A small target detection system and method based on a statistical consistency framework
Patent Information
- Application Number
- CN202611202867.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-10-30
AI Technical Summary
然而，某些复杂的背景区域在经过深度网络编码后，可能会产生较高的激活响应，但在统计学意义上，这些区域往往缺乏真实目标应具备的局部离散度和纹理支撑
[0024]经由上述的技术方案可知，与现有技术相比，本发明公开提供了一种基于统计一致性框架的小目标检测系统及方法，首先，通过引入二阶矩保持下采样机制，从物理层面抑制了卷积神经网络固有的低通滤波效应对小目标特征的模糊作用，确保了边界与纹理等关键高频信号在网络深层的存续。其次，多尺度统计矩对齐策略从理论上解决了异构特征融合时的分布失配问题，避免了传统拼接操作引入的统计混叠，使得融合后的特征表达更为稳健且具备一致的物理含义。最后，统计离散门控模块通过建立语义-统计正交方式，为模型提供了独立于语义响应之外的正交辨别维度，在VisDrone2019等复杂场景下大幅降低了误检率，使模型在维持极小参数规模（约3.1M）的同时，实现了优于SOTA模型的检测效能。
Smart Images
Figure CN123016258A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and image processing technology, and more specifically to a small target detection system and method based on a statistical consistency framework. Background Technology
[0002] With the rapid development of unmanned aerial vehicle (UAV) technology and high-resolution satellite remote sensing technology, computer vision-based target detection technology is increasingly being applied in fields such as traffic monitoring, maritime rescue, urban planning, and precision agriculture. Unlike traditional natural scene images, images captured by UAVs typically have extremely high resolution and a wide field of view. This results in the target to be detected (such as pedestrians, vehicles, small boats, etc.) occupying only a very small pixel area (usually less than 16×16 pixels) in the original image. The detection of this type of target is generally referred to as small target detection.
[0003] While general-purpose object detection frameworks, such as the YOLO series, Faster R-CNN, and the Transformer-based DETR series, have achieved great success on objects of normal scale, their performance often drops drastically when processing small targets in UAV imagery. The reason for this is that existing technologies face a severe representation bottleneck when dealing with small targets.
[0004] First, the inherent downsampling mechanism of deep neural networks leads to the irreversible loss of crucial information. To obtain a sufficient receptive field and reduce computational complexity, existing backbone networks typically employ strided convolution or pooling operations for downsampling. From a signal processing perspective, these operations are essentially low-pass filters. For large targets, losing some high-frequency details does not affect their semantic representation. However, for small targets that inherently lack visual cues, their statistically significant features, such as boundaries and textures, are mainly concentrated in the high-frequency band. After multiple layers of downsampling, the energy intensity of small targets decays rapidly, causing their features to become extremely weak in the deeper layers of the network, or even completely submerged in background clutter.
[0005] Secondly, the issue of inconsistent statistical distributions during feature fusion has been neglected for a long time. To compensate for the insufficient resolution of deep feature spaces, existing technologies generally employ Feature Pyramid Networks (FPNs) and their variants (such as PANet and BiFPN) to fuse deep semantic information with shallow spatial information through a top-down path. However, feature maps from different depths exhibit significant distributional deviations in statistical properties (such as mean and variance). Existing fusion methods often employ simple element-wise addition or channel concatenation. This crude fusion approach ignores the statistical differences between features, easily leading to the feature responses of small targets being diluted by heterogeneous background signals, resulting in severe aliasing and consequently causing a large number of missed detections.
[0006] Furthermore, semantic ambiguity in complex backgrounds leads to an extremely high false alarm rate. In UAV imagery, background regions occupy the majority of the image space and contain numerous distracting elements similar to target features, such as tree shadows, building edges, and cluttered paved surfaces. Existing detection mechanisms primarily rely on the activation amplitude of neurons to determine the target's confidence level. However, while some complex background regions may produce high activation responses after deep network encoding, statistically, these regions often lack the local dispersion and texture support that a true target should possess. This detection model, which relies solely on semantic activation and lacks statistical verification, makes it difficult for the model to distinguish between true targets and strongly interfering backgrounds in complex environments, resulting in frequent false alarms.
