Small target detection method based on FineLocNet

Abstract

According to the small target detection method based on the FineLocNet, a distribution perception boundary regression module and a hierarchical knowledge transfer module are designed on the basis of a Desection Transform framework, and through probabilistic boundary modeling and cross-scale feature cooperation, accurate positioning is achieved while efficient reasoning is kept. Specifically, the DABR models bounding box coordinates into probability distribution explicit quantitative positioning uncertainty, and dynamically focuses a key feature region through a learnable spatial offset mechanism, thereby effectively improving the small target boundary positioning precision. The HKT adopts a progressive feature alignment strategy, the semantic discrimination capability of a deep network is transmitted to shallow high-resolution features, meanwhile, a two-way fusion path of scale perception is constructed, the spatial consistency and semantic complementarity of different resolution features are ensured, specifically, HKT and DABR work cooperatively, high-level feature extraction and detection offset correction are achieved, and the detection accuracy is improved. Therefore, the target detection accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of small object detection in end-to-end image detection, specifically involving a small object detection method based on FineLocNet, which is applicable to multiple fields such as remote sensing monitoring, medical imaging, and autonomous driving. Background Technology

[0002] Small target detection, a fundamental task in computer vision, plays an indispensable role in numerous practical applications. In remote sensing, accurately identifying small-scale targets such as vehicles and buildings in satellite images is crucial for urban planning, traffic management, and military reconnaissance. In medical image analysis, accurately locating minute lesions and cell nuclei is key to early disease diagnosis and precise treatment. In autonomous driving systems, reliable detection of distant pedestrians and traffic signs is directly related to driving safety. Although deep learning-based target detection methods have made significant progress in recent years, the detection performance of small targets remains far lower than that of conventional-scale targets due to their inherent low pixel count and limited visual features, becoming a bottleneck restricting the overall performance improvement of detection systems. Therefore, developing high-precision detection algorithms specifically designed for the characteristics of small targets has significant theoretical value and broad application prospects.

[0003] However, existing small object detection methods face two core challenges. First, small objects occupy an extremely small percentage of pixels and have sparse boundary features, making it difficult for traditional bounding box regression methods to accurately locate them. Specifically, small objects typically occupy only tens or even a few pixels in an image, and their edge information is blurred and easily affected by background noise, making regression strategies based on fixed anchor boxes or center points prone to significant localization errors. Existing methods typically employ deterministic coordinate regression, which cannot effectively model the inherent uncertainty in small object localization, resulting in a lack of robustness in boundary prediction. Second, existing detection methods often require multiple downsampling operations to expand the receptive field when extracting high-level semantic features, which inevitably leads to a sharp decrease in spatial resolution and severe loss of fine-grained features of small objects. Although multi-scale feature fusion strategies alleviate this problem to some extent, it is still difficult to effectively balance the spatial accuracy and semantic richness of features when dealing with extreme scale differences. Especially for small objects, strong responses from large objects often mask their weak signals, while sophisticated feature alignment mechanisms bring significant computational overhead, making it difficult for existing methods to achieve an ideal balance between detection accuracy and inference efficiency.

[0004] A search revealed that Chinese invention patent application number CN202511426520.3 discloses an enhanced method and apparatus for small target detection in autonomous driving based on the RT-DETR framework. The method includes: constructing a small target detection network, which comprises an encoder and a decoder; inputting an autonomous driving scene image into the encoder for feature extraction to obtain multiple feature maps of different resolutions; fusing the multiple feature maps of different resolutions through a local feature fusion module to obtain a fused feature map representing image detail information and image semantic information of the autonomous driving scene image; and determining the bounding boxes of small targets in the autonomous driving scene image based on the fused feature map through a deformable cross-attention module, and outputting the detection result with the bounding boxes of the small targets.

