An infrared video small target detection method based on global dynamic query

Abstract

The application discloses an infrared video small target detection method based on global dynamic query. The application constructs a dynamic query aligner (DQAligner) framework, uses global random large displacement data enhancement to expand long tail sample distribution, and improves the adaptability of the model to the rare large motion scene. A pyramid deformable convolution with a dynamic receptive field is used to decompose complex multi-scale motion, and fine alignment of the target is realized. Further, the generalized residual form of deformable convolution is used to iteratively learn the target dynamic query storage, and the target position is searched in the frame to maintain the semantic consistency of the target in multiple frames. The modules of the DQAligner work cooperatively, realize the paradigm shift from rigid alignment to flexible matching, and significantly improve the detection performance in the large displacement and complex dynamic scene.

Description

Technical Field

[0001] This invention relates to the field of computer vision, and more particularly to a method for detecting small targets in infrared video based on global dynamic query. Background Technology

[0002] Infrared imaging, due to its ability to effectively overcome interference under complex lighting conditions, is widely used in maritime rescue, remote sensing monitoring, and early warning systems. Infrared small target detection (IRSTD) has become a research hotspot as a key technology. Generally speaking, infrared small targets are characterized by multi-scale, small area, low signal-to-noise ratio, and blurred outlines.

[0003] To address the aforementioned challenges, Moving Infrared Small Target Detection (MIRSTD) introduces multi-frame temporal cues, exhibiting stronger robustness compared to single-frame detection in environments with weak signal-to-noise ratios and complex backgrounds. Early traditional methods primarily suppressed low-frequency backgrounds and extracted moving target features by designing various spatiotemporal filters. With research advancements, single-frame-based spatiotemporal tensor decomposition methods, human visual contrast methods, and low-rank-based methods have been gradually applied to multi-frame tasks. However, in practical applications, the target's scale and motion often exhibit complex nonlinearities due to factors such as the motion of the detection platform, the target's true velocity, and scale variations. Therefore, the aforementioned assumptions are difficult to satisfy, leading to limitations in the generalization and performance of traditional algorithms.

[0004] In contrast, deep learning-based methods, with their powerful feature extraction and modeling capabilities, can adapt to more complex dynamic scenes and significantly improve target localization performance. STDMANet proposes a spatiotemporal difference multi-scale attention network to extract rich temporal multi-scale features. DTUM incorporates multi-directional motion coding into 3D convolutions to enhance the distinction between target and clutter motion. To further reduce false positives in complex scenes, methods such as TMP and ST-Trans utilize the global perception characteristics of Transformers to achieve efficient fusion of target context information across multiple frames. Notably, inspired by motion compensation techniques in video super-resolution and video object detection, recent methods such as RFR, MOCID, MOPKL, LSTFE-Net, and SSTNet achieve implicit motion compensation through deformable convolutions, motion graph structures, and long-short frame fusion, further improving adaptability to complex dynamic scenes and exhibiting better dynamic adaptability and computational efficiency compared to optical flow methods.

[0005] Despite significant progress in deep learning-based methods, limitations remain when dealing with sparse, large-displacement, or irregular target motion. Optical flow or rigid alignment strategies are prone to failure in motion-blurred or large-displacement scenarios, while deformable convolution and deformable attention, although possessing implicit alignment capabilities, often lead to sampling point drift due to the difficulty in effectively learning offsets. Furthermore, existing instance query methods introduce a large number of redundant queries through random initialization and complex matching, which is not only inefficient but also prone to instability and lacks effective modeling of time dependencies. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by proposing a video infrared small target detection method based on global dynamic query.

[0007] The objective of this invention is achieved through the following technical solution: a method for detecting small targets in infrared video based on global dynamic query, the method comprising:

[0008] For randomly sampled slices in infrared video, the long-tail sample distribution is augmented using global random large displacement data;

[0009] The enhanced slices are preprocessed into multi-scale embeddings and input into the target detection network: cross-scale resource allocation is performed on the multi-scale embeddings to obtain multi-scale coarse motion features. The multi-scale coarse motion features are simultaneously subjected to pyramid deformable alignment based on dynamic receptive field selection mechanism and random initialization of target queries to iteratively learn pixel-level high-dimensional motion feature representation of the target. The results are fused to obtain a motion mask, and finally, the video detection results are output through 3D convolution to output the segmentation results of small targets for each frame.