[0007] In summary, existing small object detection techniques primarily focus on improving performance by increasing network depth, introducing complex attention mechanisms, or sacrificing computational efficiency, but they do not fundamentally address the core challenge of statistical vanishing saliency in small objects during feature evolution. Therefore, maintaining statistical consistency of small objects throughout their evolution from the original image to the deep semantic space, and establishing a detection framework that preserves high-frequency details, eliminates distribution biases, and possesses multi-dimensional verification capabilities, are pressing technical problems in the field of computer vision. Summary of the Invention
[0008] In view of this, the present invention provides a small target detection system and method based on a statistical consistency framework, which aims to maintain the statistical significance of small targets throughout the entire process of evolution from shallow texture at the source to deep semantics, effectively resolving the problems of missed detection caused by feature dissipation and false alarms caused by semantic ambiguity.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: On one hand, the present invention provides a small target detection system based on a statistical consistency framework, comprising: The input module is used to acquire the raw image data to be detected; The detection network is used to input the original image data into a trained statistical consistency target detection network for feature extraction, specifically including: The backbone network is used to acquire multi-scale deep features through the anti-fading feature encoding module; wherein, the anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling, and uses the high-frequency prior information to recalibrate the statistical distribution of the features; The neck network is used to perform cross-layer feature fusion of the multi-scale deep features through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, and fused features with statistical consistency are obtained. The detection head network is used to input the fused features into the statistical discrete gating module, calculate the local statistical discreteness of the fused features in multiple neighborhood ranges as orthogonal verification priors, and generate discreteness gating signals. The output module is used to output the target detection results based on the discrete gating signal.
[0010] Preferably, the anti-fading feature encoding module includes a second-order moment-preserving downsampling unit and a feature statistical distribution recalibration unit; The second-order moment-preserving downsampling unit is configured to: calculate the local second-order central moment through a sliding window mechanism during the downsampling process of expanding the receptive field, extract high-frequency components as high-frequency prior features, and concatenate the high-frequency prior features with the input features in the channel dimension before performing convolutional dimensionality reduction; The feature statistical distribution recalibration unit is configured to: extract local high-frequency priors as structural priors, perform channel-level statistical distribution recalibration on the input features, sharpen the feature response using high-frequency components through adaptive soft gating, and suppress the contamination of the backbone semantic features by invalid high-frequency noise.
[0011] Preferably, the multi-scale statistical moment alignment module includes: A spatial channel alignment unit is used to acquire the baseline features and auxiliary features at the current scale, and to align the auxiliary features in terms of spatial dimension and channel dimension. The statistical moment calculation unit is used to calculate the first-order original moment and the second-order central moment of the reference feature and the aligned auxiliary feature in the channel dimension, respectively. The statistical remapping unit is used to construct a nonparametric statistical remapping mechanism, which uses the first-order original moment and the second-order central moment to map the auxiliary features from the original distribution space to the statistical space of the reference features, thereby obtaining statistically aligned features. The fusion weighting unit is used to concatenate the baseline feature and the statistical alignment feature, generate an attention map through a convolutional network, and use the attention map for weighting and residual fusion to output the fusion feature with statistical consistency.
[0012] Preferably, the statistical discrete gating module includes: The multi-neighborhood discreteness calculation unit is used to calculate the local second-order central moments of the input features in parallel within multiple preset neighborhoods of different sizes, quantify the discreteness of the features within the local regions, and obtain the multi-neighborhood discreteness feature map. The gated mask generation unit is used to jointly encode and concatenate the original input features and the discrete feature maps of different neighborhoods in the channel dimension. It learns the nonlinear mapping relationship between local discreteness and feature confidence through a lightweight convolutional network to generate a discreteness calibration mask. The feature calibration unit is used to perform a dot product weighting on the calibration mask and the original input features, and output the final calibration features through residual connection.
[0013] Preferably, in the second-moment-preserving downsampling unit, the formula for calculating the local second-order central moment is:
[0014] in, Given the input feature tensor, Indicates the kernel size as Average pooling operation, It is the ReLU activation function. This is the bias constant; The calculation formula for concatenating high-frequency prior features and input features along the channel dimension and then performing convolutional dimensionality reduction is as follows:
[0015] in, These are high-frequency prior features.
[0016] Preferably, in the feature statistical distribution recalibration unit, the calculation formula for statistical distribution recalibration of features using high-frequency priors is as follows:
[0017] in, The Hadamard product representing the channel dimension. To calculate the weights of soft gating, To extract local high-frequency components.