[0005] Upon comparison, the aforementioned prior art documents differ from this application in the following ways:

[0006] 1. Patent CN202511426520.3 uses a Local Feature Fusion (LFFM) module without dedicated semantic alignment in the feature alignment stage. It only applies spatial attention and channel attention mechanisms to the multi-resolution feature maps to generate the first processed feature map. Then, it aligns the spatial size through upsampling / downsampling to generate the fused feature map. The HKT module introduced in this patent uses overscaled dot product attention to calculate the cross-scale semantic affinity matrix and establish the semantic dependency between high-level features and low-level features. A lightweight network with two 3×3 convolutions generates a pixel-by-pixel two-dimensional offset field Δp. Based on the offset field, upsampling / downsampling is implemented to capture the spatial misalignment between semantic regions and detail regions.

[0007] 2. Patent CN202511426520.3 employs a deterministic regression approach: It generates high-quality queries through a deformable cross-attention module, extracts and weights features based on sampling offsets, and finally outputs bounding box coordinates through a feedforward neural network, filtering low-confidence results based on a confidence threshold. However, it lacks modeling of uncertainties in the localization process and has weak tolerance for small positioning deviations of small targets. This patent proposes a Distribution-Aware Boundary Regression (DABR) module, whose core is probabilistic modeling and iterative refinement: It models the bounding box coordinates as a generalized Gaussian distribution (GGD), predicts distribution parameters (scale α, shape β) to quantify positioning uncertainty, and supports multiple rounds of iterative refinement, dynamically adjusting the correction magnitude based on the uncertainty of the previous round to achieve progressive localization from coarse to fine.

[0008] 3. Patent CN202511426520.3 is based on RT-DETR and employs end-to-end supervised learning. The training objectives are classification loss and bounding box regression loss (L1 / GIoU), emphasizing a balance between real-time performance and accuracy. It does not design separate loss functions for cross-scale feature alignment and small object localization. This patent constructs a multi-task learning framework that jointly optimizes three types of losses: detection loss (Focal Loss + L1 + GIoU), HKT alignment loss (supervising cross-scale feature semantic consistency), and DABR distribution loss (negative log-likelihood loss, penalizing overestimation of uncertainty). Increasing weights are applied to each round of prediction in iterative refinement to enhance the later refinement effect. Summary of the Invention

[0009] The purpose of this invention is to solve the problems of low accuracy and poor robustness in small target detection in the prior art, and to provide a small target detection method based on FineLocNet.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A small object detection method based on FineLocNet includes the following steps:

[0012] S1. Obtain the target image dataset and preprocess the obtained dataset;

[0013] S2. Input the processed dataset into the backbone network ResNet-50 to extract multi-scale feature pyramids, including shallow and deep features;

[0014] S3. Input the multi-scale feature pyramid into the hierarchical knowledge transfer module (HKT), and achieve cross-scale semantic transfer and feature alignment through semantic attention modeling, offset field prediction, feature calibration and bidirectional feature fusion;

[0015] S4. Input the multi-scale aligned features output by HKT into the Transformer encoder to perform global context modeling;

[0016] S5. Using the Transformer decoder, preliminary detection results are generated based on the interaction between Object Queries and encoded features;

[0017] S6. Input the preliminary detection results into the Distribution Aware Boundary Regression (DABR) module, model the bounding box coordinates as a probability distribution, and output a refined bounding box prediction through uncertainty quantification and iterative refinement mechanisms.

[0018] The above steps will be explained in further detail.

[0019] Step S1: Data Preprocessing

[0020] After obtaining the small target image dataset, data preprocessing is performed. Preprocessing includes data augmentation, such as multi-scale training, random horizontal flipping, and color jittering. These operations can increase the diversity of the dataset and improve the model's generalization ability.

[0021] Step S2: Backbone Network Feature Extraction

[0022] The processed dataset is then fed into the backbone network ResNet-50 to extract a multi-scale feature pyramid. ResNet-50 progressively extracts hierarchical feature representations through four residual stages:

[0023] ;

[0024] in, , These correspond to the feature dimensions of each level. , This corresponds to the spatial resolution. The multi-scale feature pyramid exhibits a significant semantic-spatial tradeoff: shallow features ( It retains rich edge and texture details, but has a low level of semantic abstraction; deep features ( It has highly discriminative semantic information, but low spatial resolution and serious loss of fine structural information of small targets.

