An image weak and small target detection method

Abstract

This invention discloses a method for detecting weak targets in images, relating to the fields of computer vision and target detection technology. The method includes: inputting an acquired target image to a target detection network model; performing feature decomposition using a learnable biorthogonal bidirectional lifting method to obtain fused features of the lifted feature subbands; and performing target detection based on the fused features of the lifted feature subbands using a bounding box localization path and a target recognition path to obtain target recognition results and bounding box localization results. The target detection network model is trained using the acquired image dataset, based on stopband leakage suppression constraints, biorthogonal consistency constraints, and basic detection constraints. This invention constructs bounding box localization paths and target recognition paths for weak target detection in complex backgrounds, improving the accuracy of weak target detection in tasks such as UAV aerial photography, remote sensing observation, and low-altitude inspection.

Description

Technical Field

[0001] This invention belongs to the field of computer vision and target detection technology, and specifically relates to a method for detecting weak targets in images. Background Technology

[0002] In the field of computer vision target detection, whether it's target detection in drone aerial photography, target detection in remote sensing images, or target recognition in low-altitude patrols, the core objective is to accurately identify target categories and determine their spatial locations from complex backgrounds. However, it's worth noting that small targets in drone, remote sensing, and low-altitude observation images typically exhibit characteristics such as small scale, weak signal, and indistinct edge contours. In urban blocks, road areas, parking lots, or open scenes, small targets often occupy only a few pixels, and their grayscale variations, edge directions, and local shape differences easily blend with road textures, building edges, tree shadows, and dense background clutter. For these targets, the effective information available for distinguishing the background is very limited, often concentrated in a few edges, corners, weak textures, and local brightness abrupt changes. These detailed clues provide important information support for the detection network, but also place higher demands on the compression process at the network's front end: if shallow features are compressed too early or mixed with the background response during the downsampling stage, even with multi-scale compensation by the subsequent feature fusion structure and detection head, it will be difficult to fully recover the early-lost information about small targets.

[0003] Existing weak target detection technologies have many limitations:

[0004] In terms of front-end feature compression, using a single-stage detection network to perform downsampling for shallow feature compression can easily lead to aliasing of high and low frequency information during the compression process. Key discrimination clues for weak targets are mostly concentrated in high-frequency components. If the front-end compression process lacks effective frequency band division and constraints, these details may be co-encoded with background structures and high-frequency clutter, causing the target response to be compressed or mixed in the shallow stage.

[0005] In terms of frequency band decomposition, the fixed wavelet basis filter form is preset. Although it has stable frequency division and strong interpretability, it is difficult to adaptively adjust the frequency response according to the changes in target size, shooting altitude, background texture and imaging conditions in the UAV scene.

[0006] Regarding the utilization of frequency band cues at the detection end, for weak targets, target recognition relies more on local texture, edge differences, and category-related details, while bounding box localization relies more on the target's main body region, boundary continuity, and scale structure. The structural and detail cues obtained from front-end frequency band decomposition are difficult to reliably translate into detection benefits. Summary of the Invention

[0007] The purpose of this invention is to provide a method for detecting weak targets in images. By constructing stopband leakage suppression constraints, biorthogonal consistency constraints, and basic detection constraints, the stable separability between low-frequency structures and high-frequency details in the detection of weak targets is improved.

[0008] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0009] On one hand, the present invention provides a method for detecting weak targets in images, comprising:

[0010] The acquired target image to be processed is input into the target detection network model, and feature decomposition is performed through a learnable bioorthogonal bidirectional lifting method to obtain the fused features of the feature-lifted subbands;

[0011] Based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path to obtain the target recognition result and the bounding box localization result.

[0012] The training method for the target detection network model includes:

[0013] Based on the acquired baseline image to be detected, a shallow input feature map is obtained through shallow feature extraction of the target detection network model;

[0014] Based on the shallow input feature map, a learnable bioorthogonal bidirectional boosting method is used to perform feature decomposition, resulting in multiple feature boosting subbands; wherein the feature boosting subbands include low-frequency structure subbands and high-frequency detail subbands.

[0015] Based on the feature enhancement subband, update the bounding box localization path and target recognition path, and construct stopband leakage suppression constraints;

[0016] Basic detection constraints are constructed based on the bounding box localization path and the target recognition path;

[0017] Based on the predictive update parameters of the learnable bioorthogonal bidirectional boosting method, a bioorthogonal consistency constraint is constructed.

[0018] Based on stopband leakage suppression constraints, bioorthogonal consistency constraints, and basic detection constraints, an object detection network model is trained using the acquired image dataset.

[0019] Optionally, the target image to be processed is input into a target detection network model, and feature decomposition is performed using a learnable bioorthogonal bidirectional boosting method, including:

[0020] Based on the acquired target image to be processed, a shallow input feature map of the target image to be processed is obtained through normalization processing and shallow feature extraction.

[0021] Based on the shallow input feature map of the target image to be processed, feature decomposition is performed using a learnable biorthogonal bidirectional lifting method to obtain the feature-enhanced subband of the target image to be processed.

[0022] Based on the feature enhancement subbands of the target image to be processed, channel projection is performed using point convolution to obtain the fused features of the feature enhancement subbands.

