Radar target semantic segmentation method and device based on metric adaptive peak convolution

Abstract

The invention provides a radar target semantic segmentation method based on metric adaptive peak convolution. The method comprises the steps that tensors of three view angles are acquired based on original radar echo signals; sequentially obtaining K groups of candidate reference unit sets corresponding to each to-be-tested unit on each view angle tensor; obtaining K groups of sampled reference unit sets corresponding to each to-be-tested unit; obtaining a measurement score set of the K groups of uniformly sampled reference unit sets relative to the to-be-measured unit; taking a group number index corresponding to a first-order gradient maximum value of the K groups of uniformly sampled reference unit sets relative to the measurement score set of the to-be-tested unit as an optimal group number index, and taking a union set of the uniformly sampled reference unit set corresponding to the optimal group number index and the current to-be-tested unit as an optimal peak receptive field; acquiring a peak convolution result correspondingly output by each to-be-tested unit under the condition of the optimal peak receptive field; and performing radar target semantic segmentation on the tensors of the three view angles.

Description

Technical Field

[0001] The present invention relates to the technical field of radar target semantic segmentation, and in particular to a radar target semantic segmentation method and device based on metric adaptive peak convolution. Background Art

[0002] As a sensor with long-range perception capabilities, radar demonstrates superior robustness compared to visible light sensors in complex environments (such as changing weather and lighting conditions). It is also more cost-effective and adaptable than lidar in extreme weather scenarios. Thanks to the physical advantages of radar and the development of modern deep learning, deep learning-based radar signal interpretation has become a hot research topic in radar perception technology, attracting widespread attention in areas such as autonomous driving, drone reconnaissance, and ocean monitoring.

[0003] Deep learning-based radar target detection methods have advanced radar signal interpretation from the traditional "target-clutter" binary assumption to modern machine learning approaches focused on target semantic information. These include radar object detection (ROD) and radar semantic segmentation (RSS). Due to the similarities in dense representations between radar spectrograms and optical images, mainstream deep learning-based radar signal interpretation research has directly transferred convolutional networks or modules developed for optical signals (such as standard convolution, multi-scale fused convolution, dilated convolution, and deformable convolution) to radar perception tasks, achieving impressive performance. However, due to a lack of specialized design targeting the inherent characteristics of radar signals, these methods have failed to fully tap the potential of convolutional networks for radar scene understanding.

[0004] The mid-frequency response of radar signals, composed of target echoes and interference, exhibits a unique peak morphology. Therefore, most classical radar detection methods construct peak detection algorithms based on the constant false alarm rate (CFAR) criterion. Leveraging the inherent peak response characteristics of radar target signals, the recently proposed Peak Convolution (PKC) continues the principles of guard band setting and clutter estimation in classical CFAR. It explicitly embeds a similar bandpass peak enhancement mechanism within the standard convolution operator. By designing a novel "peak receptive field" (PRF), it achieves learnable clutter estimation and target screening, cleverly combining the advantages of classical radar detectors with conventional convolutional networks. However, target characteristics and associated interference in radar signals vary significantly, and the fixed guard band setting limits PKC's adaptability to signal diversity. Summary of the Invention

[0005] The present invention provides a radar target semantic segmentation method and device based on metric adaptive spike convolution, which can solve the technical problem that the current spike convolution mechanism uses a predefined spike receptive field to collect reference signals for noise estimation, resulting in performance degradation when the target frequency domain response extension is inconsistent, the noise and clutter distribution is non-uniform, and the time-varying characteristics are present.

