Small target detection method and system based on high-frequency information learning

Abstract

A small target detection method and system based on high-frequency information learning. The method comprises: acquiring an image and extracting multi-scale features, and selecting shallow features as a basis; in the encoder stage, a density-guided adaptive sampling attention module is constructed to increase the sampling rate in the target dense area to actively retain high-frequency details; in the decoder stage, an anti-over-smoothing contrast learning mechanism is introduced, a teacher queue is constructed using high-quality samples, and discriminative features smoothed by hierarchical stacking are forced to be restored through a contrast loss; finally, the results are optimized and output. The present application constructs a signal processing closed loop of "active fidelity" and "active recovery", effectively solves the problems of small target high-frequency signal being overwhelmed by noise and feature over-smoothing, and significantly improves the detection accuracy.

Description

Technical Field

[0001] This invention relates to a small target detection method and system based on high-frequency information learning, belonging to the fields of computer vision and deep learning technology, specifically to a detection method that uses a signal processing perspective to solve the problem of high-frequency signal degradation of small targets in the Transformer architecture. Background Technology

[0002] Small object detection is a key task in computer vision, widely used in fields such as autonomous driving, remote sensing image analysis, and drone monitoring. Unlike conventional object detection, the success of small object detection relies heavily on the capture and recognition of extremely weak signals, especially high-frequency information such as the object's edges and textures.

[0003] However, small target detection faces significant challenges in complex scenarios. First, natural or artificial backgrounds are often filled with high-frequency noise, which can easily obscure the high-frequency signals of small targets. Existing multi-scale fusion methods (such as FPN) attempt to preserve high-frequency information using shallow features, but they often fail to effectively separate signals from noise, resulting in the amplification of background noise.

[0004] More importantly, as Transformer-based detectors become mainstream, their core self-attention mechanism physically behaves as a low-pass filter. This mechanism aggregates information by weighted averaging of a large number of tokens. This smoothing operation inherently suppresses the crucial high-frequency components needed to distinguish small objects, tending to preserve low-frequency global information. When multiple attention layers are stacked, this effect leads to "over-smoothing" of features, making the feature map blurry and difficult to focus on small objects. While existing improvements have optimized query design or label assignment, they have not fundamentally solved the problem of high-frequency signal degradation.

[0005] Therefore, the key to improving small target detection performance lies in how to learn and retain key high-frequency target signals from noisy backgrounds within a unified framework, while also combating the inherent low-pass filtering effect of the Transformer. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings and deficiencies of existing small target detection methods, such as the easy submersion of high-frequency key signals (edges and textures) by background noise and the over-smoothing of features due to the inherent low-pass filtering characteristics of the Transformer architecture. This invention proposes a small target detection method and system based on high-frequency information learning.

[0007] This invention, from the perspective of signal processing, constructs a closed-loop processing mechanism of "active fidelity preservation" and "active recovery":

[0008] On the one hand, to address the signal overload problem in the encoder stage, this invention introduces a density-guided adaptive sampling attention module. This module is implemented based on Shannon's sampling theorem, which states that a higher sampling rate is necessary to capture rapidly changing high-frequency signals with high fidelity. Therefore, this invention does not process image features uniformly, but rather predicts the scene's density distribution and adaptively increases the sampling frequency in densely populated target areas and decreases the sampling frequency in background areas. This non-uniform sampling strategy actively separates signals from noise at the source of feature extraction, effectively preserving detailed information about small, weak targets.

[0009] On the other hand, to address the feature ambiguity problem in the decoder stage, this invention introduces an anti-oversmoothing contrastive learning mechanism. This mechanism aims to counteract the signal smoothing effect caused by the stacking of Transformer layers, transforming the decoder from a passive information aggregator into an active feature restorer. Specifically, this mechanism selects sample features with high confidence and high localization accuracy as "teacher signals," and through supervised contrastive loss, forces features that have lost discriminative power due to oversmoothing to converge towards high-fidelity features in the vector space. This process is mathematically equivalent to reversing the low-pass filtering process, thereby reconstructing clear and discriminative high-frequency target features in the deep network.

