Multi-category target semantic segmentation method and device in disordered aliasing scene and medium

Abstract

The invention discloses a multi-category target semantic segmentation method and device in a disordered aliasing scene and a medium, and relates to the technical field of semantic segmentation, and the method comprises the steps: constructing a reference data set; a reference neural network model with a U-Net network as a prototype is constructed, the reference neural network model comprises a down-sampling network and an up-sampling network, the down-sampling network is a Resnet50 network integrated with a channel attention mechanism, and the up-sampling network is a Resnet50 network integrated with a channel attention mechanism. The up-sampling network adopts a bilinear interpolation up-sampling mode integrated with a coordinate attention mechanism or adopts a bilinear interpolation up-sampling mode integrated with Ghost convolution and the coordinate attention mechanism; training and optimizing the reference neural network model through the reference data set to generate a target neural network model; and inputting a target image into the target neural network model to carry out pixel-by-pixel semantic segmentation so as to output a segmentation result. With the adoption of the method, the detection and segmentation efficiency of machine vision on the multi-category disorderly placed target objects can be effectively improved, and the calculation amount of an algorithm is greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of semantic segmentation technology, and in particular to a method, device and medium for semantic segmentation of multi-category targets in a cluttered aliasing scene. Background Art

[0002] In the current field of industrial robot-assisted assembly and processing, the complex and changeable lighting conditions at the production site and the wide variety of spare parts that are stacked in an unordered manner have significantly increased the processing difficulty of machine vision detection algorithms, especially for tasks such as accurate segmentation of multiple categories of spare parts and accurate estimation of spatial pose. Existing machine vision detection algorithms have encountered obstacles, which directly limits the efficient application and widespread promotion of industrial robots and their supporting automated sorting equipment in the sorting and loading links of assembly lines, and hinders the further development of the industrial automation process.

[0003] Accurately estimating the three-dimensional spatial pose of multi-category targets in cluttered and overlapping scenes is an indispensable prerequisite for ensuring that robots can achieve autonomous, efficient sorting and precise assembly. Accurately extracting the point cloud of the target is an important means of estimating the spatial pose of the target, and the accuracy of the target point cloud segmentation depends on the segmentation accuracy of the target object in the two-dimensional image. Especially for cluttered and overlapping scenes with multiple categories of targets, existing machine vision methods are difficult to achieve accurate segmentation of target objects, and the algorithm has high computational complexity, which leads to a decrease in the accuracy and efficiency of the spatial pose estimation of the target point cloud segmentation, restricting the high-quality development of the autonomous sorting industry. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method, device and medium for semantic segmentation of multi-category targets in a cluttered aliasing scene, which can effectively improve the segmentation performance and greatly reduce the floating-point calculation amount of the model.

[0005] In order to solve the above technical problems, the present invention provides a method for semantic segmentation of multi-category targets in a cluttered aliasing scenario, comprising: constructing a benchmark data set; constructing a benchmark neural network model based on a U-Net network as a prototype, the benchmark neural network model comprising a downsampling network and an upsampling network, wherein the downsampling network is a Resnet50 network integrated with a channel attention mechanism, and the upsampling network adopts a bilinear interpolation upsampling method integrated with a coordinate attention mechanism or a bilinear interpolation upsampling method integrated with Ghost convolution and a coordinate attention mechanism; training and optimizing the benchmark neural network model through the benchmark data set to generate a target neural network model; inputting the target image into the target neural network model for pixel-by-pixel semantic segmentation to output a segmentation result.

[0006] As an improvement of the above scheme, the downsampling network includes a convolutional layer, a maximum pooling layer, a first ECA residual module, a second ECA residual module, a third ECA residual module and a fourth ECA residual module arranged in sequence; the first ECA residual module includes three ECA residual blocks, the second ECA residual module includes four ECA residual blocks, the third ECA residual module includes six ECA residual blocks, and the fourth ECA residual module includes three ECA residual blocks; the ECA residual block includes a 1*1 convolutional layer, a 3*3 convolutional layer, a 1*1 convolutional layer and an ECA channel attention mechanism arranged in sequence.

[0007] As an improvement of the above scheme, the steps of constructing the downsampling network include: constructing a Resnet50 network and using the Resnet50 network as the downsampling network of the U-Net network; setting a channel attention mechanism in each residual block of the Resnet50 network to form an ECA residual block.