[0006] According to one aspect of the present invention, a radar target semantic segmentation method based on metric adaptive peak convolution is provided, the method comprising:

[0007] Based on the original radar echo signal, tensors of three perspectives, namely "range-Doppler", "range-azimuth" and "azimuth-Doppler" are obtained;

[0008] Sequentially obtain K sets of candidate reference units corresponding to each unit to be tested on each view tensor, thereby generating K sets of candidate spike receptive fields; wherein the K sets of candidate spike receptive fields include the current unit to be tested and the set of candidate reference units in the reference band area corresponding to the current unit to be tested;

[0009] Uniformly sampling N reference units from the K sets of candidate reference units corresponding to each unit under test to obtain K sets of sampled reference units corresponding to each unit under test, thereby generating K sets of uniformly sampled spike receptive fields; wherein the K sets of uniformly sampled spike receptive fields include the current unit under test and the uniformly sampled reference unit set within the reference band area corresponding to the current unit under test;

[0010] Obtain a set of metric scores of the K groups of uniformly sampled reference units relative to the unit to be tested based on the K groups of uniformly sampled reference units corresponding to each unit to be tested;

[0011] The group number index corresponding to the maximum value of the first-order gradient of the K groups of uniformly sampled reference unit sets relative to the measurement score set of the unit to be tested is taken as the optimal group number index, and the union of the uniformly sampled reference unit set corresponding to the optimal group number index and the current unit to be tested is taken as the optimal peak receptive field;

[0012] Obtain the spike convolution result corresponding to the output of each unit under test under the optimal spike receptive field;

[0013] Based on the peak convolution results of each unit under test on each perspective tensor under the optimal peak receptive field, radar target semantic segmentation is performed on the tensors of the three perspectives of "range-Doppler", "range-azimuth" and "azimuth-Doppler".

[0014] Preferably, the K groups of candidate peak receptive fields are obtained by the following formula:

[0015]

[0016] in,

[0017] Where, represents the K groups of candidate spike receptive fields, x c Indicates the current unit under test. Indicates the sth candidate reference unit in the reference band area corresponding to the current unit to be tested, N r represents the total number of K groups of candidate reference units, Respectively represent the horizontal and vertical widths of the protection band, Respectively represent the horizontal and vertical widths of the reference band, Represents x c The horizontal and vertical coordinates in the feature map, Respectively The horizontal and vertical coordinates in the feature map, b G represents the protection bandwidth, b R represents the reference bandwidth, p c Represents x c The coordinates in the feature map, express Coordinates in the feature map.

[0018] Preferably, the number of candidate spike receptive field groups generated by each unit to be tested is obtained by the following formula:

[0019]

[0020] Where K represents the number of candidate spike receptive field groups generated by each unit under test, Respectively represent the maximum and minimum lateral widths of the protection band, They represent the maximum and minimum longitudinal widths of the guard band, respectively.

[0021] Preferably, the peak receptive field of each group after uniform sampling is obtained by the following formula:

[0022]

[0023] Where, represents the peak receptive field after uniform sampling of the kth group, represents the kth group of i-th uniformly sampled reference units in the reference band area corresponding to the current unit to be tested, and N represents the total number of uniformly sampled reference units in each group.

[0024] Preferably, the metric score of each uniformly sampled reference unit set relative to the unit under test is obtained by the following formula:

[0025]

[0026] Where, ξk represents the metric score of the kth group of uniformly sampled reference units relative to the unit to be tested, σ(·) represents the Sigmoid activation function, and C represents the number of feature map channels.

[0027] Preferably, the optimal peak receptive field is obtained by the following formula:

[0028]

[0029] Where, represents the optimal peak receptive field, Represents the optimal number of groups index The corresponding uniformly sampled reference unit set, argmax represents the index corresponding to the maximum value retrieved from the gradient g[sort(Ξ)], g represents the difference function, sort represents the descending sorting operator, and Ξ represents the set of metric scores of the K groups of uniformly sampled reference units relative to the unit to be tested.

[0030] Preferably, the peak convolution result corresponding to the output of each unit under test under the optimal peak receptive field is obtained by the following formula:

[0031]

[0032] Where, Indicates the peak convolution result of the current unit under test under the optimal peak receptive field, W represents The learning weights of , Vec(·) represents the vectorized operator, represents the learning weight corresponding to the j-th output channel, C in 、C out Respectively represent the number of input channels and the number of output channels, Represents the field of real numbers.

[0033] According to another aspect of the present invention, a computer device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the computer program.