[0025] Step S3: Hierarchical Knowledge Transfer Module (HKT)

[0026] In multi-scale feature fusion, traditional FPN employs a strategy of directly adding features after fixed interpolation upsampling, implicitly assuming that features of different resolutions are naturally aligned in space. However, there is a systematic spatial offset between the semantic regions of high-level features (such as the target center response) and the detail regions of low-level features (such as the target edge response). This misalignment is particularly detrimental to small targets that occupy only a few pixels. To address this issue, HKT uses "semantic attention modeling". Offset field prediction The three-level cascaded structure of "feature calibration" captures cross-scale semantic dependencies and predicts semantically guided spatial offset fields through an attention mechanism, achieving accurate feature alignment.

[0027] Multi-scale semantic attention aggregation

[0028] To establish semantic relationships between high- and low-level features, a lightweight cross-scale semantic attention mechanism is first constructed. Given low-level reference features... Features to be aligned with higher levels (downsampling factor) First through Convolution projects both into a unified semantic embedding space:

[0029] ;

[0030] in, and For a learnable projection matrix, the embedding dimension To ensure computational efficiency.

[0031] The projected features are flattened into a sequence. and ,in , These represent the total number of spatial locations. Cross-scale semantic affinity matrix. Attention is calculated by scaling the dot product:

[0032] ;

[0033] The first of the matrix Each element represents the location of a high-level feature. For low-level feature locations Semantic dependency weights. Based on We perform weighted aggregation of low-level features to generate enhanced features that encode cross-scale semantic context:

[0034] ;

[0035] in, It effectively injects low-level spatial details into high-level semantic structures, providing rich semantic priors for subsequent spatial alignment.

[0036] Semantic Perceptual Offset Field Prediction

[0037] Traditional feature alignment methods (such as deformable convolution) directly predict offsets from local features, ignoring the global consistency of semantic content. To address this issue, semantic aggregated features are fused with the original high-level features to construct a semantically aware offset field prediction network.

[0038] Specifically, firstly and Stitched along the channel dimension and upsampled using bilinear interpolation. The resolution, and then with Joint Input Lightweight Migration Prediction Network :

[0039] ;

[0040] in, Represents a pixel-by-pixel two-dimensional offset field. Contains only two layers The convolutional layer (with 64 channels in the intermediate layers) has extremely low computational cost. This offset field explicitly encodes cross-scale semantic discrepancies, accurately predicting and correcting spatial misalignments between semantic and detail regions. Compared to standard deformable convolutions, this design incorporates global semantic context. Instead of relying solely on local features, offset prediction is performed, resulting in more accurate alignment of semantic boundaries.

[0041] Offset resampling and feature calibration

[0042] Based on predicted migration field High-level characteristics A two-step calibration process is performed. First, spatial resampling is performed, using the offset field to perform deformable sampling of the upsampled high-level features:

[0043] ;

[0044] in, To output any location on the feature map, Indicates its four neighboring regions, This is a bilinear interpolation weighting function. This operation ensures that the semantic regions of high-level features are spatially precisely aligned with those of low-level features.

[0045] Subsequently, using the semantic attention matrix Attention calibration is performed to suppress background noise. Convert to a spatial weighted graph:

[0046] ;

[0047] Finally, the spatial alignment features are modulated element-wise to obtain the calibrated features:

[0048] ;

[0049] in, This represents the Hadamard product. It combines accurate spatial correspondence with robust semantic consistency and will be used for subsequent bidirectional feature fusion.

[0050] Bidirectional feature fusion

[0051] Based on the above alignment process, the following was obtained A bidirectional feature fusion path was constructed. Inspired by BiFPN, two topology optimizations were performed: redundant nodes with only a single input edge were removed; and cross-layer skip connections were added between the original input and output nodes at the same level.

