Short-wave time-frequency map target detection method and system based on multi-tile fusion, and medium

Abstract

The application discloses a short-wave time-frequency map target detection method and system based on multi-graph block fusion, and a medium, and belongs to the technical field of short-wave communication, and comprises the following steps: time-frequency map construction; time-frequency map division and construction of a multi-channel joint detection input containing a center graph block and at least one adjacent graph block; performing feature extraction and context fusion processing to obtain a fused feature map; performing target detection processing on the fused feature map and outputting a candidate detection result of each graph block; performing cross-graph weighted non-maximum suppression processing, fusing overlapping candidate boxes across graph blocks, and outputting a candidate box set after weighted fusion processing; and mapping the candidate box set after weighted fusion processing to an original time-frequency map coordinate system to obtain a final detection result of a target. The method in the application can effectively avoid the loss of context information caused by dividing and processing a time-frequency map into multiple graph blocks, ensure the integrity of a target in an edge region of each graph block, and avoid the missed detection of a weak signal and the repeated detection across graph blocks.

Description

Technical Field

[0001] This invention belongs to the field of shortwave communication technology and relates to a method, system and medium for shortwave time-frequency map target detection based on multi-map tile fusion, and particularly to a method, system and computer storage medium for shortwave time-frequency map target detection based on multi-map tile context fusion. Background Technology

[0002] Shortwave communication is characterized by its long transmission distance and independence from terrestrial networks, making it widely used in fields such as aviation navigation, maritime communication, and emergency response. However, shortwave signals are susceptible to multipath fading, frequency shift, and interference noise during propagation in the ionosphere. Furthermore, the electromagnetic environment in reality is complex and dynamic, with diverse signal types (including irregular signal forms such as frequency hopping signals and burst signals), thus placing high demands on signal detection.

[0003] Currently, conventional shortwave signal detection methods generally include those based on energy detection, modulation feature analysis, or cyclostationarity characteristics. Among these, energy detection-based methods are prone to missing targets in low signal-to-noise ratio environments, resulting in insufficient detection comprehensiveness and accuracy; modulation feature-based methods rely on modulation characteristics, have insufficient adaptability to irregular signals, and have a narrow application range, limiting their use; while cyclostationarity-based detection has good robustness, its high computational complexity makes it unsuitable for real-time processing, resulting in poor timeliness.

[0004] Because conventional methods for detecting shortwave signals have the aforementioned drawbacks, researchers conceived of converting shortwave signals into time-frequency maps and using convolutional neural networks for detection. While this approach can improve shortwave signal recognition to some extent, it requires dividing the shortwave time-frequency map into multiple patches for processing. This results in a loss of contextual information across patches, leading to problems such as incomplete targets in edge regions, missed detection of weak signals, or repeated detection across patches. Consequently, it cannot meet the requirements for shortwave signal communication detection in complex electromagnetic environments and has significant limitations. Summary of the Invention

[0005] In response to one or more of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a shortwave time-frequency map target detection method, system and medium based on multi-map fusion, aiming to overcome the shortcomings of the existing shortwave signal image detection technology, improve the problems of missed detection, repeated detection and inconsistent detection results in cross-map boundary areas, and realize stable target detection and result fusion across map regions.

[0006] To achieve the above objectives, one aspect of the present invention provides a shortwave time-frequency map target detection method based on multi-map patch fusion, comprising the following steps: S1: Preprocess the acquired shortwave signal to construct a time-frequency diagram of the shortwave signal; S2: Divide the time-frequency graph into patches and select at least one adjacent patch for each central patch; stack the central patch and adjacent patches in the channel dimension to construct a multi-channel joint detection input; S3: Perform feature extraction based on multi-tile joint input on the multi-channel joint detection input, and perform fusion processing in combination with the context information of adjacent tiles to enhance the feature continuity of the edge region of the central tile, reduce the inconsistency of target information at the cross-tile boundary, and obtain the fused feature map; S4: Perform object detection processing on the fused feature map and output the candidate detection results for each patch; S5: Perform cross-map weighted nonmaximum suppression processing on the candidate detection results for each map tile, fuse overlapping candidate boxes across map tiles, and output the set of candidate boxes after weighted fusion processing; S6: Map the candidate box set to the original time-frequency map coordinate system to obtain the final detection result of the target.

[0007] As a further improvement of the present invention, in step S1, the process of constructing the time-frequency diagram specifically includes: Signal sampling over a wide bandwidth is performed using a complex IQ sampling method; A two-dimensional time-frequency graph is constructed using short-time Fourier transform. The construction process includes: configuring the window function, window length, and frame shift parameters; performing spectrum calculation and power spectrum processing; normalizing the power spectrum; and organizing the normalized two-dimensional matrix with time as the horizontal axis and frequency as the vertical axis to form a two-dimensional grayscale time-frequency graph.