[0034] By applying the technical solution of the present invention, adaptive spike convolution can optimize the bandpass filtering mechanism at the unit level by dynamically adjusting the reference unit set of each unit to be tested. This allows the radar target semantic segmentation (or radar target detection and recognition) method, model, system or device based on this convolution to exhibit greater flexibility and robustness when processing time-varying non-stationary radar signals, thereby improving the adaptability of PKC to environmental clutter and contributing to the improvement of radar target signal processing performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The accompanying drawings are included to provide a further understanding of the embodiments of the present invention, constitute a part of the specification, illustrate the embodiments of the present invention, and together with the description, explain the principles of the present invention. Obviously, the drawings described below are only some embodiments of the present invention, and those skilled in the art can derive other drawings based on these drawings without inventive effort.

[0036] Figure 1 A flowchart of a radar target semantic segmentation method based on metric-adaptive spike convolution according to an embodiment of the present invention is shown;

[0037] Figure 2 A framework diagram of an adaptive spike receptive field in a metric-based adaptive spike convolution according to an embodiment of the present invention is shown;

[0038] Figure 3 A framework diagram of a radar target semantic segmentation network based on metric adaptive spike convolution provided according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments in this application and the features in the embodiments can be combined with each other. The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is in no way intended to limit the present invention and its application or use. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0040] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0041] Unless otherwise specifically stated, the relative arrangement of the parts and steps, the numerical expressions and the numerical values ​​set forth in these embodiments do not limit the scope of the present invention. At the same time, it should be understood that, for ease of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship. The techniques, methods and equipment known to those of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods and equipment should be considered as part of the authorization specification. In all examples shown and discussed here, any specific values ​​should be interpreted as being merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values. It should be noted that similar numbers and letters represent similar items in the following figures, and therefore, once an item is defined in one figure, it does not need to be further discussed in subsequent figures.

[0042] The present invention generates multiple candidate spike receptive fields, constructs a metric to evaluate the correlation between each candidate reference unit set and the target interference noise, and selects the optimal reference unit set for noise suppression and target enhancement, thereby solving the problem of insufficient adaptability of fixed spike receptive fields in time-varying non-stationary radar signals. This is achieved specifically by the following method: performing a fast Fourier transform on the original radar echo signal to obtain tensors of three perspectives: "Range-Doppler", "Range-Angle", and "Angle-Doppler" as the input to be tested. Adaptive spike convolution operations are then performed on the units to be tested on each single perspective tensor. First, for each unit to be tested in the radar signal feature map, multiple candidate reference unit sets are generated based on the preset protection bandwidth adjustment range (minimum / maximum values ​​in the horizontal and vertical directions), and the number of reference units in each reference unit set is uniformly sampled; then, for each candidate reference unit set, the correlation between its reference unit and the unit to be tested in the feature space is calculated, and the average of the correlation scores of all reference units is taken to obtain the overall metric score of the reference unit set; after obtaining the metric score, it is sorted in descending order, and its gradient sequence is calculated, and the peak receptive field corresponding to the maximum gradient is selected as the final adaptive peak receptive field; then, based on the selected optimal peak receptive field, a PKC-like operation is performed, and each unit in the receptive field is subtracted from the copied unit to be tested to obtain the unit to be tested features with different noise suppression levels, which are then input into the two-dimensional convolution to further extract high-level semantic features. Finally, the proposed adaptive peak convolution is applied to the proposed network AdaPKC. ξ -Net is used to achieve radar semantic segmentation. The present invention retains the bandpass filtering characteristics of PKC and improves the adaptability of the model to complex signals by selecting a dynamic peak receptive field based on metrics.

[0043] In order to have a further understanding of the present invention, the following Figure 1-Figure 3 The radar target semantic segmentation method based on metric adaptive spike convolution of the present invention is described in detail.

[0044] In this embodiment, it is assumed that the feature point (i.e., the unit to be tested) on a certain layer of the feature map in the radar semantic segmentation network is Where C is the number of feature map channels, The field of real numbers.