[0018] Preferably, in the statistical remapping unit, the formula for mapping auxiliary features from their original distribution space to the statistical space of the baseline features is as follows:
[0019] in, As auxiliary features after alignment, As the baseline feature, and These are the mean and standard deviation along the channel dimension, respectively. It is the minimum bias constant.
[0020] Preferably, in the gated mask generation unit, the calculation formula for generating the discreteness calibration mask is:
[0021] in, As input features, Let be the local second-order central moment within the i-th neighborhood. Represents a lightweight convolution mapping. This represents the Sigmoid activation function; In the feature calibration unit, the calculation formula for the final calibration feature is as follows:
[0022] in, This represents the dot product. This represents a nonlinear transformation.
[0023] On the other hand, the present invention also provides a small target detection method based on a statistical consistency framework, comprising: Obtain the raw image data to be detected; The original image data is input into a trained statistical consistency target detection network for feature extraction. Multi-scale depth features are obtained in the backbone network through an anti-fading feature encoding module. The anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling and uses the high-frequency prior information to recalibrate the statistical distribution of the features. In the neck network, the multi-scale deep features are fused across layers through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, resulting in fused features with statistical consistency. In the detection head network, the fused features are input to the statistical discrete gating module, and the local statistical discreteness of the fused features in multiple neighborhood ranges is calculated as the orthogonal verification prior to generate the discreteness gating signal. The target detection result is output based on the discrete gating signal.
[0024] As can be seen from the above technical solutions, compared with the prior art, this invention discloses a small target detection system and method based on a statistical consistency framework. First, by introducing a second-order moment-preserving downsampling mechanism, the inherent low-pass filtering effect of convolutional neural networks on the blurring of small target features is suppressed at the physical level, ensuring the persistence of key high-frequency signals such as boundaries and textures in the deeper layers of the network. Second, the multi-scale statistical moment alignment strategy theoretically solves the distribution mismatch problem during heterogeneous feature fusion, avoiding the statistical aliasing introduced by traditional splicing operations, making the fused feature representation more robust and possessing consistent physical meaning. Finally, the statistical discrete gating module provides the model with an orthogonal discrimination dimension independent of the semantic response by establishing a semantic-statistical orthogonal approach, significantly reducing the false detection rate in complex scenarios such as VisDrone2019, enabling the model to achieve detection performance superior to state-of-the-art models while maintaining a very small parameter scale (approximately 3.1M). Attached Figure Description
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0026] Figure 1 A diagram of the statistical consistency framework architecture; Figure 2 Architecture diagram of the second-order moment preserving downsampling module; Figure 3 Architecture diagram of the feature statistical distribution recalibration module; Figure 4 This is a schematic diagram illustrating the principle of the multi-scale statistical moment alignment module. Figure 5 This is a schematic diagram of the statistical discrete gating module. Figure 6 Visualize the comparison results on the VisDrone2019 dataset; Figure 7 Visualize the comparison results between the baseline model and the SCF model on the AI-TOD dataset. Detailed Implementation
[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0028] Example 1
[0029] This invention provides a small target detection system based on a statistical consistency framework, such as... Figure 1 As shown, it includes: The input module is used to acquire the raw image data to be detected; The detection network is used to input the original image data into a trained statistical consistency target detection network for feature extraction, specifically including: The backbone network is used to acquire multi-scale deep features through the anti-fading feature encoding module; wherein, the anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling, and uses the high-frequency prior information to recalibrate the statistical distribution of the features; The neck network is used to perform cross-layer feature fusion of the multi-scale deep features through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, and fused features with statistical consistency are obtained. The detection head network is used to input the fused features into the statistical discrete gating module, calculate the local statistical discreteness of the fused features in multiple neighborhood ranges as orthogonal verification priors, and generate discreteness gating signals. The output module is used to output the target detection results based on the discrete gating signal.
[0030] Preferably, the anti-fading feature encoding module includes a second-order moment-preserving downsampling unit and a feature statistical distribution recalibration unit; refer to Figure 2 The second-order moment-preserving downsampling unit is configured to: calculate the local second-order central moment through a sliding window mechanism during the downsampling process of expanding the receptive field, extract high-frequency components as high-frequency prior features, and concatenate the high-frequency prior features with the input features in the channel dimension and then perform convolution dimensionality reduction. refer to Figure 3 The feature statistical distribution recalibration unit is configured to: extract local high-frequency priors as structural priors, perform channel-level statistical distribution recalibration on the input features, sharpen the feature response using high-frequency components through adaptive soft gating, and suppress the contamination of the backbone semantic features by invalid high-frequency noise.