[0008] As a further improvement of the present invention, in step S2, the at least one adjacent block includes: a left adjacent block that is adjacent to the center block in the time axis direction; and / or an upper adjacent block that is adjacent to the center block in the frequency axis direction.

[0009] As a further improvement of the present invention, for a central block that does not have a left adjacent block and / or an upper adjacent block, the adjacent area is filled by zero filling or mirror expansion. and / or When dividing a map into blocks, an overlapping area is set between adjacent blocks, and the overlap width is 5% to 15% of the block's adjacent direction dimension.

[0010] As a further improvement of the present invention, in step S3, the process of constructing the fused feature map specifically includes: A convolutional neural network is used to extract features from the central patch and its adjacent patches, and position encoding is introduced during the feature extraction process; wherein, the position encoding is used to represent the relative positional relationship of each patch in the time-frequency graph; The extracted multi-source features are applied to a spatial-channel attention mechanism, and the features are weighted according to the spatial positional relationship between the central patch and its neighboring patches to generate a fused feature map in a weighted fusion manner.

[0011] As a further improvement of the present invention, the candidate detection results obtained in step S4 further include: Non-maximum suppression is applied to multiple overlapping candidate detection boxes appearing in the same region; and / or, Candidate detection boxes are filtered based on a preset confidence threshold.

[0012] As a further improvement of the present invention, in step S5, the weighted fusion process includes: Candidate bounding boxes are filtered based on the candidate detection results for each map patch, and a set of candidate bounding boxes to be fused is selected. For a set of candidate boxes whose overlap exceeds a preset threshold, weights are calculated based on the confidence level of the candidate boxes and the location of their source tiles, and the position or category information of the candidate boxes is fused based on the weights. Perform cross-map redundancy processing to merge multiple candidate boxes from different map regions into a smaller number of merged boxes.

[0013] As a further improvement of the present invention, the method further includes a model training optimization step: S7: During the model training phase, construct cross-... Figure 1 Consistency loss function constrains the object detection network to output consistent predictions in overlapping areas of different tiles; and / or, uses high-confidence predictions from adjacent tiles to generate pseudo-labels for self-supervised training of the object detection network with insufficient training data.

[0014] Another aspect of the present invention provides a shortwave time-frequency map target detection system based on multi-map patch fusion, comprising: The time-frequency diagram construction module is used to construct a time-frequency diagram based on the acquired shortwave signal; The tile generation module is used to divide the time-frequency map into multiple tiles and construct a multi-channel joint detection input that includes a central tile and at least one adjacent tile of the central tile. The context fusion module is used to perform feature extraction and context information fusion on multi-channel joint detection input, and output a fused feature map; The object detection module is used to perform object detection processing on the fused feature map and output the candidate detection results for each patch; The cross-tile fusion module is used to perform weighted nonmaximum suppression processing on the candidate detection results of each tile, fuse overlapping candidate boxes across tiles, and output a set of candidate boxes after weighted fusion processing; The coordinate mapping module is used to map the set of candidate boxes after weighted fusion processing to the original time-frequency map coordinate system and output the final detection result of the target.

[0015] In another aspect, the present invention also provides a storage medium storing a processor-executable program, which, when executed by a processor, is used to perform the aforementioned shortwave time-frequency map target detection method based on multi-map fusion.

[0016] The aforementioned improved technical features can be combined with each other as long as they do not conflict with each other.

[0017] In summary, the beneficial effects of the above-described technical solutions conceived by this invention compared with the prior art include: The present invention provides a shortwave time-frequency map target detection method based on multi-pattern fusion, comprising the following steps: time-frequency map construction; dividing the time-frequency map and constructing a multi-channel joint detection input including a central patch and at least one adjacent patch; performing feature extraction and context fusion processing to obtain a fused feature map; performing target detection processing on the fused feature map and outputting candidate detection results for each patch; performing cross-pattern weighted nonmaximum suppression processing, fusing overlapping candidate boxes across patches, and outputting a set of candidate boxes after weighted fusion processing; mapping the set of candidate boxes after weighted fusion processing to the original time-frequency map coordinate system to obtain the final target detection result. The aforementioned method effectively avoids the loss of contextual information caused by processing the time-frequency map by dividing it into multiple patches, ensuring the integrity of targets in the edge regions of each patch, avoiding missed detection of weak signals and repeated detection across patches, and improving the quality and efficiency of shortwave signal detection. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of the shortwave time-frequency map target detection method based on multi-map fusion in Embodiment 1 of the present invention; Figure 2 This is a flowchart of the process of constructing the fused feature map in step S3 of embodiment 1 of the present invention; Figure 3 This is a detailed flowchart of obtaining candidate detection results in step S4 of embodiment 1 of the present invention; Figure 4 This is a detailed flowchart of the weighted fusion process in step S5 of embodiment 1 of the present invention; Figure 5This is a flowchart of the weighted fusion result coordinate mapping and update in step S6 of embodiment 1 of the present invention; Figure 6 This is an architecture diagram of the shortwave time-frequency map target detection system based on multi-map fusion in Embodiment 2 of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0021] In the description of this invention, it should be understood that, unless otherwise expressly specified and limited, the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," "circumferential," etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing this invention and simplifying the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0022] Furthermore, unless otherwise expressly defined, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise expressly and specifically defined.