[0023] Optionally, based on the shallow input feature map, a learnable bioorthogonal bidirectional lifting method is used for feature decomposition to obtain multiple feature-lifted subbands, including:

[0024] Based on the shallow input feature map, frequency band feature decomposition is performed along the row direction of the feature map through odd and even positions to obtain the row direction decomposed features;

[0025] Based on the row direction decomposition features, frequency band feature decomposition is performed along the column direction of the feature map through odd and even positions to obtain multiple feature enhancement sub-bands.

[0026] Optionally, based on the shallow input feature map, frequency band feature decomposition is performed along the row direction of the feature map through odd and even positions to obtain row direction decomposed features, including:

[0027] Based on the shallow input feature map, even-sampled features and odd-sampled features are obtained by dividing the feature map into even-sampled features and odd-sampled features according to the odd and even positions of the row directions.

[0028] Based on the even-sampling features, feature prediction is performed using a prediction operator to obtain the odd-sampling feature prediction result;

[0029] Based on the odd sampling feature prediction results and odd sampling features, high-frequency detail components are obtained by difference calculation.

[0030] Based on the high-frequency detail components, the low-frequency structure components are obtained by correcting the dual sampling features using the update operator.

[0031] Based on the high-frequency detail components and low-frequency structural components, the row direction decomposition features are obtained by adjusting the channel-by-channel scale using a scale factor.

[0032] Optionally, bioorthogonal consistency regularization terms can be constructed based on the prediction operator, update operator, and scaling factor in the learnable bioorthogonal bidirectional boosting method. The calculation expression is:

[0033]

[0034] in, and These are the weighting coefficients. For reference proportionality, Norm symbol; Let and k be the learnable prediction operator, the learnable update operator, and the learnable channel scaling factor, respectively; 1 and 1 are the preset basic reference values ​​for the learnable prediction operator, the learnable update operator, and the learnable channel scale factor, respectively.

[0035] Based on the bioorthogonal consistency regularization term, bioorthogonal consistency constraints are constructed, and the calculation expression is as follows:

[0036]

[0037] in, This is a bioorthogonal consistency constraint; For row direction biorthogonal consistency regularization terms; For column-direction biorthogonal consistency regularization terms.

[0038] Optionally, constructing stopband leakage suppression constraints based on the aforementioned features to enhance the sub-band includes:

[0039] Based on the aforementioned feature-enhanced subband, the spectral energy corresponding to the feature-enhanced subband is calculated using Fourier transform.

[0040] The target frequency band mask is confirmed based on the feature-enhanced subband, and the corresponding stopband region is constructed.

[0041] Based on the spectral energy, target frequency band mask, and stopband region, the stopband relative response is obtained by calculating the average stopband energy.

[0042] A stopband leakage suppression constraint is constructed based on the aforementioned stopband relative response;

[0043] The calculation expression for the bandgap leakage suppression constraint is as follows:

[0044]

[0045] in, To suppress band leakage, For activation function, To achieve the desired stopband attenuation margin, For the number of sub-bands, It is a very small constant. This is the stopband relative response.

[0046] Optionally, updating the bounding box localization path and target recognition path based on the feature enhancement sub-band includes:

[0047] Based on the low-frequency structural subband and the high-frequency detail subband, low-frequency reconstruction components and high-frequency reconstruction components are constructed through reverse reconstruction, respectively;

[0048] Based on the low-frequency reconstruction component and the high-frequency reconstruction component, the low-frequency structure guide map and the high-frequency detail guide map are calculated by the mean of the channel absolute response, respectively.

[0049] Calculate the low-frequency and high-frequency guidance reliability coefficients based on the spectral energy corresponding to the feature-enhanced subbands.

[0050] The low-frequency structure guidance map and low-frequency guidance reliability coefficient are injected into the regression path of the detection branch through gated residual method to update the bounding box positioning path;

[0051] The high-frequency detail guide map and high-frequency guide reliability coefficient are injected into the classification path of the detection branch through a gated residual method to update the target recognition path.

[0052] Optionally, the low-frequency guidance reliability coefficient and high-frequency guidance reliability coefficient are calculated based on the spectral energy corresponding to the feature-enhanced sub-band, including:

[0053] Based on the aforementioned features, the spectral energy corresponding to the sub-band is enhanced, and a probabilistic spectral distribution is obtained through normalization.

[0054] Based on the probabilistic spectrum distribution, the subband purity corresponding to the subband is improved by calculating features through a preset subband spectrum template;

[0055] The low-frequency guidance reliability coefficient and the high-frequency guidance reliability coefficient are calculated based on the subband purity using a truncation function.

[0056] Optionally, based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path, including:

[0057] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the bounding box localization path to obtain the regression path input features after bounding box localization path modulation. The calculation expression is as follows:

[0058] ;

[0059] Where Xreg is the regression path input feature after bounding box localization path modulation; β is the learnable gating coefficient of the bounding box localization path; The reliability coefficient for low-frequency guidance; To enhance the fusion features of subbands, This is a low-frequency structure guide diagram after the convolution channel is projected.

[0060] Based on the regression path input features modulated by the bounding box localization path, the bounding box localization result is obtained through regression prediction.