[0052] The complete alignment process is denoted as... The top-down path propagates higher-level semantics to lower levels step by step:

[0053] ;

[0054] in, This serves as the starting point for recursion. The bottom-up path then transmits details from lower levels back to higher levels:

[0055] ;

[0056] The final output multi-scale features The input Transformer encoder is modeled globally.

[0057] Step S4: Global Modeling of the Transformer Encoder

[0058] Multi-scale alignment features output by HKT Further global context modeling is needed to capture long-range dependencies. The multi-scale feature maps are flattened and concatenated, then incorporating learnable 2D positional encodings before being input into the Transformer encoder.

[0059] ;

[0060] in, The total length of the feature sequence. To hide dimensions, Encoding learnable locations. The encoder establishes long-range dependencies between any two spatial locations through a multi-layered self-attention mechanism:

[0061] ;

[0062] This global modeling is particularly important for understanding the relationship between small targets and their context. Features output by the encoder. They will serve as the key and value for the decoder's cross-attention.

[0063] Step S5: Preliminary detection of the Transformer decoder

[0064] The decoder receives the encoder output. and a set of learnable object queues The detection results are generated through multiple iterations. Each decoder layer sequentially contains a self-attention layer, a cross-attention layer, and a feedforward network, and its update process can be formalized as follows:

[0065] ;

[0066] The self-attention layer models the relationships between queries to avoid duplicate detections, while the cross-attention layer enables queries to interact with encoded features to aggregate target-related information.

[0067] After multiple iterations, each query The prediction head outputs preliminary classification results and bounding box parameters:

[0068] ;

[0069] in, The bounding box is represented using normalized center coordinates to indicate the number of categories. Preliminary predictions from the decoder and corresponding query features It will be used as input to the DABR module for subsequent probabilistic boundary refinement.

[0070] Step S6: Distribution-Aware Boundary Regression Module (DABR)

[0071] Bounding boxes directly output by the decoder Using deterministic regression has significant drawbacks when dealing with small targets: small targets have an extremely low pixel count, and even a slight positioning error can cause drastic fluctuations in IoU; edge information is sparse and blurry, and the true boundary itself has inherent uncertainty. Therefore, DABR, as a refinement module after the decoder, models the bounding box coordinates as a probability distribution, receiving... and Query features It outputs refined bounding box predictions through uncertainty quantification and iterative refinement mechanisms.

[0072] Uncertainty Quantification Gaussian Modeling

[0073] To capture the centroid position of the target and the degree of blurring of its boundaries, each coordinate component of the bounding box is... The model is based on independent probability distributions. DABR no longer directly outputs deterministic coordinate values, but instead relies on the initial predictions from the decoder. Predicting distribution parameters ,in This represents the offset relative to the initial prediction. This represents the uncertainty of the estimate. The final bounding box coordinates are determined by... Given the standard Gaussian distribution, the probability density function is defined as:

[0074] ;

[0075] Considering that the decay characteristics of the target boundary may deviate from the standard Gaussian distribution in real-world scenarios, a Generalized Gaussian Distribution (GGD) is further introduced to enhance the model's expressive power:

[0076] ;

[0077] in, This is a scaling parameter (positively correlated with uncertainty). For shape parameters (controlling the decay rate at the tail of the distribution). For the Gamma function. When It degenerates into a Gaussian distribution. The time follows a Laplace distribution. This probabilistic modeling allows DABR to adaptively adjust its refinement strategy based on the confidence level of the initial prediction: for large targets with relatively accurate initial predictions, the model adjusts its refinement strategy accordingly, targeting smaller targets with higher confidence levels. and The distribution is concentrated near the initial prediction; for small targets where the initial prediction is uncertain, the model output is larger. and more significant Significant revisions will be made.