[0010] Through the synergistic effect of the two core modules mentioned above, this invention systematically solves the problem of "high-frequency degradation" of small target features during transmission in deep neural networks, thereby significantly improving the performance of small target detection and achieving better detection results.

[0011] The present invention is implemented using the following specific technical solutions.

[0012] A small target detection method based on high-frequency information learning includes the following steps:

[0013] Step 1: Acquire the image to be detected and extract multi-scale feature maps through the backbone network.

[0014] Specifically, the input image is fed into a convolutional neural network (such as HGNetv2) or other backbone network to extract feature maps at different levels. The extraction process can be represented as follows:

[0015]

[0016] in, This represents the input image to be detected. This represents the extracted multi-scale feature map. Shallow features containing rich high-frequency information are selected as the basis for subsequent processing.

[0017] Step 2: In the encoder stage, a density-guided adaptive sampling attention module is built to perform hybrid sampling and aggregation of feature maps.

[0018] Specifically, to prevent high-frequency details from being suppressed during the encoding process, Shannon's sampling theorem is used to guide feature sampling. By increasing the sampling rate in dense target regions, the problem of high-frequency signals being interfered with by noise in shallow features is solved, actively preserving and enhancing the detailed features of small foreground targets. The detailed steps include the following:

[0019] Step 2.1: Generate density map.

[0020] A lightweight density prediction network is introduced to learn the target distribution of the scene from the input features and generate a density map that reflects the density of small targets, providing spatial guidance for non-uniform sampling.

[0021] Step 2.2: Implement the hybrid sampling strategy.

[0022] The design employs a dual-path sampling mechanism: firstly, global sampling, which uses query features to predict the global offset and capture contextual information. The global sampling point set... The calculation is as follows:

[0023]

[0024] in, As the original reference point, This is the predicted global offset. This represents the total number of global sampling points. This is the global sampling point index.

[0025] Second, density-guided sampling generates dense reference points in high-response regions of the density map and dynamically modulates them using local offsets and density values. (Density-guided sampling point set) The calculation is as follows:

[0026]

[0027] in, For dense reference points generated based on density maps, The density value at the reference point. For local fine offset, This represents the number of dense sampling points. Density value. It plays a modulation role: if the reference point is located in the foreground (density close to 1), the offset is fully activated to capture details; if it is located in the background (density close to 0), the sampling range is suppressed to filter out high-frequency noise.

[0028] Step 2.3: Gated feature aggregation.

[0029] By introducing a gating mechanism and a grouped Softmax strategy, the gating scalar is dynamically predicted based on the query features, and the fusion ratio of local high-frequency features and global context features is adaptively adjusted to output encoder features with enhanced signal-to-noise ratio.

[0030] Step 3: In the decoder stage, features are processed by the Transformer decoder, and an anti-oversmoothing contrastive learning mechanism is introduced for training optimization.

[0031] Specifically, to address the signal smoothing problem caused by multi-layer attention stacking, the decoder is transformed from a passive information aggregator into an active feature restorer. Discriminative learning is used to force the recovery of discriminative high-frequency features. This includes the following detailed steps:

[0032] Step 3.1 Define sample quality and dynamic partitioning.

[0033] A quality score (Qscore) based on prediction results is established, comprehensively considering classification confidence and location intersection-over-union (IoU). Quality Score The definition is as follows:

[0034]

[0035] in, For the prediction results Classification confidence, For prediction boxes With real frame The intersection and union ratio.

[0036] By setting a dynamic threshold, the predicted samples are divided in real time into high-quality samples (HQ) that retain high-frequency discrimination information and low-quality samples (LQ) that have undergone oversmoothing.

[0037] Step 3.2 Construct a high-fidelity teacher feature queue.

[0038] Maintain a dynamic feature queue to store recent high-quality sample feature embeddings as "teacher signals" or positive sample sets rich in high-frequency information.