[0008] As an improvement of the above scheme, the upsampling network includes a coordinate attention mechanism, a first decoding module, a second decoding module, a third decoding module, a fourth decoding module and an output layer arranged in sequence; the first decoding module, the second decoding module and the third decoding module all include bilinear interpolation upsampling, two 3*3 convolutional layers and a coordinate attention mechanism arranged in sequence; the fourth decoding module includes bilinear interpolation upsampling, two 3*3 convolutional layers and bilinear interpolation upsampling arranged in sequence; the output layer includes two 3*3 convolutional layers and a 1*1 convolutional layer arranged in sequence.

[0009] As an improvement of the above solution, the 3*3 convolutional layers in the first decoding module, the second decoding module, the third decoding module and the fourth decoding module are 3*3 Ghost convolutional layers.

[0010] As an improvement of the above scheme, the steps of constructing an upsampling network using a bilinear interpolation upsampling method that incorporates a coordinate attention mechanism include: constructing an upsampling network of a U-Net network, the upsampling network including four decoding modules, deconvolution upsampling and an output layer arranged in sequence, the decoding module including a deconvolution upsampling and two 3*3 convolution layers arranged in sequence; modifying the deconvolution upsampling in the upsampling network of the U-Net network to bilinear interpolation upsampling; and setting a coordinate attention mechanism before the bilinear interpolation upsampling of the decoding module.

[0011] As an improvement of the above scheme, the steps of constructing an upsampling network using a bilinear interpolation upsampling method that incorporates Ghost convolution and a coordinate attention mechanism include: constructing an upsampling network of a U-Net network, the upsampling network including four decoding modules, a deconvolution upsampling and an output layer arranged in sequence, the decoding module including a deconvolution upsampling and two 3*3 convolution layers arranged in sequence; modifying the deconvolution upsampling in the upsampling network of the U-Net network to bilinear interpolation upsampling; setting a coordinate attention mechanism before the bilinear interpolation upsampling of the decoding module; and modifying the two 3*3 convolutions of the decoding module to two 3*3 Ghost convolutions.

[0012] As an improvement of the above scheme, the step of constructing a benchmark data set includes: taking the Linemod_preprocessed data set as the benchmark data set; mapping the pixel values ​​of each target object in the benchmark data set in turn to map the pixel value of the Nth target object to N, where N is a positive integer less than 256; and dividing the benchmark data set after mapping into a training data set and a test data set.

[0013] Correspondingly, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the processor executes the computer program, the steps of the multi-category target semantic segmentation method in the above-mentioned cluttered aliasing scenario are implemented.

[0014] Correspondingly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the above-mentioned method for semantic segmentation of multi-category objects in a cluttered aliasing scenario are implemented.

[0015] The implementation of the present invention has the following beneficial effects: The present invention adopts a uniquely structured benchmark neural network model, in which a channel attention mechanism is incorporated into the benchmark neural network model for downsampling, thereby capturing the relationship between different channels and improving the ability of feature representation; at the same time, a coordinate attention mechanism is also incorporated into the benchmark neural network model for upsampling, thereby enhancing the representation of the region of interest, so as to improve the network model's insufficient ability to connect the semantic information of the target context.

[0016] Furthermore, the benchmark neural network model in the present invention is based on the U-Net network as a prototype, and the original downsampling network of the U-Net network is replaced by the Resnet50 network, and each residual block is improved, and a channel attention mechanism is introduced therein, which can effectively extract the most original and important feature information of the entire recognition model, and screen the original feature set to highlight important features, eliminate redundant features, and improve the efficiency of the entire model; at the same time, the present invention also changes the deconvolution upsampling method in the existing U-Net network to a bilinear interpolation upsampling method with stronger feature restoration ability, and introduces a coordinate attention mechanism before each bilinear interpolation upsampling operation to further enhance the representation of the region of interest, thereby upsampling the size of the feature map to a higher level to increase the detail information of the feature map, so that the network can better capture the detail information in the image.