[0078] DABR uses a lightweight, refined network to extract query features. and preliminary bounding box Predict the distribution parameters. Let... The features after splicing, among which Encoding the bounding box coordinates as high-dimensional features, the prediction process for the distribution parameters is as follows:

[0079] ;

[0080] Among them, the Softplus activation function ensures ,and The design avoids overly sharp distributions that could lead to unstable training.

[0081] Uncertainty-guided iterative refinement

[0082] DABR supports multi-round iterative refinement, with each round making finer adjustments based on the prediction results and uncertainty estimates of the previous round. Let the... The bounding box of the wheel is predicted as Uncertainty is Then the first The wheel finishing process is as follows:

[0083] ;

[0084] in, This represents the decoder's initial prediction. As iterations proceed, uncertainty increases. Gradually reduce, adjustment range Gradually reduce the size to achieve a progressive positioning from coarse to fine.

[0085] Uncertainty estimation is also used to guide feature sampling during the refinement process. An uncertainty-based spatial attention mechanism is introduced, enabling the refinement network to dynamically adjust its region of interest based on the confidence level of the current prediction: when uncertainty is high, the network searches for target boundary information over a larger spatial area; when uncertainty is low, the network focuses on fine features near the current predicted boundary. Specifically, RoIAlign predicts features from the encoder based on the current bounding box. Extracting local features The size of the sampling region is determined by uncertainty. Dynamic adjustment: ,in This is the scaling factor.

[0086] The training objective of DABR consists of two parts: distributed regression loss and iterative refinement loss.

[0087] For distributed regression, negative log-likelihood loss (NLL Loss) is used instead of the traditional... The loss, minimizing NLL, is equivalent to maximizing the probability that the true label falls within the high-probability region of the predicted distribution. For a single coordinate component... The NLL loss is defined as:

[0088] ;

[0089] The first term is the residual term, which influences the predicted mean. Approximating the true coordinates The second term is an uncertainty regularization term, which penalizes excessively large penalties. To prevent the model from "lazy," the four coordinate components of the bounding box are summed to obtain the distributed loss for a single prediction:

[0090] ;

[0091] For multi-round iterative refinement, supervision is applied to each round of prediction, and incremental loss weights are used to emphasize the importance of later refinement:

[0092] ;

[0093] in, For the number of iteration rounds, The total loss of DABR increases with each iteration round.

[0094] ;

[0095] During reasoning, use the mean of the distribution from the last round. Output as the final bounding box coordinates.

[0096] Training objectives

[0097] FineLocNet employs a multi-task learning framework that jointly optimizes the detection loss, HKT alignment loss, and DABR distribution loss.

[0098] ;

[0099] Detection loss Designed in accordance with the standard DETR: ,in Focal Loss is used for classification. and They are respectively Bounding box regression loss and generalized IoU loss. HKT alignment loss. Supervised learning of the offset field ensures semantic consistency between the aligned high-level and low-level features: ,in By blocking gradient propagation, we can ensure that the alignment process does not negatively affect the feature extraction of the backbone network.

[0100] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0101] 1. This invention proposes FineLocNet, a framework designed with a distribution-aware boundary regression module and a hierarchical knowledge transfer module. Through probabilistic boundary modeling and cross-scale feature collaboration, it achieves accurate localization while maintaining efficient inference. 2. This invention improves the accuracy and robustness of small target detection. Through multi-layer iteration and fine-tuning of bounding boxes, it solves the problems of low accuracy and poor robustness in current small target detection. 3. The overall architecture is based on DETR-type detection, eliminating the need for complex post-processing (such as NMS) and achieving high inference efficiency. The feature transfer in HKT includes a two-layer convolution and scaling dot product attention mechanism, and DABR utilizes a generalized Gaussian distribution to model the uncertainty of target detection, resulting in low overall computational overhead and suitability for real-time or edge deployment. 4. This invention is applicable to multiple fields with high requirements for small target detection accuracy, such as remote sensing monitoring, medical imaging, autonomous driving, and video surveillance. The method has good versatility and can be extended to other visual tasks requiring fine localization. Attached Figure Description