[0039] Step 3.3 Comparative recovery learning.

[0040] By employing a supervised contrastive loss function, low-quality samples with blurred features due to oversmoothing are forced to move closer to high-quality samples in the feature space, while distancing themselves from samples of other classes. The resulting gradient, generated by minimizing this loss, is backpropagated to the decoder, forcing the network to learn how to recover sharp, discriminative high-frequency features.

[0041] Step 4: Optimize the network parameters by combining the total loss function, and output the final small target detection results.

[0042] During the training phase, the model is optimized by combining conventional classification and regression losses with the contrastive recovery loss from step 3. During the inference phase, the trained model directly outputs the target's category and bounding box.

[0043] By integrating the above steps, an end-to-end detection network is constructed, comprising a hybrid encoder, decoder, and prediction head. The encoder is responsible for "actively preserving fidelity" during the feature extraction stage, while the decoder is responsible for "actively recovering" during the feature reconstruction stage. Working together, they significantly improve the detection accuracy of small targets.

[0044] Based on the above method, this invention further proposes a small target detection system based on high-frequency information learning, including a feature extraction subsystem, a density-guided coding subsystem, a feature decoding and recovery subsystem, and a prediction and optimization subsystem.

[0045] The connections between the above components are as follows:

[0046] The output of the feature extraction subsystem is connected to the input of the density-guided coding subsystem; the output of the density-guided coding subsystem is connected to the input of the feature decoding and recovery subsystem; and the output of the feature decoding and recovery subsystem is connected to the input of the prediction and optimization subsystem.

[0047] Beneficial effects

[0048] Compared with the prior art, the present invention has the following advantages:

[0049] 1. This invention utilizes Shannon's sampling theorem at the encoder end to actively increase the sampling rate in dense target regions through density-guided adaptive sampling. This method effectively separates and enhances key high-frequency signals (such as edges and textures) required for small target detection from high-frequency noise in complex backgrounds.

[0050] 2. To address the low-pass filtering characteristics of the Transformer, this invention designs an anti-oversmoothing contrastive learning mechanism at the decoder. Unlike passive feature aggregation, this mechanism uses high-quality predictions as supervision signals, forcing the network to recover the smoothed-out discriminative features through contrastive learning, significantly improving feature fidelity.

[0051] 3. A complete closed-loop protection and recovery system for high-frequency information is constructed: This invention breaks through the limitations of traditional methods that only perform feature aggregation in the spatial domain. It establishes a complete signal processing closed loop at the system level, from "active sampling for fidelity" at the encoder end to "active contrast recovery" at the decoder end. This design not only theoretically solves the physical problem of the natural attenuation of high-frequency information as the network depth increases, but also ensures the integrity of key discriminative features (such as edges and textures) of small targets in the deep feature space, providing a systematic solution to the fundamental problem of "feature vanishing" in small targets in deep learning. Attached Figure Description

[0052] Figure 1 This is the overall flowchart of the method of the present invention.

[0053] Figure 2 This is a schematic diagram of the core algorithm model described in the method of this invention.

[0054] Figure 3 This is a block diagram of the components of the system of the present invention. Detailed Implementation

[0055] To better illustrate the purpose and advantages of the present invention, the method of the invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0056] Example

[0057] like Figure 1 The flowchart shown below illustrates the small target detection method based on high-frequency information learning, which includes the following steps:

[0058] Step S1: Input image, input the image to be detected. .

[0059] Step S2: Multi-scale feature extraction.

[0060] Multi-scale feature maps are extracted using a convolutional neural network backbone (such as HGNetv2). For small object detection tasks, this embodiment focuses on selecting shallow feature maps containing rich high-frequency spatial details such as edges and textures. (For example The feature layer serves as the main input for subsequent processing to avoid the loss of geometric information of small targets in deep features.