[0017] More preferably, the present invention also replaces the two 3*3 convolutions of the decoding module with two 3*3 Ghost convolutions, which can greatly reduce the computational complexity and the floating-point operations of the network while slightly improving the network detection performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of an embodiment of a method for semantic segmentation of multi-category objects in a cluttered aliasing scene of the present invention; Figure 2 It is a schematic diagram of the structure of the down-sampling network in the present invention; Figure 3 Schematic diagram of the structure of the ECA residual block in the present invention; Figure 4 It is a structural schematic diagram of the benchmark neural network model in the present invention. DETAILED DESCRIPTION

[0019] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention will be further described in detail below with reference to the accompanying drawings.

[0020] See also Figure 1 , Figure 1 The flowchart of an embodiment of the method for semantic segmentation of multi-category objects in a cluttered aliasing scene of the present invention is shown, which includes: S101, build benchmark dataset; Accordingly, the steps to construct the benchmark dataset include: (1) Use the Linemod_preprocessed dataset as the benchmark dataset; (2) Mapping the pixel values ​​of each target object in the benchmark data set in turn, so as to map the pixel value of the Nth target object to N, where N is a positive integer less than 256; In this embodiment, the public dataset Linemod_preprocessed is used as the benchmark dataset, but this is not a limitation and can be selected according to actual conditions.

[0021] It should be noted that, in the Linemod_preprocessed data set, the pixel value of target object 1 is 21, the pixel value of target object 2 is 43, the pixel value of target object 3 is 106, the pixel value of target object 4 is 128, the pixel value of target object 5 is 170, the pixel value of target object 6 is 191, the pixel value of target object 7 is 213, the pixel value of target object 8 is 234, the pixel value of target object 9 is 255…, and the present invention needs to preprocess the Linemod_preprocessed data set.

[0022] Accordingly, the specific preprocessing steps are: mapping the pixel value of target object 1 in all label images to 1, mapping the pixel value of target object 2 to 2, and so on.

[0023] (3) Divide the mapped benchmark dataset into a training dataset and a test dataset.

[0024] Finally, the preprocessed benchmark dataset is divided into a training dataset and a test dataset required for subsequent steps.

[0025] S102, constructing a benchmark neural network model; The present invention adopts a uniquely structured benchmark neural network model, into which a channel attention mechanism is incorporated for downsampling, and a coordinate attention mechanism is incorporated for upsampling.

[0026] Accordingly, the benchmark neural network model in the present invention is based on the U-Net network as a prototype. Since the U-Net network includes a downsampling network and an upsampling network connected to each other, the benchmark neural network model in the present invention also includes a downsampling network and an upsampling network connected to each other, wherein: The downsampling network is a Resnet50 network that incorporates the channel attention mechanism; The upsampling network adopts a bilinear interpolation upsampling method that incorporates a coordinate attention mechanism or a bilinear interpolation upsampling method that incorporates Ghost convolution and a coordinate attention mechanism.

[0027] The following are detailed descriptions of the Resnet50 network (ECA-Resnet50) that incorporates the channel attention mechanism, the upsampling network that incorporates the bilinear interpolation upsampling method (CA-BilinearUpsampling) of the coordinate attention mechanism, and the upsampling network that incorporates the Ghost convolution and coordinate attention mechanism (CA-BilinearUpsampling-Ghost): 1. Resnet50 network with channel attention mechanism like Figure 2 and Figure 3 As shown, the Resnet50 network (i.e., the downsampling network) incorporating the channel attention mechanism includes a convolutional layer, a maximum pooling layer, a first ECA residual module, a second ECA residual module, a third ECA residual module, and a fourth ECA residual module arranged in sequence; wherein, the first ECA residual module includes three ECA residual blocks, the second ECA residual module includes four ECA residual blocks, the third ECA residual module includes six ECA residual blocks, and the fourth ECA residual module includes three ECA residual blocks; the ECA residual block includes a 1*1 convolutional layer, a 3*3 convolutional layer, a 1*1 convolutional layer, and an ECA channel attention mechanism (i.e., ECA-Net) arranged in sequence.

[0028] Furthermore, the activation functions BN (Batch Normalization) and ReLU (Rectified Linear Unit) are used in the ECA residual block.