[0023] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise explicitly limited. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0024] In this invention, unless otherwise explicitly specified and limited, "above" or "below" the second feature can mean that the first feature is in direct contact with the second feature, or that the first feature is in indirect contact with the second feature through an intermediate medium. Furthermore, "above," "over," and "on top" of the second feature can mean that the first feature is directly above or diagonally above the second feature, or simply that the first feature is at a higher horizontal level than the second feature. "Below," "below," and "under" the second feature can mean that the first feature is directly below or diagonally below the second feature, or simply that the first feature is at a lower horizontal level than the second feature.

[0025] Below, for reference Figures 1-6 This paper describes a shortwave time-frequency map target detection method, system, and medium based on multi-map fusion according to a preferred embodiment of the present invention, and elaborates on them in detail through the following three embodiments.

[0026] Example 1: As one aspect of the present invention, a preferred embodiment proposes a shortwave time-frequency map target detection method based on multi-map patch fusion, the steps of which are as follows: Figure 1 As shown, and specifically includes the following steps: S1: Preprocess the acquired shortwave signal to construct a time-frequency diagram of the shortwave signal; Specifically, the acquisition of shortwave signals is achieved through a shortwave receiving device, which is further preferably designed to sample electromagnetic signals covering the frequency range of 2MHz to 30MHz.

[0027] Meanwhile, the shortwave receiving device in the preferred embodiment preferably supports broadband receiving capability and adopts complex IQ sampling mode to simultaneously retain the in-phase component (I) and quadrature component (Q) at the same time point, which facilitates subsequent spectrum analysis and feature extraction.

[0028] In actual sampling, the sampling rate of the shortwave receiver can be set according to the equipment configuration and communication bandwidth requirements, which will not be elaborated here.

[0029] More specifically, for the IQ signal acquired in the preferred embodiment, it is further preferred to construct a two-dimensional time-frequency diagram using short-time Fourier transform (STFT), and the specific construction process is as follows: (1) Configure the window function and window length; In practical configuration, the preferred window function type is Kaiser window, Blackman window, or Hamming window, used to adjust spectral leakage characteristics. Simultaneously, the window length... (i.e., the number of sampling points in each Fast Fourier Transform) is preferably based on the sampling rate. With the required frequency resolution To define the relationships between them, that is... .

[0030] It should be noted that the preferred embodiments do not limit the specific values ​​of the aforementioned parameters, and the values ​​of each parameter can be flexibly adjusted according to detection requirements and computing resources. For example, in one specific implementation, the frequency resolution... It is 20Hz.

[0031] (2) Configure frame shift parameters; In actual configuration, the frame shift length is preferably set to 10% to 25% of the window length to balance time resolution and computational complexity.

[0032] More preferably, in one specific embodiment, the frame shift is preferably 12.5% ​​of the window length.

[0033] (3) Perform spectrum calculation; Perform a Fast Fourier Transform (FFT) on each frame of signal to obtain the corresponding complex spectrum data.

[0034] (4) Perform power spectrum processing; The logarithm of the amplitude of each frame's spectrum is taken to compress the dynamic range, making the energy at different amplitude levels visible and distinguishable in the image.

[0035] For example, in one specific implementation, the power spectrum is processed using... .

[0036] (5) Normalize the power spectrum; Specifically, the calculated power spectrum is subjected to minimum-maximum normalization to uniformly map the pixel value range to [0,1], and the normalization is preferably performed in the following manner:

[0037] in, These are the original pixel values; , These are the minimum and maximum values ​​of all pixels in the image, respectively. These are the normalized pixel values.

[0038] (6) Construct the time-frequency diagram; The normalized two-dimensional matrix is ​​organized with time as the horizontal axis and frequency as the vertical axis to form a two-dimensional grayscale time-frequency diagram, where the pixel grayscale value represents the relative intensity of the signal power.

[0039] The constructed time-frequency graph has the following characteristics: (1) Frequency axis resolution: The resolution of the frequency axis is given by the formula It can be determined and flexibly adjusted according to the sampling rate and window length.

[0040] (2) Time axis length: It is continuously generated by the sliding window of the STFT, and the time span is not fixed. It can be determined according to the acquisition duration and is used to preserve the time-varying characteristics of the signal.

[0041] (3) Image format: The output is a two-dimensional grayscale image, which facilitates subsequent image feature extraction and target detection processing.

[0042] S2: Divide the time-frequency graph into patches and select at least one adjacent patch for each central patch; stack the central patch and adjacent patches in the channel dimension to construct a multi-channel joint detection input; Specifically, in order to adapt to the input structure of the feature extraction network and to facilitate the use of contextual information provided by neighboring regions, the complete time-frequency map generated in step S1 is divided into patches and an adjacency structure is constructed.