[0061] Step S3: Density-Guided Adaptive Sampling in the Encoder Stage. To preserve critical high-frequency details and suppress background noise during encoding, a density-guided adaptive sampling module is introduced in the Transformer encoder layer. This module employs a hybrid sampling strategy, implemented as follows:

[0062] Step S3.1: Density Map Prediction and Sampling Point Generation. First, a lightweight density prediction network is used based on the input features. Density map of the predicted scene Based on this density map, a set of dense reference points is generated through sampling operations. :

[0063]

[0064] in This is a preset number of dense sampling points, which are mainly distributed in the predicted dense area of ​​small targets. This is an index for dense sampling points.

[0065] Step S3.2: Dual-path offset prediction. For each query on the feature map... Simultaneously execute two parallel offset prediction paths:

[0066] Global sampling path: setting A global sampling point is used to capture contextual information. This is achieved through a linear layer. Predict global offset Acting on the original reference point :

[0067]

[0068] in This is the global set of sampling points.

[0069] Density-guided sampling path: for the dense reference points generated in step 2.1 Through linear layer Predicting local fine offsets To enhance focus on high-density areas, density values ​​at a reference point are introduced. Weighted modulation of the offset:

[0070]

[0071] in To guide the sampling point set by density, use density values The sampling range is dynamically adjusted to ensure more precise high-frequency signal capture in densely populated target areas.

[0072] Step S3.3: Gated Hybrid Sampling Feature Aggregation. To adaptively balance global contextual information with local high-frequency details, a gating mechanism and a grouped softmax strategy are introduced for feature fusion. First, a scalar is predicted through an independent gating network. Used to control the activation intensity of density-guided branches:

[0073]

[0074] Next, the attention weights for global sampling points and density-guided sampling points are calculated separately. To avoid an overly smooth weight distribution, grouped softmax is used.

[0075] Global attention weights:

[0076] Density guides attention weights:

[0077] Finally, the two sets of sampling points were combined. and the corresponding weights Combine, calculate the weighted sum and pass it through the output projection layer Enhanced features :

[0078]

[0079] This step allows the encoder to significantly enhance the high-frequency detail features of small targets while preserving global semantics.

[0080] Step S4: Feature decoding and target prediction.

[0081] Enhanced features The input is processed by the Transformer decoder through layers of self-attention and cross-attention, and the prediction result is output by the prediction head. Includes category scores and bounding box .

[0082] Step S5: Anti-oversmoothing contrastive learning in the decoder stage. After the features are processed by the multi-layer Transformer decoder, an anti-oversmoothing contrastive learning mechanism is introduced for training and optimization to address the high-frequency feature smoothing problem caused by the self-attention low-pass filtering characteristics.

[0083] Step S5.1: Sample quality assessment. For each prediction output by the decoder... The quality score is calculated by combining its classification confidence and positioning accuracy. :

[0084]

[0085] in It is the confidence level when the predicted category matches the true category. It is a prediction box With real frame The intersection and union ratio.

[0086] Step S5.2: Dynamic Sample Partitioning and Feature Queue Maintenance. Set dynamic quality thresholds for each category. .like If so, the sample is determined to have retained good discriminative features. Its features are then embedded... As a "teacher signal", it is stored in the corresponding category. Dynamic feature queue In, as the positive sample set .like If the sample features are oversmoothed, it is determined that the sample features have been oversmoothed and is marked as such. Feature recovery is required. Dynamic queue The First-In-First-Out (FIFO) strategy is used for updates to ensure that the latest high-fidelity features are stored.

[0087] Step S5.3: Calculate the contrastive recovery loss. Using the concept of supervised contrastive learning, construct the loss function. The aim is to narrow down the characteristics of low-quality samples. High-quality features of the same type The distance, while pushing away from other categories of features. (from) Distance:

[0088]

[0089] in For cosine similarity, For temperature hyperparameters, ) is based on the natural constant An exponential function with base 0 is used to map similarity scores to positive weight values, thereby achieving... The probability distribution of positive samples is calculated in a formal manner. This loss is backpropagated through gradients, forcing the decoder weights to be updated, thereby learning how to recover sharp, discriminative high-frequency features from blurred features.