[0061] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the target recognition path to obtain the target recognition path modulated classification path input features. The calculation expression is as follows:

[0062] ;

[0063] Where Xcls is the input feature of the classification path after the target recognition path is modulated; α is the learnable gating coefficient of the target recognition path; For high-frequency guidance reliability coefficient; To enhance the fusion features of subbands, This is a guide map of high-frequency details after projection of the convolutional channels;

[0064] Based on the input features of the classification path after target recognition path modulation, the target recognition result is obtained through classification prediction.

[0065] Optionally, a bounding box regression loss function can be constructed based on learnable gating coefficients of the bounding box localization path;

[0066] A classification loss function is constructed based on the learnable gating coefficients of the target recognition path;

[0067] Based on the bounding box regression loss function and classification loss function, the basic detection constraints are constructed, and the calculation expression is as follows:

[0068]

[0069] in, Based on the detection constraints, The bounding box regression loss function is... This is the classification loss function.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] This invention employs a learnable biorthogonal bidirectional enhancement method to divide the target image into low-frequency structural sub-bands and multiple high-frequency detail sub-bands before it enters the subsequent backbone network. This reduces the risk of detail cues of small targets in complex backgrounds being weakened or mixed with the background response during the front-end compression stage. By training the target detection network model with constructed basic detection constraints, biorthogonal consistency constraints, and stopband leakage suppression constraints, the model maintains stable separability between low-frequency structures and high-frequency details while adapting to the target detection task. This mitigates the problems of frequency band degradation, increased stopband response, and sub-band response chaos that may occur with learnable decomposition. This invention is applicable to scenarios such as UAV aerial photography, remote sensing observation, and low-altitude inspection, improving the accuracy of small target detection.

[0072] This invention utilizes low-frequency structure guidance maps, low-frequency guidance reliability coefficients, high-frequency detail guidance maps, and high-frequency guidance reliability coefficients to construct bounding box localization paths and target recognition paths, ensuring that the structural and detail cues obtained from front-end frequency band decomposition can directly participate in the detection of weak targets, thus achieving accurate recognition of weak targets. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the overall network framework of the weak target detection method in this invention;

[0074] Figure 2 This is a schematic diagram of the learnable biorthogonal bidirectional lifting method in this invention;

[0075] Figure 3 This is a schematic diagram of the structure of the training loss function of the object detection network model in this invention;

[0076] Figure 4 This is a schematic diagram of the frequency band clue utilization mechanism for the detection branch in this invention;

[0077] Figure 5 This is a comparison chart of the performance of the target detection network model in this invention and existing target detectors;

[0078] Figure 6 This is a comparison chart of the effects of multiple modules within the target detection network model in this invention. Detailed Implementation

[0079] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0080] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0081] Example 1

[0082] This embodiment introduces a method for detecting weak targets in images, specifically including:

[0083] The acquired target image is input into the target detection network model. Feature decomposition is performed using a learnable biorthogonal bidirectional lifting method to obtain fused features of the enhanced sub-bands. The front-end compression part of the target detection network model extracts low-frequency structure sub-bands and high-frequency detail sub-bands of the target image using feature decomposition with the learnable biorthogonal bidirectional lifting method. To ensure the image can be input into the target detection network model, size and pixel normalization are performed on the acquired target image. The features obtained from the front-end compression part of the target detection network model are used for target detection through the constructed bounding box localization path and target recognition path to obtain target recognition results and bounding box localization results.

[0084] The training method for the target detection network model includes:

[0085] Based on the acquired baseline image to be detected, a shallow input feature map is obtained through shallow feature extraction of the target detection network model;

[0086] Based on the shallow input feature map, a learnable bioorthogonal bidirectional boosting method is used to perform feature decomposition, resulting in multiple feature boosting subbands; wherein the feature boosting subbands include low-frequency structure subbands and high-frequency detail subbands.

[0087] Based on the aforementioned features, sub-band enhancement is performed, and stopband leakage suppression constraints are constructed by calculating the equivalent frequency response.

[0088] The bounding box localization path and the target recognition path are constructed based on the low-frequency structural sub-band and the high-frequency detail sub-band, respectively.

[0089] Basic detection constraints are constructed based on the bounding box localization path and the target recognition path;

[0090] A bioorthogonal consistency constraint is constructed using the prediction update parameters of a learnable bioorthogonal bidirectional boosting method.

[0091] Based on stopband leakage suppression constraints, bioorthogonal consistency constraints, and basic detection constraints, an object detection network model is trained using the acquired image dataset. The image dataset can be an existing publicly available dataset used to train the object detection network model.

[0092] In summary, this embodiment reduces the risk that detailed cues of small targets in complex backgrounds will be weakened or mixed with the background response during the front-end compression stage by using a learnable biorthogonal bidirectional boosting method.

[0093] Example 2

[0094] Based on the same inventive concept as Embodiment 1, this embodiment introduces a method for detecting weak targets in images, specifically including:

[0095] The acquired target image is input into the target detection network model, and feature decomposition is performed using a learnable bioorthogonal bidirectional boosting method to obtain the fused features of the boosted subbands, including:

[0096] Based on the acquired target image to be processed, a shallow input feature map of the target image to be processed is obtained through normalization processing and shallow feature extraction.

[0097] Based on the shallow input feature map of the target image to be processed, feature decomposition is performed using a learnable biorthogonal bidirectional lifting method to obtain the feature-enhanced subband of the target image to be processed.