[0010] Furthermore, the process of obtaining randomly sampled slices in the infrared video is as follows: randomly sample T frames as a slice, and perform flipping, mirroring, and reverse enhancement on the entire slice.

[0011] Furthermore, the method of using global random large displacement data to enhance and expand the long-tail sample distribution specifically involves specifying an affine transformation matrix composed of random parameters and displacements, and applying large displacement enhancement to the frames to be detected in the slice with a preset probability to enhance the model's ability to perceive the long tail of large displacement motion.

[0012] Furthermore, the preprocessing of the enhanced slices into multi-scale embeddings specifically involves:

[0013] The input slices are reshaped into four-dimensional vectors, and then fed into a backbone parallel network constructed using ResBlocks in a time- and batch-parallel manner. Multi-scale motion feature maps are extracted from the backbone network, and a kernel size of [kernel size missing] is applied to these multi-scale motion feature maps. Step size is Overlapping patch embeddings are used to maintain the continuity of target modeling and obtain embeddings. .

[0014] Furthermore, the cross-scale resource allocation for multi-scale embedding specifically involves:

[0015] Multi-scale embedding The number of aligned channels in the 1x1 convolution input is The aligned and embedded data are concatenated together, and then subjected to global average pooling and fully connected processing to obtain channel attention maps. Spatial attention map is obtained by averaging through global channels. And after global channel averaging, further spatial pooling and fully connected processing are used to obtain the temporal attention graph. Finally, the scale, spatial, and channel attention are multiplied with the splicing features, and then a 1*1 convolution is performed to restore the original number of channels at each scale. At the same time, residual connections are added to obtain coarse motion features at multiple scales.

[0016] Furthermore, the pyramid deformable alignment based on the dynamic receptive field selection mechanism specifically refers to:

[0017] It accepts multi-scale coarse motion feature input, and simultaneously inputs reference frame features. With keyframe features ;

[0018] After convolutional modulation, the offset is generated by adaptively selecting the receptive field using a Dynamic Field Offset Generator and fusing l-1 scale offsets. .

[0019] Furthermore, the step of iteratively learning pixel-level high-dimensional motion feature representations of the target through random initialization of the target query specifically involves:

[0020] A hidden state for storing the target representation is randomly initialized at the bottom layer. , and input features The generated key vector Sum value vector Perform pixel-level cross-attention to obtain intermediate input ;

[0021] via gate cell and Perform the first update and store the target representation of the reference frame. ;

[0022] The learning objective's feature dependencies across multiple frames. Cross-attention is performed with the next reference frame, and the target representation is generated iteratively until it is stored. ;

[0023] The target is stored as the query vector and aligned features. Multiplication generates a motion mask, ensuring enhanced target consistency across multiple frames, further suppressing background motion, and maintaining target consistency constraints across multiple frames.

[0024] Furthermore, the object detection network is trained using an adaptive focusing loss, which dynamically adjusts the gradient magnitude of keyframes according to the training stage and target scale, while focusing on difficult examples.

[0025] On the other hand, an infrared video small target detection device based on global dynamic query is also provided, including a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it implements the infrared video small target detection method based on global dynamic query.

[0026] On the other hand, a computer-readable storage medium is also provided, on which a program is stored, which, when executed by a processor, implements the infrared video small target detection method based on global dynamic query.

[0027] The beneficial effects of this invention are:

[0028] 1. A global random displacement data augmentation method for sparse large displacement scenarios is proposed, and a parallelized backbone network structure is adopted to significantly improve inference efficiency and temporal receptive field;

[0029] 2. A fast multi-scale frame resource allocation and dynamic receptive field deformable alignment mechanism are proposed to work together to improve the model's multi-scale perception capability of target features;

[0030] 3. A target dynamic query storage mechanism is proposed, which not only models the dynamic changes of target representation over time, but also implicitly regulates the offset generation process through generalized residual constraints on deformable convolution inputs and outputs. Furthermore, this mechanism can adaptively perform target detection based on keyframe features when large-displacement alignment is difficult, realizing a paradigm shift from rigid alignment to flexible matching for video infrared small target segmentation methods.