[0045] like Figure 1-Figure 3 As shown, the present invention provides a radar target semantic segmentation method based on metric adaptive peak convolution, the method comprising:

[0046] The first step is to obtain tensors of three perspectives: "range-Doppler", "range-azimuth", and "azimuth-Doppler" based on the original radar echo signal. Among them, "range-Doppler", "range-azimuth", and "azimuth-Doppler" are radar signal feature maps containing "range-Doppler", "range-azimuth", and "azimuth-Doppler" data respectively; each radar signal feature map has multiple units to be measured;

[0047] Step 2: Obtain K sets of candidate reference units corresponding to each unit under test on each viewing angle tensor in turn, thereby generating K sets of candidate peak receptive fields; wherein the K sets of candidate peak receptive fields include the current unit under test and the set of candidate reference units in the reference band area corresponding to the current unit under test;

[0048] The K groups of candidate spike receptive fields are obtained by the following formula:

[0049]

[0050] in,

[0051] Where, represents the K groups of candidate spike receptive fields, x c Indicates the current unit under test. Indicates the sth candidate reference unit in the reference band area corresponding to the current unit to be tested, N r represents the total number of K groups of candidate reference units, Respectively represent the horizontal and vertical widths of the protection band, Respectively represent the horizontal and vertical widths of the reference band, Represents x c The horizontal and vertical coordinates in the feature map, Respectively The horizontal and vertical coordinates in the feature map, b G represents the protection bandwidth, b R represents the reference bandwidth, pc Represents x c The coordinates in the feature map, express Coordinates in the feature map.

[0052] Setting the reference bandwidth Fixed, and the protection bandwidth candidate set is defined as a set of K elements:

[0053]

[0054] in,

[0055] Correspondingly, there are K groups of reference units:

[0056] Where K represents the number of candidate spike receptive field groups generated by each unit under test, Respectively represent the maximum and minimum lateral widths of the protection band, Respectively represent the maximum and minimum longitudinal widths of the protection band, Indicates the lth candidate reference unit in the kth group within the reference band area corresponding to the current unit to be tested, N k Represents the total number of candidate reference units in each group, N k ≤N r .

[0057] Step 3: uniformly sample N reference units from the K sets of candidate reference units corresponding to each unit under test, obtaining K sets of sampled reference units corresponding to each unit under test, thereby generating K sets of uniformly sampled peak receptive fields; wherein the K sets of uniformly sampled peak receptive fields include the current unit under test and the uniformly sampled reference unit set in the reference band area corresponding to the current unit under test;

[0058] The peak receptive field of each group after uniform sampling is obtained by the following formula:

[0059]

[0060] Where, represents the peak receptive field after uniform sampling of the kth group, It represents the reference unit after uniform sampling of the kth group i in the reference band area corresponding to the current unit to be tested. N represents the total number of reference units after uniform sampling of each group, N≤N k , generally N=16, Indicates that the current unit under test corresponds to the kth set of uniformly sampled reference units.

[0061] Step 4: Based on the K sets of uniformly sampled reference units corresponding to each unit under test, a set of metric scores (i.e., correlation values) of the K sets of uniformly sampled reference units relative to the unit under test is obtained;

[0062] The metric score of each set of uniformly sampled reference units relative to the unit under test is obtained by the following formula:

[0063]

[0064] Where, ξ k represents the metric score of the kth group of uniformly sampled reference units relative to the unit to be tested, σ(·) represents the Sigmoid activation function, and ξ k Normalized to the (0,1) interval, C represents the number of feature map channels.

[0065] Step 5: The group number index corresponding to the maximum value of the first-order gradient of the K groups of uniformly sampled reference unit sets relative to the measurement score set of the unit to be tested is used as the optimal group number index, and the union of the uniformly sampled reference unit set corresponding to the optimal group number index and the current unit to be tested is used as the optimal peak receptive field;

[0066] The optimal peak receptive field is obtained by the following formula:

[0067]

[0068] Where, represents the optimal peak receptive field, Represents the optimal number of groups index The corresponding uniformly sampled reference unit set, argmax represents the index corresponding to the maximum value retrieved from the gradient g[sort(Ξ)], g represents the difference function, sort represents the descending sorting operator, and Ξ represents the set of metric scores of the K groups of uniformly sampled reference units relative to the unit to be tested.