[0098] Based on the feature enhancement subbands of the target image to be processed, channel projection is performed using point convolution to obtain the fused features of the feature enhancement subbands.

[0099] Based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path to obtain the target recognition result and the bounding box localization result.

[0100] The training method for the target detection network model includes:

[0101] Based on the acquired baseline image to be detected, a shallow input feature map is obtained through shallow feature extraction of the target detection network model;

[0102] Based on the shallow input feature map, a learnable bioorthogonal bidirectional boosting method is used for feature decomposition to obtain multiple feature boosting subbands. These feature boosting subbands include low-frequency structure subbands and high-frequency detail subbands, specifically including:

[0103] Based on the shallow input feature map, frequency band feature decomposition is performed along the row direction of the feature map through odd and even positions to obtain row direction decomposition features. Based on the shallow input feature map, even sampling features and odd sampling features are obtained by dividing the feature map through odd and even positions along the row direction.

[0104] Based on the even-sampling features, feature prediction is performed using a prediction operator to obtain the odd-sampling feature prediction result;

[0105] Based on the odd sampling feature prediction results and odd sampling features, high-frequency detail components are obtained by difference calculation.

[0106] Based on the high-frequency detail components, the low-frequency structure components are obtained by correcting the dual sampling features using the update operator.

[0107] Based on the high-frequency detail components and low-frequency structural components, the row direction decomposition features are obtained by adjusting the channel-by-channel scale using a scale factor.

[0108] Based on the row-direction decomposition features, frequency band feature decomposition is performed along the column direction of the feature map through odd and even positions to obtain multiple feature enhancement sub-bands. The decomposition method used when decomposing the row-direction features is the same as the method for frequency band feature decomposition based on the shallow input feature map along the row direction of the feature map through odd and even positions.

[0109] The bioorthogonal consistency constraint is constructed based on the prediction operator, update operator, and scaling factor in the learnable bioorthogonal bidirectional boosting method.

[0110] The stopband leakage suppression constraint is constructed based on feature-enhanced subbands, specifically including:

[0111] Based on the aforementioned feature-enhanced subband, the spectral energy corresponding to the feature-enhanced subband is calculated using Fourier transform.

[0112] The target frequency band mask is confirmed based on the feature-enhanced subband, and the corresponding stopband region is constructed.

[0113] Based on the spectral energy, target frequency band mask, and stopband region, the stopband relative response is obtained by calculating the average stopband energy.

[0114] A stopband leakage suppression constraint is constructed based on the relative stopband response.

[0115] Based on the low-frequency structural subband and the high-frequency detail subband, low-frequency reconstruction components and high-frequency reconstruction components are constructed through reverse reconstruction, respectively;

[0116] Based on the low-frequency reconstruction component and the high-frequency reconstruction component, the low-frequency structure guide map and the high-frequency detail guide map are calculated by the mean of the channel absolute response, respectively.

[0117] The low-frequency and high-frequency guidance reliability coefficients are calculated based on the spectral energy corresponding to the feature-enhanced subbands, including:

[0118] Based on the aforementioned features, the spectral energy corresponding to the sub-band is enhanced, and a probabilistic spectral distribution is obtained through normalization.

[0119] Based on the probabilistic spectrum distribution, the subband purity corresponding to the subband is improved by calculating features through a preset subband spectrum template;

[0120] The low-frequency guidance reliability coefficient and the high-frequency guidance reliability coefficient are calculated based on the subband purity using a truncation function.

[0121] The low-frequency structure guidance map and low-frequency guidance reliability coefficient are injected into the regression path of the detection branch through gated residual method to update the bounding box positioning path;

[0122] The high-frequency detail guide map and high-frequency guide reliability coefficient are injected into the classification path of the detection branch through a gated residual method to update the target recognition path.

[0123] Based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path, including:

[0124] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the bounding box localization path to obtain the regression path input features after bounding box localization path modulation. The calculation expression is as follows:

[0125]

[0126] Where Xreg is the regression path input feature after bounding box localization path modulation; β is the learnable gating coefficient of bounding box localization path; is the low-frequency guidance reliability coefficient; is the fusion feature of feature enhancement subband; and is the low-frequency structure guidance map after convolution channel projection.

[0127] Based on the regression path input features modulated by the bounding box localization path, the bounding box localization result is obtained through regression prediction.

[0128] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the target recognition path to obtain the target recognition path modulated classification path input features. The calculation expression is as follows:

[0129]

[0130] Where Xcls is the input feature of the classification path after the target recognition path is modulated; α is the learnable gating coefficient of the target recognition path; is the high-frequency guidance reliability coefficient; is the fusion feature of the feature enhancement sub-band; and is the high-frequency detail guidance map after the convolution channel projection. Based on the input feature of the classification path after the target recognition path is modulated, the target recognition result is obtained through classification prediction.

[0131] The basic detection constraints are constructed based on the learnable gating coefficients of the bounding box localization path and the learnable gating coefficients of the target recognition path.

[0132] Based on stopband leakage suppression constraints, bioorthogonal consistency constraints, and basic detection constraints, an object detection network model is trained using the acquired image dataset.