[0090] Step S6: Comprehensive Optimization and Output. Construct the total loss function. :

[0091]

[0092] This includes conventional detection losses, such as This represents the classification loss, used to supervise the accuracy of the predicted box's category. It improves the network's ability to distinguish between the target and the background by penalizing incorrectly classified predictions. This represents the bounding box regression loss, used to directly supervise the offset between the predicted box and the ground truth box coordinates, ensuring the numerical accuracy of the localization results, as well as the contrast recovery loss mentioned above. Minimize during training phase Optimize network parameters. During the inference phase, directly utilize the pre-trained network to output the target category and bounding box, eliminating the need to calculate contrastive loss.

[0093] Figure 2 This is a diagram of the core algorithm model structure. It shows the complete data flow from input features to the dual-path sampling and aggregation of the density-guided adaptive sampling attention module, then to the decoder output, and finally to feedback learning through an anti-oversmoothing contrastive learning mechanism. These two parts work together, from signal preservation at the encoding end to signal recovery at the decoding end, systematically solving the problem of missing high-frequency information with discriminative power in small target detection.

[0094] Figure 3 This is a block diagram of the system components of the present invention. The present invention provides a small target detection system based on high-frequency information learning. The system consists of five main subsystems: a feature extraction subsystem, a density-guided encoding subsystem, a feature decoding and recovery subsystem, a prediction and optimization subsystem, and a target detection prediction subsystem. Specifically, the feature extraction subsystem extracts multi-scale visual features from the original image, providing a basic feature map for subsequent processing. The density-guided encoding subsystem includes a density prediction module and an adaptive sampling module. The density prediction module generates a scene density map; the adaptive sampling module, based on the density distribution, performs dense sampling and enhancement of high-frequency information regions while preserving the global context, outputting high-fidelity encoded features. The feature decoding and recovery subsystem includes a Transformer decoding module and a contrastive learning module. The decoding module generates preliminary prediction results; the contrastive learning module counteracts feature smoothing and restores the discriminative power of features by maintaining a dynamic feature queue and calculating contrastive loss. The prediction and optimization subsystem calculates classification, regression, and contrastive losses, updates network parameters, and outputs the final detection results.

[0095] The connections between the subsystems are as follows: the output of the data preprocessing subsystem is connected to the input of the feature extraction subsystem, used to input the preprocessed image into the backbone network for basic feature calculation; the output of the feature extraction subsystem is connected to the input of the density-guided coding subsystem, passing the extracted multi-scale feature maps to the encoding stage; the output of the density-guided coding subsystem is connected to the input of the feature decoding and reconstruction subsystem, providing high-fidelity encoded features enhanced by density-guided sampling; the output of the feature decoding and reconstruction subsystem is connected to the input of the object detection and prediction subsystem, outputting preliminary detection and prediction results containing category and location information; subsequently, the output of the object detection and prediction subsystem is connected to the input of the prediction and optimization subsystem, used to calculate the classification loss. Bounding box regression loss and comparative recovery loss Finally, the output of the prediction and optimization subsystem is reconnected to the feature decoding and recovery subsystem through a feedback loop. By comparing the gradient generated by the recovery loss, backpropagation is performed to dynamically update the decoder parameters and maintain a high-quality feature queue, thereby achieving active recovery of feature discriminativeness.

[0096] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A small target detection method based on high-frequency information learning, characterized in that, Includes the following steps: Step 1: Obtain the image to be detected, extract multi-scale feature maps through the backbone network, and select shallow features containing rich high-frequency information as basic features; Step 2: In the encoder stage, a density-guided adaptive sampling attention module is constructed to generate a scene density map and perform hybrid sampling and aggregation to actively preserve the high-frequency details of the dense target region. Step 3: In the decoder stage, features are processed by the Transformer decoder, and an anti-oversmoothing contrastive learning mechanism is introduced to establish a dynamic feature queue. Contrastive learning is used to force the recovery of discriminative high-frequency features that have been smoothed due to layer stacking. Step 4: Calculate the total loss function to optimize the network parameters, and output the category and bounding box of the small target based on the trained model to complete the detection.