[0043] More specifically, the time-frequency graph fully preserves the target frequency band range along the frequency axis, and continuously generates frames along the time axis using a sliding window based on short-time Fourier transform, forming a two-dimensional image that naturally extends along the time axis. Based on this, the image is further cropped along the time axis using a sliding window method to construct several uniformly sized image patches.

[0044] For example, in one specific implementation, the size of each tile is preferably set to a fixed M×M pixels, such as 640×640 pixels, where the width corresponds to the time axis direction and the length corresponds to the frequency axis direction.

[0045] Meanwhile, when segmenting maps, it is preferable to set overlapping areas between adjacent maps to preserve the time-frequency structure information across maps. In practice, the map overlap width is preferably a certain percentage of the size of the adjacent map dimensions (time axis or frequency axis), for example, 5% to 15%.

[0046] Furthermore, each currently detected patch is designated as the center patch, and patches adjacent to it in the time axis and / or frequency axis direction are extracted as adjacent patches to construct a joint input structure.

[0047] It is easy to understand that the "center" in the center block of the preferred embodiment is relative to the adjacent blocks in its dual-axis direction (time axis and frequency axis), and does not refer to the center position of the time-frequency graph.

[0048] In a specific implementation, the adjacent blocks of the central block preferably include the left adjacent block and the upper adjacent block, which together constitute a three-block joint detection unit. The left adjacent block refers to the block preceding the central block in the time axis direction, and the upper adjacent block refers to the block directly adjacent to the central block in the frequency axis direction.

[0049] In addition, if the central patch is located at the edge (i.e. there are no left adjacent patches and / or upper adjacent patches), it is preferable to use zero-fill or mirror expansion to fill the adjacent area, so as to keep the joint detection structure consistent.

[0050] Furthermore, in the preferred embodiment, the segmented image patches are two-dimensional grayscale images. Preferably, a feature extraction network is used to process the generated image patches, and the central image patch and its adjacent image patches are stacked in the channel dimension to form a multi-channel input image. This structure is used to provide contextual information of the current region and its neighborhood to the subsequent feature extraction and detection modules, so as to realize joint detection and contextual feature fusion.

[0051] In actual settings, the tile division in the preferred embodiment is not limited to tile size, sliding step size, number or direction of adjacent tiles, and also supports further expansion of multi-scale tile input, pyramid fusion structure and attention mechanism.

[0052] More preferably, in the preferred embodiment, a confidence map or spatial uncertainty map is also constructed based on the detection results to reflect the prediction differences in the time-frequency map region, and the adjacent map tile selection strategy is adjusted accordingly to make the adjacent map tile construction have certain adaptive characteristics.

[0053] S3: Perform feature extraction and context fusion processing on the multi-channel joint detection input to obtain a fused feature map; Specifically, based on the multi-channel joint detection input constructed in step S2, the input network model is further subjected to feature extraction based on multi-tile joint input, and fusion processing is performed in combination with the context information of adjacent tiles to enhance the feature continuity of the edge region of the central tile, reduce the inconsistency of target information at the boundary of cross tiles, and finally obtain the fused feature map.

[0054] The process further includes, for example Figure 2 As shown, the preferred method includes the following sub-steps: S31: Perform patch feature extraction and location information encoding; In a preferred embodiment, a convolutional neural network (CNN) is preferably used to extract features from the central patch and its adjacent patches.

[0055] As an example, the aforementioned extraction process preferably includes multi-layer convolution and pooling structures to obtain feature representations at different spatial scales.

[0056] Meanwhile, during the feature extraction process, a positional encoding structure is preferably introduced to embed the relative position of the patch in the original time-frequency map (e.g., the positional relationship between the center patch, the left neighbor patch, and the upper neighbor patch) to help the network distinguish feature information from different source regions.

[0057] S32: Perform spatial-channel attention fusion processing, and weight the features according to the spatial positional relationship between the central patch and adjacent patches to obtain a feature map with fused contextual information and output it.

[0058] Based on the above multi-source feature extraction, a spatial and channel-based attention mechanism for multi-source features is further optimized. This mechanism is as follows: By assigning different weights to different regions or channels in the feature map, the network can make fuller use of the contextual information of adjacent patches, giving higher attention to regions related to the signal target and lower attention to background noise regions, thereby forming a fused context-enhanced feature map.

[0059] The above processing yields a feature map that integrates contextual information, preserving the spatial structural differences between the central patch and its neighboring patches. This fused feature map is then used as input to the subsequent object detection module for tasks such as bounding box location prediction, category classification, and confidence calculation.

[0060] More preferably, to facilitate engineering implementation and model training, the preferred embodiment also includes an optimization strategy for the feature extraction and context fusion process in step S3, which preferably includes at least one of the following mechanisms: (1) Feature sharing mechanism For overlapping areas between tiles, convolutional features can be cached and reused to avoid repeatedly processing the same area during the sliding window process, thereby reducing redundant computation.