[0133] Example 3

[0134] Based on the same inventive concept as Embodiment 2, this embodiment describes an implementation process for detecting weak targets in images, such as... Figure 1The target detection network model shown includes a learnable biorthogonal bidirectional lifting decomposition module BiBLD, a spectral separability constraint module based on equivalent frequency response EFR-SSC, and a frequency band cue utilization mechanism module DFCI oriented towards detection tasks. The spectral separability constraints include: stopband leakage suppression constraints, biorthogonal consistency constraints, and basic detection constraints. DFCI transforms the front-end BiBLD decomposition results into low-frequency structure guidance information and high-frequency detail guidance information, and then introduces high-resolution detection branches in a residual manner.

[0135] The specific implementation steps are as follows: Figure 1 As shown, the target image input to the target detection network model is subjected to BiBLD feature enhancement decomposition to obtain the low-frequency structure subband. and high-frequency detail subband , , Where LL represents the low-frequency sub-band channel, and LH, HL, and HH are all high-frequency sub-band channels; specifically defined as follows: HH represents a dual high-frequency sub-band with high frequencies in both row and column directions; HL represents a sub-band with high frequencies in the row direction and low frequencies in the column direction; LH represents a sub-band with low frequencies in the row direction and high frequencies in the column direction; LL represents a dual low-frequency sub-band with low frequencies in both row and column directions; the feature enhancement sub-band is projected through 1×1 convolution to obtain the fused features of the feature enhancement sub-band not acquired by the detection head. The Head is the detection head of the object detection network model. The Head is connected to BiBLD through the Neck network.

[0136] BiBLD can employ a one-dimensional Predict-Update bidirectional lifting unit for feature enhancement decomposition, as detailed below. Figure 2 As shown: The input feature X is divided into even and odd sampling, resulting in even-sampled features. Sampling features :

[0137]

[0138] in, and These represent even-position sampling and odd-position sampling, respectively. Even-sampling features mainly preserve local low-frequency approximations, while odd-sampling features contain information about the variations between adjacent positions.

[0139] Predicting odd-sampled features using even-sampled features yields the odd-sampled feature prediction results. :

[0140]

[0141] in, This indicates a learnable predictive operator. Its parameters. The difference between the true odd-sampled features and the predicted results is used as the high-frequency detail component. We can obtain:

[0142]

[0143] High-frequency detail components It primarily preserves variation information between even and odd sampling locations that cannot be explained by local low-frequency approximations, including edges, textures, corners, and abrupt changes in local brightness. After obtaining the high-frequency residuals, the lifting step uses the high-frequency detail component residual information to correct the even-sampling features, thus obtaining the low-frequency structural components. :

[0144]

[0145] in, This indicates a learnable update operator, with θu as its parameter. This step feeds back information related to local structure from the high-frequency detail components to the low-frequency components, enabling the low-frequency components to maintain a more stable main area, contour continuity, and background structure while reducing resolution.

[0146] For low-frequency structural components and high-frequency detail components Performing scale adjustments in the opposite direction yields low-frequency and high-frequency subbands:

[0147]

[0148]

[0149] in, This represents the learnable channel scaling factor. Its parameters, This indicates channel-wise multiplication. This design allows the response amplitudes of low-frequency structures and high-frequency details to be adjusted within a controlled range, avoiding excessively strong or weak responses from any component during training, thereby improving the numerical stability of the boosting decomposition. In this embodiment, frequency band feature decomposition is performed along the row direction of the shallow input feature map through odd and even positions to obtain row-direction decomposed features; based on the row-direction decomposed features, frequency band feature decomposition is performed along the column direction of the feature map through odd and even positions to obtain multiple feature boosting sub-bands. The feature boosting sub-bands are output through LL, LH, HL, and HH, respectively. , , , .

[0150] The spectral separability constraint EFR-SSC is constructed based on the equivalent frequency response of each sub-band, specifically as follows: Figure 3As shown, it includes:

[0151] Based on the aforementioned feature-enhanced subband, the spectral energy corresponding to the feature-enhanced subband is calculated using Fourier transform.

[0152] Feature-enhanced subbands Confirm target frequency band mask And construct the corresponding stopband region ;

[0153] Based on the spectral energy The target frequency band mask and stopband region are used to calculate the relative stopband response by means of the stopband energy. ;

[0154] Based on the stopband relative response, a stopband leakage suppression constraint is constructed. ; For activation function, To achieve the desired stopband attenuation margin, This represents the number of sub-bands. When the stopband strong response is already below the passband average by a set margin, that sub-band will no longer be subject to additional penalties.

[0155] The low-frequency and high-frequency guidance reliability coefficients are calculated based on the spectral energy corresponding to the feature-enhanced subbands, specifically including:

[0156] Based on the aforementioned features, the spectral energy corresponding to the sub-band is enhanced, and a probabilistic spectral distribution is obtained through normalization.

[0157]

[0158] in, For the spectral energy distribution, For probabilistic spectral distribution, This is a minimal constant used to avoid a denominator of zero. is a coordinate constant used to represent frequency coordinates.

[0159] Based on the probabilistic spectral distribution, the purity of the corresponding subband is improved by calculating features using a preset subband spectral template:

[0160]

[0161] in, Subband purity, including low-frequency purity. and high frequency purity , , ;

[0162] The construction of the preset sub-band spectrum templates includes: determining the lateral soft low-pass response, lateral soft high-pass response, longitudinal soft low-pass response, and longitudinal soft high-pass response using preset normalized frequency coordinates, lateral amplitude, longitudinal amplitude, division threshold, and transition bandwidth; and calculating the soft spectrum templates corresponding to the four sub-bands based on the lateral soft low-pass response, lateral soft high-pass response, longitudinal soft low-pass response, and longitudinal soft high-pass response. .