[0031] refer to Figure 4 The multi-scale statistical moment alignment module includes: A spatial channel alignment unit is used to acquire the baseline features and auxiliary features at the current scale, and to align the auxiliary features in terms of spatial dimension and channel dimension. The statistical moment calculation unit is used to calculate the first-order original moment and the second-order central moment of the reference feature and the aligned auxiliary feature in the channel dimension, respectively. The statistical remapping unit is used to construct a nonparametric statistical remapping mechanism, which uses the first-order original moment and the second-order central moment to map the auxiliary features from the original distribution space to the statistical space of the reference features, thereby obtaining statistically aligned features. The fusion weighting unit is used to concatenate the baseline feature and the statistical alignment feature, generate an attention map through a convolutional network, and use the attention map for weighting and residual fusion to output the fusion feature with statistical consistency.
[0032] refer to Figure 5 The statistical discrete gating module includes: The multi-neighborhood discreteness calculation unit is used to calculate the local second-order central moments of the input features in parallel within multiple preset neighborhoods of different sizes, quantify the discreteness of the features within the local regions, and obtain the multi-neighborhood discreteness feature map. The gated mask generation unit is used to jointly encode and concatenate the original input features and the discrete feature maps of different neighborhoods in the channel dimension. It learns the nonlinear mapping relationship between local discreteness and feature confidence through a lightweight convolutional network to generate a discreteness calibration mask. The feature calibration unit is used to perform a dot product weighting on the calibration mask and the original input features, and output the final calibration features through residual connection.
[0033] Furthermore, in the second-order moment preserving downsampling unit, the formula for calculating the local second-order central moment is:
[0034] in, Given the input feature tensor, Indicates the kernel size as Average pooling operation, It is the ReLU activation function. This is the bias constant; The calculation formula for concatenating high-frequency prior features and input features along the channel dimension and then performing convolutional dimensionality reduction is as follows:
[0035] in, These are high-frequency prior features.
[0036] Furthermore, in the feature statistical distribution recalibration unit, the calculation formula for statistical distribution recalibration of features using high-frequency priors is as follows:
[0037] in, The Hadamard product representing the channel dimension. To calculate the weights of soft gating, To extract local high-frequency components.
[0038] Specifically, in the statistical remapping unit, the formula for mapping auxiliary features from their original distribution space to the statistical space of the baseline features is as follows:
[0039] in, As auxiliary features after alignment, As the benchmark feature, and These are the mean and standard deviation along the channel dimension, respectively. It is the minimum bias constant.
[0040] Preferably, in the gated mask generation unit, the calculation formula for generating the discreteness calibration mask is:
[0041] in, As input features, Let be the local second-order central moment within the i-th neighborhood. Represents a lightweight convolution mapping. This represents the Sigmoid activation function; In the feature calibration unit, the calculation formula for the final calibration feature is as follows:
[0042] in, This represents the dot product. This represents a nonlinear transformation.
[0043] Example 2
[0044] The Statistical Consistency Framework (SCF) proposed in this invention aims to solve the performance bottleneck caused by feature decay when deep networks process small objects. Its core is to maintain the statistical saliency of small objects throughout the entire process of evolution from source texture to deep semantics.
[0045] In implementing the statistical consistency-based small target detection method described in this invention, an end-to-end deep convolutional neural network architecture needs to be constructed first. This architecture follows a three-stage design paradigm: backbone network, neck network, and detection head. In terms of hardware, this embodiment is implemented on a server with high-performance computing capabilities, configured with a computing platform featuring dual NVIDIA GeForce RTX 5090 GPUs, 32GB of video memory, and a high-performance central processing unit. The software environment uses PyTorch 2.8.0 and the CUDA 12.8 deep learning framework.
[0046] Reference Figure 1This paper presents the overall architecture of the statistical consistency framework of this invention. The framework consists of three core stages: a decay-resistant feature encoding stage in the backbone network, a multi-scale statistical moment alignment and fusion stage in the neck network, and a statistical discrete gating inference stage in the detection head network. The original image to be detected is first input into the backbone network for multi-scale feature extraction. In each downsampling operation with halved spatial resolution, the decay-resistant feature encoding mechanism retains high-frequency prior information through the second-moment preserving downsampling unit, and performs prior-guided distribution correction of semantic features using a feature statistical distribution recalibration unit at a deeper level. After the multi-scale feature map output from the backbone network enters the neck network, the multi-scale statistical moment alignment module performs statistical distribution alignment on the feature maps from different depths, eliminating statistical biases between layers and completing the fusion. The fused features are finally passed to the detection head, where the statistical discrete gating module uses local statistical discreteness to generate a gating signal for secondary correction of the features. The detection head outputs the category prediction and bounding box coordinates.
