into multiple fourth feature maps; The feature reduction layer transforms each fourth feature map into a one-dimensional feature vector. Each feature vector is processed by the fully connected layer to obtain the fully connected vector corresponding to each feature vector; The feature upscaling layer transforms each fully connected vector into a fifth feature map, wherein each fifth feature map has the same dimension as each fourth feature map. The high-level processed feature map is obtained by stitching together the various fifth feature maps through the feature stitching layer.

6. The method according to claim 1, characterized in that, Before inputting the target image into the image segmentation model, the method further includes: Obtain training images and input them into the image segmentation model: The training images are processed by multiple downsampling networks to obtain downsampling feature maps; The downsampled feature map is processed by the feature processing network to obtain the processed feature map; The processed feature map is processed by multiple upsampling networks to obtain a segmentation feature map; The loss between the segmentation feature map and the labels of the training image is calculated using the divergence loss function, and the image segmentation model is optimized based on the loss to complete the training of the image segmentation model.

7. A position-sensitive image segmentation device, characterized in that, include: The construction module is configured to construct a position-sensitive network, construct a downsampling network and an upsampling network using multiple residual units and the position-sensitive network, construct a feature processing network using multiple residual units, and construct an image segmentation model using multiple downsampling networks, the feature processing network, and multiple upsampling networks; The location-sensitive network is constructed by sequentially connecting a height grouping layer, a first window sliding layer, a first feature dimensionality reduction layer, a first fully connected layer, a first feature dimensionality increase layer, and a first feature concatenation layer to obtain a height processing branch; sequentially connecting a width grouping layer, a second window sliding layer, a second feature dimensionality reduction layer, a second fully connected layer, a second feature dimensionality increase layer, and a second feature concatenation layer to obtain a width processing branch; sequentially connecting a channel grouping layer, a third window sliding layer, a third feature dimensionality reduction layer, a third fully connected layer, a third feature dimensionality increase layer, and a third feature concatenation layer to obtain a channel processing branch; and connecting the input side of the feature fusion layer to the output side of the parallelized height processing branch, width processing branch, and channel processing branch to obtain the location-sensitive network. The acquisition module is configured to acquire the target image to be segmented and input the target image into the image segmentation model. The downsampling module is configured to process the target image through multiple downsampling networks to obtain a target downsampling feature map; The processing module is configured to process the target downsampled feature map through the feature processing network to obtain the target processed feature map; The upsampling module is configured to process the target processing feature map through multiple upsampling networks to obtain a target segmentation feature map; The determination module is configured to determine the segmented image corresponding to the target object in the target image based on the segmentation feature map.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 6.

Images