[0163] The low-frequency guidance reliability coefficient is calculated based on the subband purity using a truncation function. and high-frequency guidance reliability coefficient The calculation formula is:

[0164]

[0165]

[0166] in, () represents the truncation function. Used to control the intensity of low-frequency structural guidance information introduction. This is used to control the intensity of the introduction of high-frequency detail guidance information. If the purity of the high-frequency subbands obtained from the current front-end decomposition is low, the impact of high-frequency guidance information on the classification branch will be weakened. If the low-frequency structure is more stable, its auxiliary effect on the regression branch will be enhanced.

[0167] Construct a bioorthogonal consistency regularization term based on the learnable prediction operator, the learnable update operator, and the learnable channel scaling factor;

[0168] When the learnable prediction operator, learnable update operator, and learnable channel scaling factor are respectively adopted and k, and its default base reference value is When the sum is 1, the bioorthogonal consistent regularization term for:

[0169]

[0170] in, and These are the weighting coefficients. For reference proportionality, The norm symbol is used; since the shallow input feature map is subjected to feature enhancement decomposition along the row and column directions in this embodiment, the bioorthogonal consistency constraint is used. The consistency regularization terms are calculated based on the row and column directions and then added together. The calculation expression for the bioorthogonal consistency constraint is as follows:

[0171]

[0172] in, This is a bioorthogonal consistency constraint; For row direction biorthogonal consistency regularization terms; For column-direction biorthogonal consistency regularization terms.

[0173] Feature-enhanced subbands Confirm target frequency band mask And construct the corresponding stopband region ;

[0174] Based on the spectral energy The target frequency band mask and stopband region are used to calculate the relative stopband response by means of the stopband energy. ;

[0175] Based on the stopband relative response, a stopband leakage suppression constraint is constructed. ; For activation function, To achieve the desired stopband attenuation margin, This represents the number of sub-bands. When the stopband strong response is already below the passband average by a set margin, that sub-band will no longer be subject to additional penalties.

[0176] Based on the low-frequency structural subband and high-frequency detail subband, low-frequency reconstruction components are constructed through reverse reconstruction. and high-frequency reconstruction components , The eigenvalue decomposition function of BiBLD ;

[0177] The low-frequency structure guidance map is calculated based on the mean absolute response of the channel based on the low-frequency reconstruction component and the high-frequency reconstruction component. and high-frequency detail guide chart ,in This is the mean calculation function;

[0178] The low-frequency structure guidance map and low-frequency guidance reliability coefficient are injected into the regression path of the high-resolution detection branch through gated residuals to update the bounding box positioning path; in this embodiment, the high-resolution detection branch used is the P2 detection branch of the YOLOv11s detection framework.

[0179] The high-frequency detail guide map and high-frequency guide reliability coefficient are injected into the classification path of the high-resolution detection branch through a gated residual method to update the target recognition path.

[0180] Based on the feature enhancement sub-band fusion features of the target image to be processed, target detection is performed using bounding box localization paths and target recognition paths, including:

[0181] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the bounding box localization path to obtain the regression path input features after bounding box localization path modulation. The calculation expression is as follows:

[0182]

[0183] Where Xreg is the regression path input feature after bounding box localization path modulation; β is the learnable gating coefficient of the bounding box localization path; The reliability coefficient for low-frequency guidance; To enhance the fusion features of subbands, This is a low-frequency structure guide diagram after the convolution channel is projected.

[0184] Based on the regression path input features modulated by the bounding box localization path, the bounding box localization result is obtained through regression prediction. ; For the target detection network model, the regression prediction function of the detection head

[0185] Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the target recognition path to obtain the target recognition path modulated classification path input features. The calculation expression is as follows:

[0186]

[0187] Where Xcls is the input feature of the classification path after the target recognition path is modulated; α is the learnable gating coefficient of the target recognition path; For high-frequency guidance reliability coefficient; To enhance the fusion features of subbands, This is a guide map of high-frequency details after projection of the convolutional channels;

[0188] like Figure 4 As shown, the low-frequency structure guidance diagram after projection of the convolutional channels. High-frequency detail guide map after convolution channel projection , For the regression path convolution projection function, For classification path convolution projection function; For detecting the size alignment function of branches; Enhance the fusion features of subbands.

[0189] Based on the input features of the classification path after target recognition path modulation, the target recognition result is obtained through classification prediction. ; This is the classification prediction function for the detection head of the object detection network model.

[0190] Basic detection constraints The learningable gating coefficients of the bounding box localization path and the target recognition path are constructed, specifically: the bounding box regression loss function is constructed based on the learningable gating coefficients of the bounding box localization path.

[0191] A classification loss function is constructed based on the learnable gating coefficients of the target recognition path;

[0192] Based on the bounding box regression loss function and classification loss function, the basic detection constraints are constructed, and the calculation expression is as follows:

[0193]

[0194] in, Based on the detection constraints, The bounding box regression loss function is... This is the classification loss function.