[0029] Accordingly, the steps for building a Resnet50 network with channel attention mechanism include: (1) Construct a Resnet50 network and use it as the downsampling network of the original U-Net network; (2) An ECA (Efficient Channel Attention) channel attention mechanism is set in each residual block of the Resnet50 network to form an ECA residual block.

[0030] Aiming at the complex target object features processed by robots in actual industrial application scenarios, the present invention replaces the original downsampling network of the U-Net network with the Resnet50 network, and improves each residual block based on the Resnet50 network: the ECA channel attention mechanism is introduced to capture the relationship between different channels, thereby improving the feature representation capability. It should be noted that the core goal of the Resnet50 network that incorporates the channel attention mechanism is to accurately capture the most core and critical feature layer of the target object at the initial stage.

[0031] Therefore, the Resnet50 network that incorporates the channel attention mechanism can use the convolutional neural network to extract the most original and important feature information of the entire recognition model, reduce the size of the feature map to a higher level, and integrate ECA-Net into the feature extraction network to screen the original feature set, highlight important features, eliminate redundant features, and improve the efficiency of the entire model. This step constitutes the original feature set, and the subsequent steps are based on this feature set.

[0032] 2. Upsampling network with bilinear interpolation upsampling method integrating coordinate attention mechanism The upsampling network of the bilinear interpolation upsampling method integrating the coordinate attention mechanism includes a coordinate attention mechanism, a first decoding module, a second decoding module, a third decoding module, a fourth decoding module and an output layer arranged in sequence; wherein the first decoding module, the second decoding module and the third decoding module all include bilinear interpolation upsampling, two 3*3 convolutional layers and a coordinate attention mechanism arranged in sequence; the fourth decoding module includes bilinear interpolation upsampling, two 3*3 convolutional layers and bilinear interpolation upsampling arranged in sequence; the output layer includes two 3*3 convolutional layers and a 1*1 convolutional layer arranged in sequence.

[0033] Accordingly, the steps for constructing an upsampling network using bilinear interpolation upsampling with coordinate attention mechanism include: (1) constructing an upsampling network of the U-Net network, wherein the upsampling network includes four decoding modules, a deconvolution upsampling layer and an output layer arranged in sequence, and the decoding module includes a deconvolution upsampling layer and two 3*3 convolution layers arranged in sequence; (2) Modify the deconvolution upsampling in the upsampling network of the U-Net network to bilinear interpolation upsampling; The deconvolution upsampling method in the upsampling network of the existing U-Net network is changed to a bilinear interpolation upsampling method with stronger feature restoration capability; (3) Set the CA (Coordinate Attention) coordinate attention mechanism before the bilinear interpolation upsampling of the decoding module.

[0034] The CA coordinate attention mechanism is introduced before each bilinear interpolation upsampling operation to further enhance the representation of the region of interest, so as to improve the insufficient ability of the U-Net network to connect the semantic information of the target context.

[0035] Therefore, the bilinear interpolation upsampling method that incorporates the coordinate attention mechanism can upsample the size of the feature map to a higher level to increase the detail information of the feature map, allowing the network to better capture the detail information in the image.

[0036] 3. Upsampling network with bilinear interpolation upsampling method integrating Ghost convolution and coordinate attention mechanism like Figure 4 As shown, the upsampling network of the bilinear interpolation upsampling method integrating Ghost convolution and coordinate attention mechanism includes a coordinate attention mechanism, a first decoding module, a second decoding module, a third decoding module, a fourth decoding module and an output layer arranged in sequence; the first decoding module, the second decoding module and the third decoding module all include bilinear interpolation upsampling, two 3*3Ghost convolution layers and a coordinate attention mechanism arranged in sequence; the fourth decoding module includes bilinear interpolation upsampling, two 3*3Ghost convolution layers and bilinear interpolation upsampling arranged in sequence; the output layer includes two 3*3 convolution layers and a 1*1 convolution layer arranged in sequence.

[0037] Accordingly, the construction steps of the bilinear interpolation upsampling method that incorporates Ghost convolution and coordinate attention mechanism include: (1) constructing an upsampling network of the U-Net network, wherein the upsampling network includes four decoding modules, a deconvolution upsampling layer and an output layer arranged in sequence, and the decoding module includes a deconvolution upsampling layer and two 3*3 convolution layers arranged in sequence; (2) Modify the deconvolution upsampling in the upsampling network of the U-Net network to bilinear interpolation upsampling; (3) Set the CA coordinate attention mechanism before the bilinear interpolation upsampling of the decoding module.