[0031] 4. Based on the target centroid provided by the segmentation results, the azimuth and pitch of the target relative to the center of the field of view can be calculated. This can not only guide manual adjustment of the field of view size, but also allow the servo drive mechanism to automatically track the target based on the miss distance. Attached Figure Description

[0032] Figure 1 A flowchart of an infrared video small target detection technology based on global dynamic query is provided for an embodiment of the present invention;

[0033] Figure 2 This is a block diagram illustrating the implementation of the infrared video small target detection model based on global dynamic query provided in an embodiment of the present invention.

[0034] Figure 3 This invention provides an improved parallelized backbone network for small target detection in videos, as described in an embodiment of the invention.

[0035] Figure 4 A cross-scale resource allocator provided in an embodiment of the present invention;

[0036] Figure 5 A diagram illustrating the deformable alignment of a pyramid based on a dynamic receptive field selection mechanism, provided in an embodiment of the present invention.

[0037] Figure 6 A block diagram of the global dynamic target query mechanism for iterative learning provided in this embodiment of the invention;

[0038] Figure 7 The diagram shows the comparison effect between this invention and other methods in dynamic motion and cluttered background scenes.

[0039] Figure 8 This is a schematic diagram of the apparatus provided in an embodiment of the present invention. Detailed Implementation

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] like Figure 1 As shown in the figure, this embodiment provides a flowchart of a motion-robust infrared video small target detection technology, and the steps are as follows:

[0042] S1. Obtain the infrared video sequence set, and augment the long-tail sample distribution by using global random large displacement data to randomly sample slices in the infrared video.

[0043] Specifically, T frames are randomly sampled as a slice, and the entire slice is flipped, mirrored, and inverted to enhance the motion and positional diversity of the target.

[0044] The random parameters angle θ and displacement The constructed affine transformation matrix specifies that large displacement enhancement is applied to the frames to be detected in the slice (the last frame in this invention is the key frame to be detected) with a 20% probability, improving the model's ability to perceive the long tail of large displacement motion. Here, the displacements dx and dy are randomly selected from 50 to 100 pixels. The affine transformation here is as follows, where x and y are the pixel coordinates before enhancement. .

[0045] S2, such as Figure 2 As shown, an object detection network is constructed, including a Fast Scale Resource Assigner (FSRA) module, a Dynamic Field Deformable Alignment (DFDA) module, and a pixel-level Class Query Memory (CQM) module.

[0046] like Figure 3 As shown, the target detection network is an improved parallel backbone network for small target detection in videos.

[0047] The inputs for constructing the target network include:

[0048] a) Reshape the input slice into a four-dimensional vector By parallelizing time and batch processing inputs into a frame-parallel backbone network constructed with residual blocks, the bandwidth usage for model training and inference is significantly improved. This substantially increases the batch normalization statistics to B*T, reduces the noise impact of small batches, and implicitly expands the model's temporal receptive field, allowing gradient calculations to utilize contextual information from multiple time points simultaneously.

[0049] b) Extract multi-scale motion feature maps from the backbone network. .

[0050] c) The execution core size is Step size is Overlapping patch embeddings are used to maintain the continuity of target modeling, resulting in embedded... P represents the patch size, the superscript l indicates the scale, the superscripts K and V indicate the key and value generated by this feature during attention, and the subscripts indicate the frame sequence order.

[0051] like Figure 4 As shown, the FSRA module includes:

[0052] The FSRA module allocates channels, spatial and frame resource relationships for each scale. First, multi-scale embedding is performed. The input is a 1x1 convolution with Ca aligned channels to balance the contribution of each scale and prevent high-resolution but weakly featured layers from being ignored. These layers are then concatenated and processed through global average pooling and fully connected layers to obtain channel attention maps. Spatial attention map is obtained by averaging through global channels. And after global channel averaging, further spatial pooling and fully connected processing are used to obtain the temporal attention graph. Finally, the scale, spatial, and channel attention are multiplied by the stitched features, and then a 1*1 convolution is performed to restore the original number of channels at each scale, while residual connections are added. This process can be described as follows:

[0053]

[0054]

[0055]

[0056]

[0057] in, This is the multi-scale feature map output by this module. This represents a 1x1 convolution. This represents the Hadamard product. In this way, FSRA can quickly allocate resources for each scale, enhance the feature information of small targets through attention, and achieve multi-scale target perception.