[0061] (2) Lightweight extraction of adjacent blocks A shallow network structure is used to extract features from adjacent plots to provide auxiliary contextual information, while the main convolutional computation is concentrated on the central plot, thereby optimizing the allocation of computational resources.

[0062] (3) Multi-scale fusion mechanism During the training phase, it is preferable to use different window sizes (e.g., 480×480, 640×640, or 800×800) to segment the time-frequency map in order to construct multi-scale feature representations; during the inference phase, a certain tile size can be fixed to maintain consistency in the detection operation.

[0063] (4) Frequency domain weighted modulation mechanism Based on the frequency direction energy distribution or prediction confidence distribution of the time-frequency plot, the channels in the fusion features are weighted to enhance the network's feature response to certain frequency domain regions.

[0064] S4: Perform object detection processing on the fused feature map and output the candidate detection results for each patch; Based on the fused feature map obtained in step S3, it is input into the target detection sub-network to perform target detection processing for the current map region.

[0065] Specifically, the target detection process in the preferred embodiment is as follows: Figure 3 As shown, and preferably includes the following steps: S41: Input the fused feature map into the object detection sub-network, predict the target location and category in the current map region, and output the corresponding candidate detection results; S42: Perform candidate target detection within the two-dimensional range of the current tile and output preliminary screening information; In a preferred embodiment, the output preliminary screening information includes the bounding box information, category information, and confidence information of the time-frequency target.

[0066] S43: Perform redundant box processing within the block; To avoid multiple overlapping candidate detection boxes in the same area, it is preferable to perform non-maximum suppression (NMS) on the candidate boxes to reduce the number of overlapping candidate boxes.

[0067] For example, in one specific implementation, it is preferable to process multiple overlapping candidate boxes based on the overlap and confidence relationship between candidate boxes, and retain one representative candidate box.

[0068] S44: Perform candidate box filtering; Candidate boxes are filtered based on a preset confidence threshold, retaining candidate detection results above the threshold and filtering out low-confidence boxes to reduce the impact of invalid candidate regions.

[0069] Through the design of steps S41 to S44 above, the location and category information of candidate targets can be obtained within the current tile space coverage area, the candidate detection results of each tile can be obtained, and the basic input can be provided for subsequent cross-tile detection fusion.

[0070] Furthermore, in actual implementation, the preferred embodiment does not limit the specific structure of the target detection sub-network, nor does it limit the specific representation of the candidate box output. As long as the requirements are met, it is acceptable, and will not be elaborated here.

[0071] S5: Perform cross-map weighted nonmaximum suppression processing on the candidate detection results for each map patch, fuse overlapping candidate boxes across map patches, and output the set of candidate boxes after weighted fusion processing.

[0072] Specifically, the processing procedure in step S5 is preferably as follows: Figure 4 As shown, it includes the following sub-steps: S51: Perform candidate box filtering for the candidate detection results of each patch, and select candidate boxes that overlap or represent the same target area as the set of candidate boxes to be merged.

[0073] S52: For a set of candidate boxes whose overlap exceeds a preset threshold, calculate the weights based on the confidence level of the candidate boxes and the location of their source tiles, and then perform fusion processing on the position or category information of the candidate boxes based on the weights.

[0074] In a specific implementation, the weight is preferably represented as the ratio of the confidence level of the candidate box to the sum of the confidence levels of the candidate box set, and the coordinates of the candidate boxes are weighted based on this weight to obtain the fused boxes.

[0075] Specifically, assume the set of overlapping candidate boxes is The corresponding confidence level is At this point, the coordinates of the merged target bounding box... The preferred method is to calculate using the following formula:

[0076] In the formula, Indicates the first i The geometric parameters of the candidate boxes; through this weighting strategy, the candidate boxes with higher confidence contribute more to the final position, thereby correcting the positional bias caused by low-confidence boxes.

[0077] S53: Perform cross-map redundancy processing, merging multiple candidate boxes from different map regions into a smaller number of fused boxes, thereby reducing the problem of duplicate detection across maps. Afterwards, output a weighted fused set of candidate boxes.

[0078] Specifically, the set of candidate boxes processed by weighted fusion serves as the input to subsequent processing logic, used to describe the approximate location and category information of the target.

[0079] In the actual design and execution of this step, the specific implementation method of weighted nonmaximum suppression and the form of the weight function are preferably not limited, as long as the aforementioned requirements are met. Similarly, in the preferred embodiment, the fusion strategy, threshold setting method, or confidence weight form are not limited, and other weighting or voting mechanisms can also be used to achieve similar processing, which will not be elaborated here.

[0080] The above-mentioned weighted nonmaximum suppression process can alleviate the problem of cross-block duplicate detection caused by tile segmentation and make the candidate box fusion process have a certain degree of contextual consistency.

[0081] S6: Map the candidate box set after weighted fusion processing to the original time-frequency map coordinate system to obtain the final detection result of the target.