[0047] In the specific construction of the backbone network, this invention implements an anti-fading feature encoding mechanism, the core motivation of which is to correct the representation bottleneck in the evolution of deep networks. Let the original input image to be detected be... The object detection task aims to learn a parameterized mapping function. This maps the image domain to a target space containing category and location information. In deep convolutional neural networks, this process is achieved through hierarchical feature extraction. Let... Indicates the first The feature map of the layer, where The evolution of features can be modeled as a nonlinear transformation as shown in the following equation:
[0048] In the above formula, Represents a nonlinear transformation, while Indicates step size is The downsampling operation. Although this mechanism can extract abstract semantic information, for small targets with extremely limited spatial support (e.g., less than 32×32 pixels), successive downsampling leads to a rapid decay in their statistical significance.
[0049] To address this issue, the anti-fading feature encoding mechanism first implements a second-order moment-preserving downsampling (SoMPD) unit. In standard downsampling strategies, high-frequency components in the spectrum exceeding the new Nyquist frequency undergo irreversible aliasing or disappearance. However, key features for the localization and classification of small targets (such as edges and textures) are primarily found in high-frequency signals. From a statistical signal processing perspective, these features correspond to high-variance regions in the statistical distribution. In practice, the SoMPD unit, for the input feature tensor... Firstly, in the local neighborhood Extracting regions of significant change. This process is defined using the generalized second-order statistical operator as shown in the following equation:
[0050] in, Indicates the kernel size as Average pooling operation, It is the ReLU activation function. For a very small bias constant (e.g.) To ensure the stability of numerical computation, this operator accurately captures the local second-order central moments, i.e., local variance, by calculating the difference between the mean of the feature squares and the square of the feature mean. This allows for the extraction of high-frequency prior features. It contains rich boundary information. Then, Input features The concatenation is performed along the channel dimension, and spatial compression and dimension alignment are performed using parameterized convolutional layers, as shown in the mathematical definition:
[0051] By introducing second-order statistical moments, the SoMPD unit expands the receptive field while actively preserving high variance information, thereby maintaining the statistical significance of small targets during the physical space compression process.
[0052] As the backbone network evolves further into deeper layers, due to the extremely low proportion of small target pixels, deep networks often exhibit a strong semantic induction bias towards large-scale backgrounds. This implementation addresses this issue through a Feature Statistical Distribution Recalibration (FSDR) unit. This unit explicitly decouples high-frequency information from the semantic backbone and uses it as a structural prior to reconstruct the feature distribution. In specific implementation, for input features... The local high-frequency components were extracted again using statistical computing. ,Right now Subsequently, this prior information is used to recalibrate the statistical distribution of the features, as shown in the following mathematical formula:
[0053] in, This represents the Hadamard product of the channel dimension, while As a statistical soft-gated weight, it adaptively optimizes during training. For channels containing the target texture, The value tends towards positive, utilizing high-frequency components to sharpen the originally blurred feature response; however, for channels dominated by a stable background or pure noise, It converges to zero, ensuring that invalid high-frequency noise does not contaminate the backbone semantic features. The FSDR module ensures that small targets with weak semantic responses can be effectively propagated to deeper layers of the network through explicit distribution calibration.
[0054] After feature extraction, multi-scale feature maps are fused into the neck network. Traditional multi-scale linear fusion (such as addition or concatenation) often ignores the statistical distribution offset between layers. This invention implements a multi-scale statistical moment alignment (MsSMA) fusion strategy, the core of which lies in eliminating the distribution bias between layers through nonparametric mapping. Let the baseline feature at the current scale be... Auxiliary features that provide multi-scale information are Before the actual fusion, the spatial and channel dimensions must first be aligned:
[0055] in, Used for channel alignment. and These represent adaptive average pooling and bilinear interpolation, respectively. After dimension alignment, this invention utilizes global statistical moments for measurement. and The statistical bias between them. Specifically, the mean of the auxiliary features along the channel dimension is calculated. with standard deviation And it is projected onto the statistical space of the baseline features using the following nonparametric statistical remapping mechanism:
[0056] Features after alignment Compared with the benchmark features Channel splicing is performed, and the data is fed into a lightweight convolutional network to generate a spatially saliency attention map. The final fusion features are obtained by significance-weighted summation and residual connection, as shown in the following equation:
[0057] This strategy achieves a balance between statistical consistency and content complementarity, significantly reducing aliasing during heterogeneous feature fusion.