[0038] (4) Modify the two 3*3 convolutions of the decoding module into two 3*3 Ghost convolutions.

[0039] It should be noted that in the existing U-Net network, the feature matrix will be spliced ​​after the upsampling operation, and then two 3*3 convolution operations will be performed to extract more useful information; the present invention replaces these two 3*3 convolutions with Ghost convolution, which can generate a lot of required information from the original features at a very low cost, while greatly reducing the computational complexity and the floating-point operations of the network, and slightly improving the network detection performance.

[0040] Therefore, the bilinear interpolation upsampling method that incorporates Ghost convolution and coordinate attention mechanism can effectively improve the model's feature restoration ability in upsampling operations, greatly reduce the number of floating-point operations, and improve the model's computing efficiency and speed.

[0041] S103, training and optimizing the benchmark neural network model using the benchmark data set to generate a target neural network model; The benchmark data set constructed in step S101 is used to perform multiple iterations of training and parameter adjustment on the benchmark neural network model; at the same time, the model parameter fitting is used during the iteration process to achieve better model performance.

[0042] S104, inputting the target image into the target neural network model for pixel-by-pixel semantic segmentation to output a segmentation result.

[0043] It can be seen from the above that the method for semantic segmentation of multi-category targets in cluttered aliasing scenarios of the present invention proposes a more accurate and efficient object semantic segmentation method in cluttered aliasing scenarios of multi-category targets on the basis of the existing U-Net network, and improves the existing network structure in three aspects: integrating the ECA channel attention mechanism into the Resnet50 network (downsampling network) that integrates the channel attention mechanism, integrating the CA coordinate attention mechanism into the bilinear interpolation upsampling method that integrates Ghost convolution and coordinate attention mechanism, and replacing the 3x3 convolution in the upsampling network, strengthening the feature extraction network, improving the segmentation performance, and greatly reducing the floating-point calculation amount of the model. The computational efficiency and speed of the model are also improved, providing theoretical support for the promotion and application of industrial robots in processing and assembly environments.

[0044] Correspondingly, the present invention also discloses a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the processor executes the computer program, the steps of the multi-category target semantic segmentation method in the above-mentioned cluttered aliasing scenario are implemented.

[0045] At the same time, the present invention also discloses a computer-readable storage medium on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the multi-category target semantic segmentation method in the above-mentioned cluttered aliasing scene are implemented.

[0046] The effects of the present invention are further described below in conjunction with specific comparative examples: 1. The multi-category target semantic segmentation method (U-Net+ECA-ResNet50+CA-BilinearUpsampling-Ghost) in the cluttered aliasing scene of the present invention is compared with the visual segmentation algorithm based on the U-Net model. The segmentation results are shown in Table 1 below: Table 1

[0047] It can be seen from Table 1 above that the floating-point operations (FLOPs) of the multi-category target semantic segmentation method in the cluttered aliasing scenario of the present invention is only 29.6% of that of the U-Net model.

[0048] 2. The multi-category target semantic segmentation method in the cluttered aliasing scene of the present invention is compared with several visual segmentation algorithms based on DeepLabV3+, RefineNet, U-Net+ResNet50, U-Net+ECA-ResNet50 and U-Net models. The segmentation results of each algorithm for the same scene are shown in Table 2 below: Table 2

[0049] It can be seen from Tables 1 and 2 above that the multi-category target semantic segmentation method in the cluttered aliasing scene of the present invention has improvements in the four evaluation indicators of accuracy (Accuracy), F1 score (F1-Score), mean intersection-over-union (mIou), and category average pixel accuracy (mPA), and the floating-point operation FLOPs is much less than that of other algorithms.

[0050] Therefore, the multi-category target semantic segmentation method in the cluttered aliasing scene of the present invention can effectively improve the detection and segmentation efficiency of machine vision for multi-category disorderly placed targets, especially reduce the amount of algorithm calculation, and provide smart factories and intelligent robots with higher accuracy and more efficient segmentation results.