[0195] In this embodiment, the YOLOv11s detection framework is used as the basic framework of the target detection network, such as... Figure 1 As shown, the fusion features of the feature enhancement subband are obtained through the high-resolution detection branch of YOLOv11s. The high-resolution detection branch is preferably the P2 detection branch used for weak target detection, while the P3, P4, and P5 detection branches maintain the original detection head structure. Figure 4 The diagram illustrates a band cue utilization mechanism for the high-resolution detection branch.

[0196] The training of the object detection network model includes: based on stopband leakage suppression constraints. Biorthogonal consistency constraints and basic testing constraints An object detection network model is trained using the acquired image dataset. The model is trained using a pre-defined loss function target value. The calculation expression is:

[0197]

[0198] in, These are the weighting coefficients for the bioorthogonal consistency constraint. These are the weighting coefficients for the stopband leakage suppression constraint. The image dataset obtained can be an existing dataset.

[0199] When the target image to be processed is the VisDrone dataset, compared with different types of detectors such as Faster R-CNN, SSD, and SwinTransformer, the target detection network model trained in this embodiment maintains advantages in accuracy (P), recall (R), and mAP metrics. Figure 5As shown, the image weak target detection method in this embodiment can not only improve the detection capability of weak targets, but also help improve the stability of bounding box localization.

[0200] When BiBLD, EFR-SSC, and DFCI are used individually and in combination, the accuracy (P), recall (R), and mAP metrics are as follows: Figure 6 As shown, experimental results when BiBLD, EFR-SSC, and DFCI are used together demonstrate that BiBLD provides a frequency-aware foundation for weak target detection in images, but its performance improvement is limited when used alone. Introducing EFR-SSC significantly improves accuracy P, indicating that this module effectively constrains the frequency response and suppresses redundant noise, thereby enhancing the stability and purity of feature representation. However, its contribution to recall and overall detection accuracy is relatively limited. In contrast, the introduction of DFCI brings a more significant performance improvement. The BiBLD+DFCI combination significantly outperforms the results of using BiBLD alone in terms of recall, mAP@0.5, and mAP@0.5:0.95, indicating that DFCI can effectively utilize the frequency detail information extracted from the front end and convert it into discriminative gain in the detection head. Therefore, DFCI is a key module for effectively utilizing detail information. When all three are used simultaneously, the model achieves the optimal mAP@0.5:0.95, indicating that the complete method has the best overall performance. Overall, in this embodiment, BiBLD, EFR-SSC, and DFCI synergistically improve the performance of the target detection network model from three levels: frequency perception construction, frequency representation optimization, and utilization of detail information.

[0201] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0202] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0203] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0204] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0205] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for detecting weak targets in an image, characterized in that, include: The acquired target image to be processed is input into the target detection network model, and feature decomposition is performed through a learnable bioorthogonal bidirectional lifting method to obtain the fused features of the feature-lifted subbands; Based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path to obtain the target recognition result and the bounding box localization result. The training method for the target detection network model includes: Based on the acquired baseline image to be detected, a shallow input feature map is obtained through shallow feature extraction of the target detection network model; Based on the shallow input feature map, a learnable bioorthogonal bidirectional boosting method is used to perform feature decomposition, resulting in multiple feature boosting subbands; wherein the feature boosting subbands include low-frequency structure subbands and high-frequency detail subbands. Based on the feature enhancement subband, update the bounding box localization path and target recognition path, and construct stopband leakage suppression constraints; Basic detection constraints are constructed based on the bounding box localization path and the target recognition path; Based on the predictive update parameters of the learnable bioorthogonal bidirectional boosting method, a bioorthogonal consistency constraint is constructed. Based on stopband leakage suppression constraints, bioorthogonal consistency constraints, and basic detection constraints, an object detection network model is trained using the acquired image dataset.

2. The image weak target detection method according to claim 1, characterized in that, The target image to be processed is input into the target detection network model, and feature decomposition is performed using a learnable bioorthogonal bidirectional boosting method, including: Based on the acquired target image to be processed, a shallow input feature map of the target image to be processed is obtained through normalization processing and shallow feature extraction. Based on the shallow input feature map of the target image to be processed, feature decomposition is performed using a learnable biorthogonal bidirectional lifting method to obtain the feature-enhanced subband of the target image to be processed. Based on the feature enhancement subbands of the target image to be processed, channel projection is performed using point convolution to obtain the fused features of the feature enhancement subbands.

3. The image weak target detection method according to claim 1, characterized in that, Based on the shallow input feature map, a learnable bioorthogonal bidirectional boosting method is used for feature decomposition to obtain multiple feature boosting sub-bands, including: Based on the shallow input feature map, frequency band feature decomposition is performed along the row direction of the feature map through odd and even positions to obtain the row direction decomposed features; Based on the row direction decomposition features, frequency band feature decomposition is performed along the column direction of the feature map through odd and even positions to obtain multiple feature enhancement sub-bands.