[0102] Figure 1 This is a flowchart of the Fine-LocNet detection process in an embodiment of the present invention;

[0103] Figure 2 This is a diagram of the Fine-LocNet network structure in an embodiment of the present invention;

[0104] Figure 3 This is a structural diagram of the HKT in an embodiment of the present invention;

[0105] Figure 4This is a structural diagram of the DABR in an embodiment of the present invention. Detailed Implementation

[0106] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The specific implementation methods of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0107] like Figure 1 The diagram shown is a flowchart of the detection process in this embodiment. Figure 2 This is the network structure diagram of this embodiment. This embodiment provides a small target detection method based on FineLocNet, which includes the following steps during detection:

[0108] S1. First, obtain a dataset of small target images. Preprocess the dataset, including data augmentation operations such as multi-scale training, random horizontal flipping, and color jittering. This allows the processed data to be better used by the model for feature learning, improving training accuracy. Specifically, multi-scale training allows the model to train on images of different scales, enhancing its adaptability to targets of varying sizes; random horizontal flipping increases dataset diversity, preventing overfitting to the target's orientation; and color jittering alters the image's color features, making the model focus more on key features such as the target's shape and texture.

[0109] S2. The data is input into the ResNet-50 backbone network to extract a multi-scale feature pyramid. The input image is processed through convolutional and pooling layers to obtain a preliminary feature map, which retains the basic contour and color information of the image. As the residual stage progresses, the semantic information of the feature map gradually increases, while the spatial resolution gradually decreases. For example, in the fourth residual stage, the spatial resolution of the feature map is only 1 / 16 of that of the input image, but it contains rich high-level semantic information, such as the object category and semantic relationships.

[0110] S3. Input the multi-scale feature pyramid into the hierarchical knowledge transfer module (HKT), such as... Figure 3 The diagram shows the structure of HKT. HKT achieves cross-scale semantic transfer and feature alignment through semantic attention modeling, offset field prediction, feature calibration, and bidirectional feature fusion. In the semantic attention modeling stage, a lightweight cross-scale semantic attention mechanism is constructed to establish semantic relationships between high- and low-level features. Given a low-level reference feature... Features to be aligned with higher levels (downsampling factor) First through Convolution projects both into a unified semantic embedding space:

[0111] ;

[0112] in, and For a learnable projection matrix, the embedding dimension To ensure computational efficiency, the projected features are flattened into a sequence. and ,in , These represent the total number of spatial locations. Cross-scale semantic affinity matrix. Attention is calculated by scaling the dot product:

[0113] ;

[0114] The first of the matrix Each element represents the location of a high-level feature. For low-level feature locations The semantic dependency weights are used to inject low-level detailed information into high-level semantic features. In the offset field prediction stage, a semantically aware offset field prediction network accurately predicts and corrects spatial misalignments between semantic and detailed regions. Offset field The calculation process is as follows:

[0115] ;

[0116] in, Represents a pixel-by-pixel two-dimensional offset field. Contains only two layers Convolution (64 channels in the intermediate layer) has extremely low computational cost. The offset field predicts based on the global semantic context, thus enabling more accurate alignment of features at different scales. During feature calibration, spatial resampling and attention calibration ensure precise spatial alignment of the semantic regions of high-level features with low-level features and suppress background noise. Spatial resampling utilizes the offset field to perform deformable sampling of the upsampled high-level features:

[0117] ;

[0118] in, To output any location on the feature map, Indicates its four neighboring regions, This is a bilinear interpolation weight function. Subsequently, the semantic attention matrix is ​​used... Attention calibration is performed to suppress background noise. Convert to a spatial weighted graph:

[0119] ;

[0120] Finally, the spatial alignment features are modulated element-wise to obtain the calibrated features:

[0121] ;

[0122] in, This represents the Hadamard product. In the bidirectional feature fusion stage, cross-scale semantic transfer and feature fusion are achieved through top-down and bottom-up paths, enabling features at different scales to complement each other and improve feature expressiveness. The top-down path progressively transmits high-level semantics to lower levels:

[0123] ;

[0124] in, This serves as the starting point for recursion. The bottom-up path then transmits details from lower levels back to higher levels:

[0125] ;

[0126] S4. Input the multi-scale aligned features output by HKT into the Transformer encoder for global context modeling. The Transformer encoder establishes long-range dependencies between any two spatial locations through a multi-layer self-attention mechanism:

[0127] ;

[0128] in, The total length of the feature sequence. To hide dimensions, This is for learnable position encoding. Features output by the encoder. This global modeling is particularly important for small target detection, because small targets are often easily affected by background noise, and global contextual information can help the model better locate small targets.

[0129] S5. Using the Transformer decoder, preliminary detection results are generated based on the interaction between Object Queries and encoded features. The decoder receives the encoder output. and a set of learnable object queues The detection results are generated through multiple iterations. Each decoder layer sequentially contains a self-attention layer, a cross-attention layer, and a feedforward network, and its update process can be formalized as follows:

[0130] ;

[0131] After multiple iterations, each query The prediction head outputs preliminary classification results and bounding box parameters:

[0132] ;

[0133] in, The bounding box is represented using normalized center coordinates to indicate the number of categories. For example, for a small target, the decoder can predict the target's category and approximate location based on encoded features and object queues.

[0134] S6. Input the preliminary detection results into the Distribution-Aware Boundary Regression (DABR) module, such as... Figure 4 The diagram shows the structure of DABR, which models the bounding box coordinates as a probability distribution and outputs refined bounding box predictions through uncertainty quantization and iterative refinement mechanisms. In the uncertainty quantization Gaussian modeling stage, each coordinate component of the bounding box... The model is based on independent probability distributions. For a standard Gaussian distribution, the probability density function is defined as:

[0135] ;

[0136] Considering that the decay characteristics of the target boundary may deviate from the standard Gaussian distribution in real-world scenarios, a Generalized Gaussian Distribution (GGD) is further introduced to enhance the model's expressive power:

[0137] ;

[0138] in, This is a scaling parameter (positively correlated with uncertainty). For shape parameters (controlling the decay rate at the tail of the distribution). Let be the Gamma function. In the uncertainty-guided iterative refinement phase, DABR supports multiple rounds of iterative refinement, with each round making finer adjustments based on the prediction results and uncertainty estimates of the previous round. Let the th... The bounding box of the wheel is predicted as Uncertainty is Then the first The wheel finishing process is as follows:

[0139] ;

[0140] in, This represents the decoder's initial prediction. As iterations proceed, uncertainty increases. Gradually reduce, adjustment range The loss is gradually reduced to achieve a progressive positioning from coarse to fine. In the target optimization phase, negative log-likelihood loss (NLL Loss) is used instead of traditional... The loss, minimizing NLL, is equivalent to maximizing the probability that the true label falls within the high-probability region of the predicted distribution. For a single coordinate component... The NLL loss is defined as:

[0141] ;

[0142] The first term is the residual term, which influences the predicted mean. Approximating the true coordinates The second term is an uncertainty regularization term, which penalizes excessively large penalties. To prevent the model from "lazy," the four coordinate components of the bounding box are summed to obtain the distributed loss for a single prediction:

[0143] ;

[0144] For multi-round iterative refinement, supervision is applied to each round of prediction, and incremental loss weights are used to emphasize the importance of later refinement:

[0145] ;

[0146] in, For the number of iteration rounds, It increases with each iteration round.