2. The method as described in claim 1, characterized in that, The density-guided adaptive sampling attention module in step 2 includes: a density map generation unit: using a density prediction network to learn the distribution of scene targets from input features and generate a density map reflecting the density of small targets; a dual-path sampling mechanism: including a global sampling path and a density-guided sampling path. The global sampling path uses query features to predict global offsets to capture contextual information, while the density-guided sampling path generates dense reference points in high-response regions of the density map and dynamically modulates them by combining local offsets and density values; and gated feature aggregation: predicting gated scalars through a gated network, calculating attention weights by combining a grouped Softmax strategy, and adaptively adjusting the fusion ratio of local high-frequency features and global contextual features.

3. The method of claim 2, wherein, In the density-guided sampling path, for the generated dense reference points The calculation formula of the sampling point set of the dense reference points is: wherein, is the number of dense sampling points, is the dense reference point, is the predicted local fine offset, is the density value at the reference point used to weight modulate the offset to control the sampling range.

4. The method of claim 2, wherein, In the gating feature aggregation, the enhanced features The calculation formula is: wherein, is the output projection layer, is the total set of sampling points combining the global set of sampling points and the density-guided set of sampling points , are the corresponding attention weights computed by a grouped Softmax, wherein the weights of the density-guided part are controlled by a gating scalar .

5. The method of claim 1, wherein, The anti-oversmoothing contrastive learning mechanism in step 3 includes: Sample quality assessment and dynamic partitioning: A quality score is calculated by combining the classification confidence and location intersection-union ratio of the prediction results. Based on a dynamic threshold, the predicted samples are divided into high-quality samples that retain high-frequency discriminative information and low-quality samples that have undergone oversmoothing; High-fidelity teacher feature queue: A dynamic feature queue is maintained for each category, storing the feature embeddings of high-quality samples as teacher signals, and updated using a first-in-first-out strategy; Contrastive recovery learning: A supervised contrastive loss function is used to narrow the distance between low-quality sample features and high-quality features of the same category, while widening the distance between features of other categories.

6. The method of claim 5, wherein, The mass fraction The calculation formula is: in, For the predicted results, To predict the classification confidence when a category matches the true category, The intersection-union ratio (IU) of the predicted bounding box and the ground truth bounding box; if Greater than or equal to dynamic threshold If the sample is high quality, it is considered a high-quality sample.

7. The method of claim 5, wherein, The loss function of the contrastive recovery learning is: in, Features of low-quality samples A high-quality feature set of the same type, For feature sets of other categories, Represents cosine similarity. This refers to temperature hyperparameters.

8. The method as described in claim 1, characterized in that, The total loss function in step 4 is: L = 1 - 1 / 2 * (1 - 1 / 2 * (1 - 1 / 2 wherein, is a classification loss, is a bounding box regression loss, is a generalized intersection over union loss, is a contrastive recovery loss, is a weight coefficient of the contrastive recovery loss. 9.A small object detection system based on high-frequency information learning, characterized in that, include: Feature extraction subsystem: used to extract multi-scale features from the input image and select shallow feature maps as the basis; Density-guided coding subsystem: includes a density prediction module and an adaptive sampling module, used to generate scene density maps and output enhanced coding features by an active fidelity strategy that increases the sampling rate in dense target areas; Feature Decoding and Recovery Subsystem: Includes Transformer decoding module and contrastive learning module, used to generate prediction results, and performs active feature recovery to counteract feature smoothing by maintaining a dynamic feature queue and calculating contrastive loss; Prediction and Optimization Subsystem: Used to calculate classification, regression and contrastive recovery loss, update network parameters, and output the final target category and bounding box.

Images