4. The image weak target detection method according to claim 3, characterized in that, Based on the shallow input feature map, frequency band feature decomposition is performed along the row direction of the feature map through odd and even positions to obtain the row direction decomposed features, including: Based on the shallow input feature map, even-sampled features and odd-sampled features are obtained by dividing the feature map into even-sampled features and odd-sampled features according to the odd and even positions of the row directions. Based on the even-sampling features, feature prediction is performed using a prediction operator to obtain the odd-sampling feature prediction result; Based on the odd sampling feature prediction results and odd sampling features, high-frequency detail components are obtained by difference calculation. Based on the high-frequency detail components, the low-frequency structure components are obtained by correcting the dual sampling features using the update operator. Based on the high-frequency detail components and low-frequency structural components, the row direction decomposition features are obtained by adjusting the channel-by-channel scale using a scale factor.

5. The image weak target detection method according to claim 4, characterized in that, Based on the prediction operator, update operator, and scaling factor in the learnable bioorthogonal bidirectional boosting method, bioorthogonal consistency regularization terms are constructed respectively. The calculation expression is:

6. Among them, and These are the weighting coefficients. For reference proportionality, Norm symbol; Let and k be the learnable prediction operator, the learnable update operator, and the learnable channel scaling factor, respectively; 1 and 1 are the preset basic reference values ​​for the learnable prediction operator, the learnable update operator, and the learnable channel scale factor, respectively. Based on the bioorthogonal consistency regularization term, bioorthogonal consistency constraints are constructed, and the calculation expression is as follows:

7. Among them, This is a bioorthogonal consistency constraint; For row direction biorthogonal consistency regularization terms; For column-direction biorthogonal consistency regularization terms.

8. The image weak target detection method according to claim 1, characterized in that, Based on the aforementioned features, subband enhancement is used to construct stopband leakage suppression constraints, including: Based on the aforementioned feature-enhanced subband, the spectral energy corresponding to the feature-enhanced subband is calculated using Fourier transform. The target frequency band mask is confirmed based on the feature-enhanced subband, and the corresponding stopband region is constructed. Based on the spectral energy, target frequency band mask, and stopband region, the stopband relative response is obtained by calculating the average stopband energy. A stopband leakage suppression constraint is constructed based on the aforementioned stopband relative response; The calculation expression for the bandgap leakage suppression constraint is as follows:

9. Among them, To suppress band leakage, For activation function, To determine the desired stopband attenuation margin, For the number of sub-bands, It is a very small constant. This is the stopband relative response.

10. The image weak target detection method according to claim 6, characterized in that, Based on the aforementioned feature enhancement sub-band, the bounding box localization path and target recognition path are updated, including: Based on the low-frequency structural subband and the high-frequency detail subband, low-frequency reconstruction components and high-frequency reconstruction components are constructed through reverse reconstruction, respectively; Based on the low-frequency reconstruction component and the high-frequency reconstruction component, the low-frequency structure guide map and the high-frequency detail guide map are calculated by the mean of the channel absolute response, respectively. Calculate the low-frequency and high-frequency guidance reliability coefficients based on the spectral energy corresponding to the feature-enhanced subbands. The low-frequency structure guidance map and low-frequency guidance reliability coefficient are injected into the regression path of the detection branch through gated residual method to update the bounding box positioning path; The high-frequency detail guide map and high-frequency guide reliability coefficient are injected into the classification path of the detection branch through a gated residual method to update the target recognition path.

11. The image weak target detection method according to claim 7, characterized in that, The low-frequency and high-frequency guidance reliability coefficients are calculated based on the spectral energy corresponding to the feature-enhanced subbands, including: Based on the aforementioned features, the spectral energy corresponding to the sub-band is enhanced, and a probabilistic spectral distribution is obtained through normalization. Based on the probabilistic spectrum distribution, the subband purity corresponding to the subband is improved by calculating features through a preset subband spectrum template; The low-frequency guidance reliability coefficient and the high-frequency guidance reliability coefficient are calculated based on the subband purity using a truncation function.

12. The image weak target detection method according to claim 7, characterized in that, Based on the fusion features of the enhanced sub-bands, target detection is performed using the bounding box localization path and the target recognition path, including: Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the bounding box localization path to obtain the regression path input features after bounding box localization path modulation. The calculation expression is as follows: ； Where Xreg is the regression path input feature after bounding box localization path modulation; β is the learnable gating coefficient of the bounding box localization path; The reliability coefficient for low-frequency guidance; To enhance the fusion features of subbands, This is a low-frequency structure guide diagram after the convolution channel is projected. Based on the regression path input features modulated by the bounding box localization path, the bounding box localization result is obtained through regression prediction. Based on the fusion features of the feature-enhanced subbands, feature modulation is performed through the target recognition path to obtain the target recognition path modulated classification path input features. The calculation expression is as follows: ； Where Xcls is the input feature of the classification path after the target recognition path is modulated; α is the learnable gating coefficient of the target recognition path; For high-frequency guidance reliability coefficient; To enhance the fusion features of subbands, This is a guide map of high-frequency details after projection of the convolutional channels; Based on the input features of the classification path after target recognition path modulation, the target recognition result is obtained through classification prediction.

13. The image weak target detection method according to claim 9, characterized in that, Construct a bounding box regression loss function based on learnable gating coefficients of the bounding box localization path; A classification loss function is constructed based on the learnable gating coefficients of the target recognition path; Based on the bounding box regression loss function and classification loss function, the basic detection constraints are constructed, and the calculation expression is as follows:

14. Among them, Based on the detection constraints, The bounding box regression loss function is... This is the classification loss function.

Images