[0069] Step 6: Obtain the peak convolution result corresponding to the output of each unit under test under the optimal peak receptive field;

[0070] The following formula is used to obtain the corresponding output spike convolution result of each unit under test under the optimal spike receptive field:

[0071]

[0072] Where, Indicates the peak convolution result of the current unit under test under the optimal peak receptive field, W represents The learning weights of , Vec(·) represents the vectorized operator, represents the learning weight corresponding to the j-th output channel, C in、C out Respectively represent the number of input channels and the number of output channels, Represents the field of real numbers.

[0073] Step 7: Based on the peak convolution results corresponding to the output of each unit under test in each view tensor under the optimal peak receptive field, radar target semantic segmentation is performed on the "range-Doppler", "range-azimuth", and "azimuth-Doppler" view tensors. The specific semantic segmentation method can adopt existing methods and will not be detailed in this invention.

[0074] The present invention also provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the above methods when executing the computer program.

[0075] The adaptive spike convolution operator designed in this patent is an improvement on the spike convolution operator. This operator innovatively introduces the idea of ​​adaptive reference unit selection in the classic CFAR detector into the convolution operation, and dynamically adjusts the spike receptive field of each unit in a data-driven manner, so that it can adapt to challenges such as target frequency response expansion, non-uniform distribution of interference noise, and time variability. Its core lies in constructing a multi-scale candidate reference unit set and realizing dynamic screening at the unit level based on feature space correlation measurement. Compared with the convolution operator with a fixed spike receptive field, AdaPKC ξ The model's generalization ability for multi-scenario radar signals has been significantly improved, breaking through the bottleneck of traditional fixed receptive field convolution's insufficient adaptability to radar signal diversity.

[0076] like Figure 3 As shown in the figure, for the multi-view radar semantic segmentation (RSS) task of radar dataset, the ReDA-PKC module is replaced by AdaPKC based on the design of PKCIn-Net. ξ , a multi-view RSS model was constructed and named AdaPKC ξ -Net, its overall framework is as follows Figure 3 AdaPKC ξ The rest of the components of -Net are consistent with PKCIn-Net: i) The designed network takes multi-view radar frequency domain tensors of range-Doppler (RD), range-azimuth (RA) and azimuth-Doppler (AD) as input, consists of three encoding branches and two decoding branches, and performs semantic segmentation on RD and RA views simultaneously; ii) The basic encoding module of the encoding branch of each single view contains two 3D convolution blocks to fully mine the temporal information and obtain higher-level feature representations; iii) The basic encoding module in the encoding branch is followed by AdaPKC ξmodule, and adopts the Atrous Spatial Pyramid Module (ASPP) module to aggregate multi-scale spatial information; iv) After the encoding stage, the multi-view features are fused through the Latent Space Encoder (LSE) and passed to the decoder to generate the final prediction result.

[0077] In summary, the present invention provides a radar target semantic segmentation method and device based on metric adaptive spike convolution. By dynamically selecting the optimal spike receptive field through a data-driven adaptive mechanism, the present invention solves the performance degradation of the fixed receptive field convolution mechanism when the target frequency domain response is inconsistently extended and the noise and clutter distribution is non-uniform and time-varying. Compared with existing radar semantic segmentation methods based on deep learning, the present invention has a significant performance improvement.

[0078] For ease of description, spatially relative terms such as "above", "above", "on the upper surface of", "above", etc. may be used herein to describe the spatial positional relationship of a device or feature to other devices or features as shown in the figures. It should be understood that spatially relative terms are intended to include different orientations of the device in use or operation in addition to the orientation described in the figures. For example, if the device in the drawings is inverted, the device described as "above other devices or structures" or "above other devices or structures" will be positioned as "below other devices or structures" or "below other devices or structures". Thus, the exemplary term "above" can include both "above" and "below". The device can also be positioned in other different ways (rotated 90 degrees or in other orientations), and the spatially relative descriptions used here are interpreted accordingly.

[0079] In addition, it should be noted that the use of terms such as "first" and "second" to limit components is only for the convenience of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of the present invention.