[0051] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for semantic segmentation of multi-category objects in a cluttered aliasing scene, characterized in that: include: Build a benchmark dataset; Construct a benchmark neural network model based on the U-Net network, wherein the benchmark neural network model includes a downsampling network and an upsampling network, wherein the downsampling network is a Resnet50 network incorporating a channel attention mechanism, and the upsampling network adopts a bilinear interpolation upsampling method incorporating a coordinate attention mechanism or a bilinear interpolation upsampling method incorporating Ghost convolution and a coordinate attention mechanism; Training and optimizing the benchmark neural network model using the benchmark data set to generate a target neural network model; The target image is input into the target neural network model for pixel-by-pixel semantic segmentation to output a segmentation result.

2. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene according to claim 1, characterized in that: The downsampling network includes a convolutional layer, a maximum pooling layer, a first ECA residual module, a second ECA residual module, a third ECA residual module and a fourth ECA residual module which are arranged in sequence; The first ECA residual module includes three ECA residual blocks, the second ECA residual module includes four ECA residual blocks, the third ECA residual module includes six ECA residual blocks, and the fourth ECA residual module includes three ECA residual blocks; The ECA residual block includes a 1*1 convolutional layer, a 3*3 convolutional layer, a 1*1 convolutional layer and an ECA channel attention mechanism arranged in sequence.

3. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene as claimed in claim 2, characterized in that: The steps of constructing the downsampling network include: Construct a Resnet50 network, and use the Resnet50 network as a downsampling network of the U-Net network; A channel attention mechanism is set in each residual block of the Resnet50 network.

4. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene according to claim 1, characterized in that: The upsampling network includes a coordinate attention mechanism, a first decoding module, a second decoding module, a third decoding module, a fourth decoding module and an output layer which are arranged in sequence; The first decoding module, the second decoding module and the third decoding module all include bilinear interpolation upsampling, two 3*3 convolutional layers and a coordinate attention mechanism arranged in sequence; The fourth decoding module includes bilinear interpolation upsampling, two 3*3 convolutional layers and bilinear interpolation upsampling arranged in sequence; The output layer includes two 3*3 convolutional layers and a 1*1 convolutional layer arranged in sequence.

5. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene as claimed in claim 4, characterized in that: The 3*3 convolutional layers in the first decoding module, the second decoding module, the third decoding module and the fourth decoding module are 3*3 Ghost convolutional layers.

6. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene as claimed in claim 4, characterized in that: The steps to construct an upsampling network using bilinear interpolation upsampling with coordinate attention mechanism include: Construct an upsampling network of the U-Net network, wherein the upsampling network includes four decoding modules, a deconvolution upsampling layer and an output layer arranged in sequence, and the decoding module includes a deconvolution upsampling layer and two 3*3 convolution layers arranged in sequence; Modify the deconvolution upsampling in the upsampling network of the U-Net network into bilinear interpolation upsampling; A coordinate attention mechanism is set before the bilinear interpolation upsampling of the decoding module.

7. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene as claimed in claim 5, characterized in that: The steps for constructing an upsampling network using bilinear interpolation upsampling with Ghost convolution and coordinate attention mechanism include: Construct an upsampling network of the U-Net network, wherein the upsampling network includes four decoding modules, a deconvolution upsampling layer and an output layer arranged in sequence, and the decoding module includes a deconvolution upsampling layer and two 3*3 convolution layers arranged in sequence; Modify the deconvolution upsampling in the upsampling network of the U-Net network into bilinear interpolation upsampling; Setting a coordinate attention mechanism before bilinear interpolation upsampling of the decoding module; The two 3*3 convolutions of the decoding module are modified into two 3*3 Ghost convolutions.

8. The method for semantic segmentation of multi-category objects in a cluttered aliasing scene as claimed in claim 1, characterized in that: The steps of constructing the benchmark dataset include: The Linemod_preprocessed dataset is used as the benchmark dataset; Mapping the pixel values ​​of each target object in the reference data set in turn, so as to map the pixel value of the Nth target object to N, where N is a positive integer less than 256; The benchmark data set after mapping processing is divided into a training data set and a test data set.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for semantic segmentation of multi-category objects in a cluttered aliasing scenario described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for semantic segmentation of multi-category objects in a cluttered aliasing scenario described in any one of claims 1 to 8 are implemented.

Images