[0058] like Figure 5 As shown, the Deformable Pyramid Alignment (DFDA) module includes:

[0059] The results after passing through the cross-scale resource allocator will simultaneously enter the DFDA based on the dynamic receptive field selection mechanism and the pixel-level high-dimensional motion feature representation of the target through iterative learning with randomly initialized target queries.

[0060] DFDA accepts multi-scale coarse motion feature inputs. Taking a certain scale l as an example... The system includes the assigned reference frame features and keyframe features, which are then processed through convolution and bilinear interpolation operations to restore the spatial resolution to the corresponding scale, thus obtaining the reference frame features. With keyframe features As input, after convolutional modulation, the receptive field is adaptively selected through a dynamic receptive field selection and fusion mechanism to generate the offset. In addition, the l-1 scale offset is fused to generate the l-scale offset. In this invention, a reference frame represents a historical frame in a slice, excluding the last frame to be detected, and a key frame represents the last frame in the slice. The features of the reference frame and the key frame are obtained by patch embedding and recovery after FSRA allocation.

[0061] In particular, for modulated deformable convolutions with kernel sizes kh*kw, alignment features It can be obtained using the following formula:

[0062]

[0063] Where P0 represents a position of the aligned feature, N=kh*kw represents all sampling points at the corresponding position, and Pn represents the value of nth of the sampling grid G ​​= (−⌊kh / 2⌋, −⌊kw / 2⌋), (−⌊kh / 2⌋, −⌊kw / 2⌋+1),..., (⌊kh / 2⌋, ⌊kw / 2⌋). ⌊·⌋ is the floor function. This represents the weight at position nth of the convolution kernel. and These represent the learnable offset and learnable modulation scalar, respectively, learned through the DFO Generator:

[0064]

[0065] in , , This indicates a channel split operation. The fractional results are obtained using bilinear interpolation to obtain accurate values. Considering the overparameterization phenomenon caused by simply increasing the kernel size, the expansion rate is used. The expanded convolution is equivalent to a receptive field of 3 to 49. The goal of the DFO Generator is to enable neurons to adaptively adjust the receptive field size according to the stimulus content, thereby adapting to target motion under multi-scale displacement. To achieve this, a gating mechanism is used to carry the information flow of different branches. First, the features of all branches are added and fused, and then global information is embedded through global average pooling. Reducing the channel dimension and global interaction can have side effects on attention learning; therefore, a simple and efficient 1D convolution is used to capture the nonlinear interaction of local channels. Then, the channel dependencies of all branches are recovered from the compressed channel representation, and a soft gating mechanism across branches enables the network to adaptively select information at different spatial scales to generate offsets under dynamic receptive fields.

[0066] In addition to the dynamic receptive field, this invention also designs a pyramid structure to support the multi-scale motion feature decomposition of targets and background in the backbone network. The top layer focuses on larger-scale target motion, and then the offsets and features are progressively propagated to the bottom layer to focus on finer-grained target motion, achieving motion compensation and unifying target semantics at different scales in a coarse-to-fine manner. Input L-level features and Except for the top layer, which directly generates offsets and alignment features, all other layers receive upsampled offsets and alignment features from the layer above. This process can be described as follows:

[0067]

[0068]

[0069] in denoted as bilinear interpolation upsampling, and DCN represents modulated deformable convolution. Compared to optical flow-based methods, DFDA mitigates target blurring and energy loss caused by inaccurate offset estimation through pyramidal dynamic deformable convolution. It can adaptively select the receptive field according to the target motion scale and establishes an implicit motion compensation from coarse to fine.

[0070] The CQM module includes: a reference Figure 6 The iterative learning of pixel-level high-dimensional motion feature representations of the target will be implemented using the CQM module.