[0058] Finally, in the detection head section, to address the issue that background clutter in UAV imagery can easily trigger high activation responses, leading to false detections, this invention implements a Statistical Discrete Gated Module (SDGM). This module utilizes local statistical dispersion as the orthogonal validation dimension. In implementation, for the input features... Parallel computation of multiple preset neighborhood ranges The local second-order central moments are used to obtain a set of multi-neighborhood discrete feature maps. As shown in the following formula:
[0059] In small target regions, due to drastic changes in structural edges, pixel values deviate significantly from the mean than from the background. SDGM jointly encodes the original features with these discrete feature maps to learn and generate a discrete-gated signal mask. :
[0060] in, express Activation function This represents a lightweight convolutional mapping. The final calibrated features are output through the dot product residual operation between this gated mask and the features, as shown in the following equation:
[0061] This process establishes a dual-reset reliability of semantic discriminative power and statistical consistency, effectively filtering false alarm noise lacking texture support without significantly increasing computational burden.
[0062] During the model training phase, this embodiment employs random initialization and trains from scratch. The input image resolution is set to 640×640 pixels, the SGD optimizer is used, the initial learning rate is 0.01, the momentum is set to 0.937, and the weight decay is... The training process includes a 3-epoch warmup phase, coupled with a linear learning rate decay strategy, for a total of 500 training iterations. During inference and prediction, the model outputs calibrated features through the SDGM, and the decoupled detection head outputs the final class prediction probability and bounding box coordinates, achieving high-precision statistically consistent object detection. (See Table 1.) Figure 6 , Figure 7 As shown in the table, the best results are displayed in bold. Experiments show that the accuracy is significantly improved when dealing with extremely small targets (such as the AI-TOD dataset).
[0063] Table 1. Comparison with state-of-the-art methods on VisDrone 2019.
[0064]
[0065] On the other hand, the present invention also provides a small target detection method based on a statistical consistency framework, comprising: Obtain the raw image data to be detected; The original image data is input into a trained statistical consistency target detection network for feature extraction. Multi-scale depth features are obtained in the backbone network through an anti-fading feature encoding module. The anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling and uses the high-frequency prior information to recalibrate the statistical distribution of the features. In the neck network, the multi-scale deep features are fused across layers through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, resulting in fused features with statistical consistency. In the detection head network, the fused features are input to the statistical discrete gating module, and the local statistical discreteness of the fused features in multiple neighborhood ranges is calculated as the orthogonal verification prior to generate the discreteness gating signal. The target detection result is output based on the discrete gating signal.
[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the relevant parts are described in the method section.
[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A small target detection system based on a statistical consistency framework, characterized in that, include: The input module is used to acquire the raw image data to be detected; The detection network is used to input the original image data into a trained statistical consistency target detection network for feature extraction, specifically including: The backbone network is used to acquire multi-scale deep features through the anti-fading feature encoding module; wherein, the anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling, and uses the high-frequency prior information to recalibrate the statistical distribution of the features; The neck network is used to perform cross-layer feature fusion of the multi-scale deep features through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, and fused features with statistical consistency are obtained. The detection head network is used to input the fused features into the statistical discrete gating module, calculate the local statistical discreteness of the fused features in multiple neighborhood ranges as orthogonal verification priors, and generate discreteness gating signals. The output module is used to output the target detection results based on the discrete gating signal.
2. The small target detection system based on a statistical consistency framework according to claim 1, characterized in that, The anti-attenuation feature encoding module includes a second-order moment-preserving downsampling unit and a feature statistical distribution recalibration unit; The second-order moment-preserving downsampling unit is configured to: calculate the local second-order central moment through a sliding window mechanism during the downsampling process of expanding the receptive field, extract high-frequency components as high-frequency prior features, and concatenate the high-frequency prior features with the input features in the channel dimension before performing convolutional dimensionality reduction; The feature statistical distribution recalibration unit is configured to: extract local high-frequency priors as structural priors, perform channel-level statistical distribution recalibration on the input features, sharpen the feature response using high-frequency components through adaptive soft gating, and suppress the contamination of the backbone semantic features by invalid high-frequency noise.