[0082] Specifically, based on the weighted fusion set of candidate boxes, coordinate mapping and result update operations are performed to ensure that the candidate boxes correspond to the coordinate system of the original time-frequency diagram. This process is as follows: Figure 5 As shown, a further preferred embodiment includes the following sub-steps: S61: Based on the starting offset position of the patch in the original time-frequency graph, map the position of the candidate box inside the patch to the coordinate system of the original graph; During the tile segmentation process, each tile corresponds to a region in the original time-frequency map; at this time, the mapping relationship is preferably represented as: Original image position = tile initial offset + tile internal position The tile initial offset is used to describe the positional relationship of the tile relative to the original time-frequency map; the tile internal position is used to describe the relative position of the candidate box within the tile; in the preferred embodiment, the specific mathematical form of the coordinate mapping is not limited.

[0083] S62: Update the candidate boxes after coordinate mapping according to the coordinate system of the original time-frequency diagram.

[0084] In a specific implementation scheme, for the same or adjacent candidate targets from different map tiles, a weighted fusion strategy is preferably used to process the mapped results to reduce candidate box redundancy caused by repeated detection across map tiles. The fused or filtered candidate boxes are preferably used as structured results representing the target location and category information.

[0085] Through the above coordinate mapping and result update operations, the target detection results at the tile level can be unified to the coordinate system of the original time-frequency map, which facilitates subsequent processing or annotation.

[0086] In a preferred embodiment of the present invention, the representation of candidate boxes, fusion strategy or coordinate system type are not limited, nor are the execution timing or integration form of mapping processing limited.

[0087] Through the corresponding design of the aforementioned steps S1 to S6, the core detection path of signal acquisition and time-frequency map generation, tile segmentation and adjacency construction, multi-tile context fusion, single-tile target detection, cross-tile weighted fusion, and coordinate mapping can be realized, thereby achieving automatic detection of signal targets in shortwave time-frequency maps. By utilizing the fusion of cross-tile context information, the continuity of boundary targets is improved, thereby obtaining stable cross-tile detection results in shortwave communication image detection, which is fully applicable to application scenarios such as shortwave communication reconnaissance, spectrum monitoring, and electromagnetic situational awareness.

[0088] Furthermore, it is preferable to introduce cross-tile consistency constraints and pseudo-label mechanisms during the training phase, so as to utilize the correlation information of overlapping regions across tiles during model training, and to improve the inference efficiency of the model in conjunction with optimization strategies during the deployment phase.

[0089] Specifically, the aforementioned training optimization mechanism preferably includes: (1) Constructing cross Figure 1 Consistency loss function: In areas where tiles overlap, it is preferable to introduce cross-area mapping. Figure 1 Consistency loss is used to ensure that the network's predictions remain consistent across different maps. Among these, the aforementioned cross-regional... Figure 1 Consistency loss preferably consists of multiple loss terms, including but not limited to: Position-dependent loss: used to represent the spatial relationship between candidate boxes and reference boxes; Category correlation loss: used to represent the difference between predicted class distributions; Confidence difference loss: used to represent the confidence difference between candidate boxes and different tiles.

[0090] In a preferred embodiment, the number, form, and specific construction of the weighting function of the loss term are not limited.

[0091] In a specific implementation, the aforementioned three loss terms are preferably combined in a weighted manner to form a consistency loss, which is used to optimize the network during the training phase.

[0092] (2) Self-supervised mechanism of pseudo-labels: In one specific implementation plan, a pseudo-label self-supervised mechanism is preferably introduced to expand the training samples, which is then used to perform self-supervised training on the target detection network with insufficient training data.

[0093] Specifically, when training data is insufficient, pseudo-labels are preferably generated based on high-confidence prediction results in adjacent patches, and the pseudo-labels are filtered in combination with confidence information. The filtered pseudo-labels are used to assist training, thereby utilizing more data samples without manual annotation.

[0094] More specifically, during the screening process, there are no restrictions on the specific value of the confidence threshold; it can be set according to actual needs to distinguish between acceptable and unacceptable pseudo-labels.

[0095] Through the above training mechanism, the network parameters can be adjusted by utilizing the relationship between cross-tile regions during the model training stage, and the training set can be expanded by pseudo-labels in scenarios with limited data, further enhancing the detection model's adaptability to the time-varying nature of shortwave signals and complex background scenarios.

[0096] Example 2: As another aspect of the present invention, a shortwave time-frequency map target detection system based on multi-map fusion is also provided. The system operation logic corresponds to the shortwave time-frequency map target detection method in Embodiment 1, with a focus on protecting the corresponding module structure.