[0071] The CQM module implicitly constrains the generation of dynamic offsets from a global perspective, while further enhancing the distinction between target and background features. A hidden state storing the target representation is randomly initialized at the underlying level. , with DFDA input features The generated key vector Sum value vector Perform pixel-level cross-attention to obtain intermediate input Then, through the gate cell... and Perform the first update and store the target representation of the reference frame. Then, the feature dependencies of the learning target across multiple frames are studied. After performing cross-attention with the next reference frame, the target representation is continuously generated iteratively and finally stored. Next, the target is stored as the query vector and aligned features. Multiplication generates a motion mask, ensuring enhanced target consistency across multiple frames, further suppressing background motion, and maintaining target consistency constraints across multiple frames. This process can be described as follows:

[0072]

[0073]

[0074]

[0075] in This indicates the final output characteristics of the CQM module. This indicates a cyclic gating unit operation, which achieves efficient information fusion and memory update of target stored features through unique update and reset gates. Specifically,

[0076]

[0077]

[0078]

[0079]

[0080] in This represents the Sigmoid activation function, and tanh represents the hyperbolic tangent activation function. To update the gate output, It is the output of the reset door. This is the candidate hidden state.

[0081] Finally, the target query obtained from iterative learning is used to perform cross-attention with the DFDA output to generate multi-frame motion masks. Then, the masks and the DFDA output are multiplied element-wise again for self-attention modulation. Finally, the video detection results are output through 3D convolution.

[0082] S3. Train the object detection network using adaptive focal loss. Adaptive focal loss (A daptive FocalLoss) includes: dynamically adjusting the gradient magnitude of keyframes based on the training stage and target scale, while focusing on hard examples.

[0083] As an example, the present invention also provides some comparative cases to highlight the effectiveness comparison of the methods. Infrared images are input into a trained target detection network, and evaluation results such as cross-union ratio, F1 score, detection probability, and false alarm rate are obtained.

[0084] The Intersection over Union (IoU) ratio, as part of the evaluation results, measures the model's ability to describe the overall shape of the target, and is expressed by the following formula:

[0085]

[0086] Where N represents the number of samples. Indicates the actual number of pixels. Indicates the number of pixels in the positive label. This represents the pixels that are predicted to be positive.

[0087] The evaluation result F1 specifically refers to the combined performance of the model in imbalanced samples, specifically the precision (Prec) and recall (Rec), as shown in the following formula:

[0088]

[0089] The detection probability (Pd) of the evaluation result is specifically defined as the ability of the evaluation model to correctly locate the target, which is defined as correctly predicting the target T. pred account for all target T all The proportion is given by the following formula:

[0090]

[0091] In this invention, if the deviation of the centroid of the detected target is less than 3 pixels, it is considered a correctly predicted target.

[0092] The false alarm rate (Fa) of the evaluation result is specifically defined as the model's ability to suppress noise, expressed as the incorrectly predicted target pixel T. false Image pixel size (P) all The proportion is given by the following formula:

[0093]

[0094] As shown in Table 1, the comparison results with existing algorithms on the mainstream dataset IRDST demonstrate that the present invention exhibits significant performance advantages for infrared small target detection in low signal-to-noise ratio scenarios.

[0095] Referring to Table 2, this invention demonstrates a number of parameters, floating-point operations, and an acceptable frame rate per second comparable to current algorithms.

[0096] See Figure 7 Red indicates a correct detection, yellow indicates a false detection, and blue indicates a missed detection. In scenes with cluttered backgrounds and significant displacement, such as… Figure 7 In the scenarios (a) and (b) where buildings move within the field of view, only this invention can correctly detect them between two frames. In cases (c) and (d) where target displacement is extreme and cloud interference is present, all video detection methods other than this invention fail to detect the target; only certain single-frame detection methods (e.g., DNANet and ALCNet) can correctly detect both frames. This demonstrates that this invention is robust to detecting small infrared targets in multi-scale dynamic scenes and provides more accurate detection results than other inventions.

[0097] Table 1. Detection performance of this invention on the IRDST full video dataset and low signal-to-noise ratio subset.

[0098]

[0099] Corresponding to the aforementioned embodiment of an infrared video small target detection method based on global dynamic query, the present invention also provides an embodiment of an infrared video small target detection device based on global dynamic query.

[0100] See Figure 8The present invention provides an infrared video small target detection device based on global dynamic query, comprising a memory and one or more processors. The memory stores executable code, and when the processor executes the executable code, it is used to implement an infrared video small target detection method based on global dynamic query in the above embodiment.