3. The small target detection system based on a statistical consistency framework according to claim 1, characterized in that, The multi-scale statistical moment alignment module includes: A spatial channel alignment unit is used to acquire the baseline features and auxiliary features at the current scale, and to align the auxiliary features in terms of spatial dimension and channel dimension. The statistical moment calculation unit is used to calculate the first-order original moment and the second-order central moment of the reference feature and the aligned auxiliary feature in the channel dimension, respectively; The statistical remapping unit is used to construct a nonparametric statistical remapping mechanism, which uses the first-order original moment and the second-order central moment to map the auxiliary features from the original distribution space to the statistical space of the reference features, thereby obtaining statistically aligned features. The fusion weighting unit is used to concatenate the baseline feature and the statistical alignment feature, generate an attention map through a convolutional network, and use the attention map for weighting and residual fusion to output the fusion feature with statistical consistency.
4. The small target detection system based on a statistical consistency framework according to claim 1, characterized in that, The statistical discrete gating module includes: The multi-neighborhood discreteness calculation unit is used to calculate the local second-order central moments of the input features in parallel within multiple preset neighborhoods of different sizes, quantify the discreteness of the features within the local regions, and obtain the multi-neighborhood discreteness feature map. The gated mask generation unit is used to jointly encode and concatenate the original input features and the discrete feature maps of different neighborhoods in the channel dimension. It learns the nonlinear mapping relationship between local discreteness and feature confidence through a lightweight convolutional network to generate a discreteness calibration mask. The feature calibration unit is used to perform a dot product weighting on the calibration mask and the original input features, and output the final calibration features through residual connection.
5. A small target detection system based on a statistical consistency framework according to claim 2, characterized in that, In the second-order moment-preserving downsampling unit, the formula for calculating the local second-order central moment is: in, Given the input feature tensor, Indicates the kernel size as Average pooling operation, It is the ReLU activation function. This is the bias constant; The calculation formula for concatenating high-frequency prior features and input features along the channel dimension and then performing convolutional dimensionality reduction is as follows: in, These are high-frequency prior features.
6. A small target detection system based on a statistical consistency framework according to claim 2, characterized in that, In the feature statistical distribution recalibration unit, the calculation formula for statistical distribution recalibration of features using high-frequency priors is as follows: in, The Hadamard product representing the channel dimension. To calculate the weights of soft gating, To extract local high-frequency components.
7. A small target detection system based on a statistical consistency framework according to claim 3, characterized in that, In the statistical remapping unit, the formula for calculating the mapping of auxiliary features from their original distribution space to the statistical space of the baseline features is as follows: in, As auxiliary features after alignment, As the benchmark feature, and These are the mean and standard deviation along the channel dimension, respectively. It is the minimum bias constant.
8. A small target detection system based on a statistical consistency framework according to claim 4, characterized in that, In the gated mask generation unit, the calculation formula for generating the discreteness calibration mask is as follows: in, For input features, Let be the local second-order central moment within the i-th neighborhood. Represents a lightweight convolution mapping. This represents the Sigmoid activation function; In the feature calibration unit, the calculation formula for the final calibration feature is as follows: in, This represents the dot product. This represents a nonlinear transformation.
9. A small target detection method based on a statistical consistency framework, characterized in that, include: Obtain the raw image data to be detected; The original image data is input into a trained statistical consistency target detection network for feature extraction. Multi-scale depth features are obtained in the backbone network through an anti-fading feature encoding module. The anti-fading feature encoding module explicitly extracts and retains local high-frequency prior information during downsampling and uses the high-frequency prior information to recalibrate the statistical distribution of the features. In the neck network, the multi-scale deep features are fused across layers through a multi-scale statistical moment alignment module; wherein, by calculating and aligning the global statistical moments of features at different levels, statistical distribution bias in the cross-layer feature fusion process is eliminated, resulting in fused features with statistical consistency. In the detection head network, the fused features are input to the statistical discrete gating module, and the local statistical discreteness of the fused features in multiple neighborhood ranges is calculated as the orthogonal verification prior to generate the discreteness gating signal. The target detection result is output based on the discrete gating signal.