[0097] Specifically, the system architecture of the shortwave time-frequency map target detection system in the preferred embodiment is as follows: Figure 6 As shown, it includes: (1) Time-frequency diagram construction module, used to construct a time-frequency diagram based on the acquired shortwave signal; In practical configuration, the time-frequency graph construction module is also equipped with a signal acquisition module, which is used to acquire shortwave electromagnetic signals covering the target frequency band and output sampled data in complex IQ form to preserve the amplitude and phase information of the signal. Simultaneously, in the preferred embodiment, the time-frequency graph construction module is used to perform short-time Fourier transform (STFT) or other time-frequency analysis processing on the IQ signals to generate a time-frequency graph, and to perform preprocessing such as normalization on the generated two-dimensional image data.

[0098] (2) A tile generation module, used to divide the time-frequency map into multiple tiles and construct a multi-channel joint detection input containing a central tile and at least one adjacent tile of the central tile; In actual operation, the tile generation module performs a sliding window segmentation operation on the preprocessed time-frequency graph to generate several tiles of the same size, and constructs tile adjacency information based on the positional relationship of the tiles in the original time-frequency graph. On this basis, individual tiles and their adjacent tiles are stacked along the channel dimension to obtain multi-channel joint detection input.

[0099] (3) Context fusion module, used to perform feature extraction and context information fusion on multi-channel joint detection input, and output fused feature map; As an example, in the preferred embodiment, the context fusion module further performs feature extraction based on a feature extraction network and achieves context fusion based on a location information encoding structure.

[0100] More specifically, the context fusion module is used to input the central patch and its adjacent patches into the feature extraction network, and introduces positional encoding during the feature extraction process; whereby the positional encoding is used to represent the relative positional relationship of each patch in the time-frequency map. Based on this, a spatial-channel attention mechanism is applied to the extracted multi-source feature maps to generate a fused feature map in a weighted fusion manner.

[0101] (4) Target detection module, used to perform target detection processing on the fused feature map and output the candidate detection results for each map patch; wherein, the candidate detection results preferably include the spatial location information, category information or confidence information of the target.

[0102] (5) Cross-tile fusion module, which performs weighted nonmaximum suppression processing on the candidate detection results of each tile, fuses overlapping candidate boxes across tiles, and outputs a set of candidate boxes after weighted fusion processing; (6) Coordinate mapping module, used to map the candidate box set after weighted fusion processing to the original time-frequency map coordinate system and output the final detection result of the target.

[0103] In practice, coordinate mapping is preferably performed based on the relationship between the starting offset position of the block and its relative position within the block.

[0104] More preferably, in order to improve the model inference efficiency of the object detection system during the deployment phase, the system in the preferred embodiment further includes a training optimization module, which is used to introduce cross-tile consistency constraints and pseudo-label mechanisms during the model training phase, and to expand the training data or adjust the network parameters by utilizing the association information between tiles.

[0105] More specifically, in the preferred embodiment, the cross-tile consistency constraint is used to constrain the target detection network to output consistent prediction results in overlapping areas of different tiles, thereby maintaining the consistency of the prediction results output.

[0106] The pseudo-label mechanism is used to utilize high-confidence prediction results from neighboring tiles as pseudo-labels when labeled data is limited, and then use these pseudo-labels as the target detection network for self-supervised training.

[0107] By combining the above modules, a processing flow for shortwave signals can be realized, from acquisition, time-frequency construction, tile generation, context fusion, candidate detection, result mapping to training and optimization. This enables shortwave time-frequency map target detection based on multi-tile context information fusion, ensuring consistency and accuracy in cross-tile detection.

[0108] Example 3: Regarding the method in Embodiment 1, the method steps can be implemented by an electronic device executing program instructions, or by software, hardware, or a combination of both.

[0109] The program instructions can be stored in a computer-readable storage medium, and the electronic device involved includes a processor and a memory. The processor executes the program instructions in the memory to implement the method steps in Embodiment 1. Conversely, the software is preferably deployed on a server, embedded device, or other processing device with computing capabilities.

[0110] Therefore, as another aspect of the present invention, a storage medium is also provided, wherein a processor-executable program is stored, which, when executed by a processor, is used to perform the shortwave time-frequency map target detection method based on multi-map tile context information fusion described in Embodiment 1.

[0111] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A shortwave time-frequency map target detection method based on multi-map patch fusion, characterized in that, Includes the following steps: S1: Preprocess the acquired shortwave signal to construct a time-frequency diagram of the shortwave signal; S2: Divide the time-frequency graph into blocks and select at least one adjacent block for each central block obtained from the division; The central patch and adjacent patches are stacked along the channel dimension to construct a multi-channel joint detection input; where: The at least one adjacent tile includes: a left adjacent tile adjacent to the center tile in the time axis direction; and / or, an upper adjacent tile adjacent to the center tile in the frequency axis direction; and For a central tile that does not have a left adjacent tile and / or an upper adjacent tile, the adjacent area is filled using zero-fill or mirror expansion. S3: Perform feature extraction based on multi-tile joint input on the multi-channel joint detection input, and perform fusion processing by combining the context information of adjacent tiles to enhance the feature continuity of the edge region of the central tile and reduce the inconsistency of target information at cross-tile boundaries, thereby obtaining a fused feature map; the construction process of the fused feature map specifically includes: A convolutional neural network is used to extract features from the central patch and its adjacent patches, and position encoding is introduced during the feature extraction process; wherein, the position encoding is used to represent the relative positional relationship of each patch in the time-frequency graph; The extracted multi-source features are applied to a spatial-channel attention mechanism, and the features are weighted according to the spatial positional relationship between the central patch and its neighboring patches to generate a fused feature map in a weighted fusion manner. S4: Perform object detection processing on the fused feature map and output the candidate detection results for each patch; S5: Perform cross-map weighted nonmaximum suppression processing on the candidate detection results for each map tile, fuse overlapping candidate boxes across map tiles, and output the set of candidate boxes after weighted fusion processing; S6: Map the candidate box set to the original time-frequency map coordinate system to obtain the final detection result of the target.