[0101] The embodiment of the infrared video small target detection device based on global dynamic query provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 8 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is an infrared video small target detection device based on global dynamic query provided by the present invention. (Except for...) Figure 8 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0102] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0103] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0104] refer to Figure 8 The present invention also provides a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, it implements an infrared video small target detection method based on global dynamic query as described in the above embodiments.

[0105] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device of any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices of any data processing device. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0106] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the aforementioned infrared video small target detection method based on global dynamic query.

[0107] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0108] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. This application is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for detecting small targets in infrared video based on global dynamic query, characterized in that, The method includes: For randomly sampled slices in infrared video, the long-tail sample distribution is augmented using global random large displacement data; The enhanced slices are preprocessed into multi-scale embeddings and input into the target detection network: cross-scale resource allocation is performed on the multi-scale embeddings to obtain multi-scale coarse motion features. The multi-scale coarse motion features are simultaneously subjected to pyramid deformable alignment based on dynamic receptive field selection mechanism and random initialization of target queries to iteratively learn pixel-level high-dimensional motion feature representation of the target. The results are fused to obtain a motion mask, and finally, the video detection results are output through 3D convolution to output the segmentation results of small targets for each frame.

2. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The process of obtaining randomly sampled slices in the infrared video is as follows: randomly sample T frames as a slice, and then flip, mirror, and reverse the entire slice for enhancement.

3. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The method of using global random large displacement data to enhance and expand the long-tail sample distribution specifically involves specifying an affine transformation matrix composed of random parameters and displacements, and applying large displacement enhancement to the frames to be detected in the slice with a preset probability to enhance the model's ability to perceive the long tail of large displacement motion.

4. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The specific steps of preprocessing the enhanced slices into multi-scale embeddings are as follows: The input slices are reshaped into four-dimensional vectors, and the time and batch parallelization is applied to the backbone parallel network constructed with ResBlock. Multi-scale motion feature maps are extracted from the backbone network, and a kernel size of P is applied to the multi-scale motion feature maps of patch size P. Step size is Overlapping patch embeddings are used to maintain the continuity of target modeling, resulting in preprocessed embeddings, where the superscript l represents the scale.

5. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The specific steps for cross-scale resource allocation in multi-scale embedding are as follows: The number of channels in the multi-scale embedding input is aligned using a 1*1 convolution. The aligned embeddings are then concatenated and subjected to global average pooling and fully connected layers to obtain channel attention maps. A spatial attention map is obtained by global channel averaging, and a temporal attention map is obtained by further spatial pooling and fully connected layers after global channel averaging. Finally, the scale, spatial, and channel attention are multiplied with the concatenated features and then subjected to a 1*1 convolution to restore the original number of channels at each scale. Residual connections are added to obtain coarse multi-scale motion features.

6. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The pyramid deformable alignment based on the dynamic receptive field selection mechanism specifically refers to: It accepts multi-scale coarse motion feature input, and simultaneously inputs reference frame features and keyframe features; After convolutional modulation, the offset is generated by adaptively selecting the receptive field and fusing the l-1 scale offset through a dynamic field offset generator.

7. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The specific method for iteratively learning pixel-level high-dimensional motion feature representations of the target through random initialization of the target query is as follows: A hidden state for storing the target representation is randomly initialized at the bottom layer. , and input features The generated key vector Sum value vector Perform pixel-level cross-attention to obtain intermediate input ; via gate cell and Perform the first update and store the target representation of the reference frame. ; The learning objective's feature dependencies across multiple frames. Cross-attention is performed with the next reference frame, and the target representation is generated iteratively until it is stored. ; The target is stored as the query vector and aligned features. Multiplication generates a motion mask, ensuring enhanced target consistency across multiple frames, further suppressing background motion, and maintaining target consistency constraints across multiple frames.

8. The infrared video small target detection method based on global dynamic query according to claim 1, characterized in that, The object detection network is trained using adaptive focusing loss, which dynamically adjusts the gradient magnitude of keyframes according to the training stage and target scale, while focusing on difficult examples.

9. An infrared video small target detection device based on global dynamic query, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that, When the processor executes the executable code, it implements an infrared video small target detection method based on global dynamic query as described in any one of claims 1-8.

10. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements an infrared video small target detection method based on global dynamic query as described in any one of claims 1-8.

Images