[0147] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A small target detection method based on FineLocNet, characterized in that, Includes the following steps: S1. Obtain the target image dataset and preprocess the obtained dataset; S2. Input the preprocessed dataset into the backbone network ResNet-50 to initially extract a multi-scale feature pyramid, including shallow and deep features. The feature extraction process is as follows: ； in, , These correspond to the feature dimensions of each level. , The corresponding spatial uncertainty-guided iterative refinement resolution; S3. Input the multi-scale feature pyramid into the hierarchical knowledge transfer module, and achieve cross-scale semantic transfer and feature alignment through semantic attention modeling, offset field prediction, feature calibration and bidirectional feature fusion; S4. Input the multi-scale aligned features output by the hierarchical knowledge transfer module into the Transformer encoder to perform global context modeling; S5. Using the Transformer decoder, preliminary detection results are generated based on the interaction between Object Queries and encoded features; S6. Input the preliminary detection results into the distribution-aware boundary regression module, model the boundary box coordinates as a probability distribution, and output a refined boundary box prediction through uncertainty quantification and iterative refinement mechanism.

2. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S1, the preprocessing is data augmentation processing.

3. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S3, the semantic attention modeling of the hierarchical knowledge transfer module establishes semantic associations between high- and low-level features by constructing a lightweight cross-scale semantic attention mechanism. The cross-scale semantic affinity matrix A is calculated by scaling dot product attention. ； The first of the matrix Each element represents the location of a high-level feature. For low-level feature locations The semantic dependency weights are calculated; based on A, the low-level features are weighted and aggregated to generate enhanced features that encode cross-scale semantic context. ； in, .

4. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S3, the offset field prediction uses a semantically aware offset field prediction network to accurately predict and correct the spatial misalignment between the semantic region and the detail region; offset field The calculation process is as follows: ； in, Represents a pixel-by-pixel two-dimensional offset field. Contains two layers convolution.

5. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S3, the feature calibration achieves accurate feature alignment through spatial resampling and attention calibration; spatial resampling utilizes the offset field to perform deformable sampling on the upsampled high-level features: ； in, To output any location on the feature map, Indicates its four neighboring regions, This is the bilinear interpolation weighting function; Subsequently, using the semantic attention matrix Perform attention calibration to suppress background noise, Convert to a spatial weighted graph: ； Finally, the spatial alignment features are modulated element-wise to obtain the calibrated features: ； in, This represents the Hadamard product.

6. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S3, the bidirectional feature fusion achieves cross-scale semantic transfer and feature fusion through top-down and bottom-up paths; The top-down path propagates higher-level semantics to lower levels step by step: ； in, This is the starting point of the recursion; The bottom-up path transmits details of the lower-level space back to the upper levels: ； The final output multi-scale features The input Transformer encoder is modeled globally.

7. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S6, the distribution-aware boundary regression module models the bounding box coordinates as independent probability distributions. For a standard Gaussian distribution, the probability density function is defined as: ； Considering that the decay characteristics of the target boundary may deviate from the standard Gaussian distribution in real-world scenarios, a generalized Gaussian distribution is further introduced to enhance the model's expressive power: ； in, For scale parameters, For shape parameters, For the Gamma function; when It degenerates into a Gaussian distribution. The time follows a Laplace distribution.

8. The small target detection method based on FineLocNet according to claim 1, characterized in that, In step S6, the distribution-aware boundary regression module supports multi-round iterative refinement, with each round of refinement making more precise adjustments based on the prediction results and uncertainty estimates of the previous round; let the bounding box prediction of the t-th round be... Uncertainty is Then the refinement process in round t+1 is as follows: ； in, This is the initial prediction from the decoder; as iterations proceed, uncertainty increases. Gradually reduce, adjustment range Gradually reduce the size to achieve a progressive positioning from coarse to fine.

9. A small target detection method based on FineLocNet according to any one of claims 1 to 8, characterized in that, The training objective employs a multi-task learning framework, jointly optimizing the detection loss, HKT alignment loss, and DABR distribution loss: ； Detection loss Designed in accordance with the standard DETR: ,in Focal Loss is used for classification. and They are respectively Bounding box regression loss and generalized IoU loss; HKT alignment loss Supervised learning of the offset field ensures semantic consistency between the aligned high-level and low-level features: ,in By blocking gradient propagation, we can ensure that the alignment process does not negatively affect the feature extraction of the backbone network.

Images