2. The shortwave time-frequency map target detection method based on multi-map patch fusion according to claim 1, characterized in that, In step S1, the construction process of the time-frequency diagram specifically includes: Signal sampling over a wide bandwidth is performed using a complex IQ sampling method; A two-dimensional time-frequency graph is constructed using short-time Fourier transform. The construction process includes: configuring the window function, window length, and frame shift parameters; performing spectrum calculation and power spectrum processing; normalizing the power spectrum; and organizing the normalized two-dimensional matrix with time as the horizontal axis and frequency as the vertical axis to form a two-dimensional grayscale time-frequency graph.

3. The shortwave time-frequency map target detection method based on multi-map patch fusion according to claim 1, characterized in that, In step S2, when dividing the map, an overlapping area is set between adjacent maps, and the overlap width is 5% to 15% of the adjacent direction dimension of the map.

4. The shortwave time-frequency map target detection method based on multi-map patch fusion according to any one of claims 1 to 3, characterized in that, The candidate detection results obtained in step S4 also include: Non-maximum suppression is applied to multiple overlapping candidate detection boxes appearing in the same region; and / or, Candidate detection boxes are filtered based on a preset confidence threshold.

5. The shortwave time-frequency map target detection method based on multi-map patch fusion according to any one of claims 1 to 3, characterized in that, In step S5, the weighted fusion process includes: Candidate bounding boxes are filtered based on the candidate detection results for each map patch, and a set of candidate bounding boxes to be fused is selected. For a set of candidate boxes whose overlap exceeds a preset threshold, weights are calculated based on the confidence level of the candidate boxes and the location of their source tiles, and the position or category information of the candidate boxes is fused based on the weights. Perform cross-map redundancy processing to merge multiple candidate boxes from different map regions into a smaller number of merged boxes.

6. The shortwave time-frequency map target detection method based on multi-map patch fusion according to any one of claims 1 to 3, characterized in that, The method also includes a model training optimization step: S7: During the model training phase, construct a cross-graph consistency loss function to constrain the object detection network to output consistent prediction results in the overlapping areas of different graph tiles; and / or, use high-confidence prediction results in adjacent graph tiles to generate pseudo-labels to perform self-supervised training on the object detection network with insufficient training data.

7. A shortwave time-frequency map target detection system based on multi-map patch fusion, characterized in that, include: The time-frequency diagram construction module is used to construct a time-frequency diagram based on the acquired shortwave signal; The tile generation module is used to divide the time-frequency map into multiple tiles and construct a multi-channel joint detection input containing a central tile and at least one adjacent tile of the central tile; wherein, the at least one adjacent tile includes: a left adjacent tile adjacent to the central tile in the time axis direction; and / or, an upper adjacent tile adjacent to the central tile in the frequency axis direction; and for the central tile that does not have a left adjacent tile and / or an upper adjacent tile, the adjacent area is filled in by zero-filling or mirror expansion. The context fusion module is used to perform feature extraction and context information fusion on the multi-channel joint detection input, and output a fused feature map. The specific process of the context fusion module outputting the fused feature map includes: using a convolutional neural network to extract features from the central patch and its adjacent patches, and introducing position encoding during the feature extraction process; wherein the position encoding is used to represent the relative positional relationship of each patch in the time-frequency map; applying a spatial-channel attention mechanism to the extracted multi-source features, and weighting the features according to the spatial positional relationship between the central patch and adjacent patches, generating a fused feature map in a weighted fusion manner. The object detection module is used to perform object detection processing on the fused feature map and output the candidate detection results for each patch; The cross-tile fusion module is used to perform weighted nonmaximum suppression processing on the candidate detection results of each tile, fuse overlapping candidate boxes across tiles, and output a set of candidate boxes after weighted fusion processing; The coordinate mapping module is used to map the weighted fusion candidate box set to the original time-frequency map coordinate system and output the final detection result of the target.

8. A storage medium storing a processor-executable program, characterized in that, When executed by a processor, the program is used to perform the shortwave time-frequency map target detection method based on multi-map fusion as described in any one of claims 1 to 6.

Images