[0080] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A radar target semantic segmentation method based on metric adaptive peak convolution, characterized in that: The method comprises: Based on the original radar echo signal, tensors of three perspectives, namely "range-Doppler", "range-azimuth" and "azimuth-Doppler" are obtained; Sequentially obtain K sets of candidate reference units corresponding to each unit to be tested on each view tensor, thereby generating K sets of candidate spike receptive fields; wherein the K sets of candidate spike receptive fields include the current unit to be tested and the set of candidate reference units in the reference band area corresponding to the current unit to be tested; Uniformly sampling N reference units from the K sets of candidate reference units corresponding to each unit under test to obtain K sets of sampled reference units corresponding to each unit under test, thereby generating K sets of uniformly sampled spike receptive fields; wherein the K sets of uniformly sampled spike receptive fields include the current unit under test and the uniformly sampled reference unit set within the reference band area corresponding to the current unit under test; Obtain a set of metric scores of the K groups of uniformly sampled reference units relative to the unit to be tested based on the K groups of uniformly sampled reference units corresponding to each unit to be tested; The group number index corresponding to the maximum value of the first-order gradient of the K groups of uniformly sampled reference unit sets relative to the measurement score set of the unit to be tested is taken as the optimal group number index, and the union of the uniformly sampled reference unit set corresponding to the optimal group number index and the current unit to be tested is taken as the optimal peak receptive field; Obtain the spike convolution result corresponding to the output of each unit under test under the optimal spike receptive field; Based on the peak convolution results of each unit under test on each view tensor under the optimal peak receptive field, radar target semantic segmentation is performed on the tensors of the three view angles of "range-Doppler", "range-azimuth" and "azimuth-Doppler".

2. The method according to claim 1, characterized in that The K groups of candidate spike receptive fields are obtained by the following formula: in, Where, represents the K groups of candidate spike receptive fields, x c Indicates the current unit under test. Indicates the sth candidate reference unit in the reference band area corresponding to the current unit to be tested, N r represents the total number of K groups of candidate reference units, Respectively represent the horizontal and vertical widths of the protection band, Respectively represent the horizontal and vertical widths of the reference band, Represents x c The horizontal and vertical coordinates in the feature map, Respectively The horizontal and vertical coordinates in the feature map, b G represents the protection bandwidth, b R represents the reference bandwidth, p c Represents x c The coordinates in the feature map, express Coordinates in the feature map.

3. The method according to claim 1, characterized in that The number of candidate spike receptive field groups generated by each unit to be tested is obtained by the following formula: Where K represents the number of candidate spike receptive field groups generated by each unit under test, Respectively represent the maximum and minimum lateral widths of the protection band, They represent the maximum and minimum longitudinal widths of the guard band, respectively.

4. The method according to claim 1, wherein The peak receptive field of each group after uniform sampling is obtained by the following formula: Where, represents the peak receptive field after uniform sampling of the kth group, represents the kth group of i-th uniformly sampled reference units in the reference band area corresponding to the current unit to be tested, and N represents the total number of uniformly sampled reference units in each group.

5. The method according to claim 1, wherein The metric score of each set of uniformly sampled reference units relative to the unit under test is obtained by the following formula: Where, ξ k represents the metric score of the kth group of uniformly sampled reference units relative to the unit to be tested, σ(·) represents the Sigmoid activation function, and C represents the number of feature map channels.

6. The method according to claim 1, characterized in that The optimal peak receptive field is obtained by the following formula: Where, represents the optimal peak receptive field, Represents the optimal number of groups index The corresponding uniformly sampled reference unit set, argmax represents the index corresponding to the maximum value retrieved from the gradient g[sort(Ξ)], g represents the difference function, sort represents the descending sorting operator, and Ξ represents the set of metric scores of the K groups of uniformly sampled reference units relative to the unit to be tested.

7. The method according to claim 1, characterized in that The following formula is used to obtain the corresponding output spike convolution result of each unit under test under the optimal spike receptive field: Where, Indicates the peak convolution result of the current unit under test under the optimal peak receptive field, W represents The learning weights of , Vec(·) represents the vectorized operator, represents the learning weight corresponding to the j-th output channel, C in 、C out Respectively represent the number of input channels and the number of output channels, Represents the field of real numbers.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

Images