A dual-branch infrared small target detection method based on neural low-rank background modeling

Abstract

The application discloses a kind of double-branch infrared small target detection methods based on neural low rank background modeling, it is related to target detection technical field, including the following steps: preparing data set;Double-branch detection model is constructed;Network model is trained;Loss function and evaluation index are designed;Model is fine-tuned;Model is saved.The application adopts double-branch architecture: branch one passes through neural low rank background modeling, utilizes multilayer perception mechanism to build parameterized background tensor, combines neural regularization to constrain the local smoothness and time continuity of background, effectively suppresses background noise;Branch two introduces double perception enhancement and resolution adaptive amplification module, and the details of target are strengthened by channel reorganization upsampling and residual connection reinforcement.The physical size of design joint target, spatial coordinate information and focus loss function are accurately optimized The detection accuracy of model provides an innovative solution for efficient, robust infrared small target detection.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target detection, and particularly relates to a dual-branch infrared small target detection method based on neural low-rank background modeling. BACKGROUND

[0002] Target detection is an important research direction in the field of computer vision, aiming to detect the position and contour of specific objects in images or videos. With the development of machine vision, target detection is widely used in security monitoring, unmanned aerial vehicle inspection, industrial automation and other scenarios, and has great strategic significance for the intelligent transformation of various fields. Among them, in the fields of military reconnaissance, night vision and night combat, intelligent security, the detection technology of small unmanned aerial vehicles, floating objects and other infrared small targets plays a crucial role. However, due to the characteristics of infrared small targets, such as low pixel value, low contrast, complex background and adjacent small targets, there are many unique challenges in detecting infrared images. Traditional infrared image target detection methods have the advantages of all-weather operation and strong robustness in the face of complex backgrounds, but still face efficiency and accuracy problems when dealing with infrared small targets. Therefore, it is still a challenging task to achieve efficient and robust infrared small target detection with the aid of infrared imaging, and it is also the focus of current research by scholars.

[0003] Chinese patent publication No. "CN116468928B" entitled "Infrared small target detection method based on data expansion and weak feature enhancement" first uses an improved feature extraction network to reduce small target feature loss, and designs an efficient feature enhancement module to strengthen the target region representation without significantly increasing the computational complexity. At the same time, the method introduces a balanced sample training strategy to effectively solve the problem of extremely unbalanced positive and negative samples. Finally, the method further integrates feature information of different scales through a multi-level feature fusion structure to generate high-quality detection results.

[0004] However, this method still faces significant challenges when dealing with infrared small targets. First, the method lacks consideration of adaptability to complex background scenes, and when the complexity of target and background distribution increases, the detection performance will decrease significantly; second, the existing feature extraction backbone network has limited feature expression ability for low-contrast small targets, resulting in loss of key target features; in addition, although a sample balancing strategy is used, there is still a risk of overfitting under extremely unbalanced data, affecting the generalization ability of the model in complex environments. Therefore, designing an infrared small target detection method that can effectively deal with complex background interference, reduce target feature loss and solve the problem of overfitting of the detection model is an urgent problem to be solved by the present application. SUMMARY

[0005] The technical scheme for solving the above technical problems of the present application is to provide a dual-branch infrared small target detection method based on neural low-rank background modeling, comprising the following steps:

[0006] S1, preparing a data set: five public data sets are adopted, wherein data set one, data set two and data set three are used for network training and fine-tuning; data set four and data set five are used for model testing;

[0007] S2, constructing a dual-branch detection model: including a neural low-rank background modeling module, a dual-perception enhancement and resolution adaptive magnification module, a multi-scale contrast adaptive enhancement module and a multi-scale dynamic detector;

[0008] S3, training the network model: input the training data of S1 into the neural low-rank background modeling module to extract background features, and input the features enhanced by the dual-perception enhancement module into the fusion module; at the same time, input the training data into the dual-perception enhancement module for enhancement, and then input the enhanced data into the multi-scale contrast adaptive enhancement module for adaptive weight learning, and then input the data into the fusion module for multi-scale information fusion training;

[0009] S4, designing a loss function and evaluation index: a joint loss function composed of a physical size sensitive loss, a spatial coordinate sensitive loss and a focal loss is used to optimize the model parameters, and the performance of the model is evaluated through multi-dimensional indexes;

[0010] S5, fine-tuning the model: based on data set two, the model parameters are optimized;

[0011] S6, saving the model: solidifying the network parameters after fine-tuning to obtain the final detection model.

[0012] Further, in S1: data set one is MDvsFA data set; data set two is SIRST-Aug data set; data set three is IRSTD-1k data set; data set four is NUDT-SIRST data set; and data set five is SIRST data set.

[0013] Further, in S2, the neural low-rank background modeling module comprises:

[0014] The neural function realized by the multi-layer perception machine represents the background tensor through a parameterized factor function, and the factor function is defined in the height, width and time dimensions, respectively;

[0015] The neural regularization module applies local smoothness and temporal regularization constraints by calculating the gradient of the background tensor in three-dimensional space, and its mathematical expression is:

[0016]

[0017] Among them, and respectively represent the height, width and time dimension gradient.

[0018] Further, in the S2, the dual perception enhancement and resolution adaptive magnification module comprises:

[0019] Perception block: dynamically adjust the channel weight by convolution expansion channel followed by SE module, when the input and output channel are consistent, use residual connection;

[0020] Up-sampling module: improve the resolution by channel reorganization, using bicubic interpolation method, the scale factor is 3;

[0021] Jump connection: fuse the traditional interpolation image and the output feature of perception block, and reserve the spatial information.

[0022] Further, in the S2, the multi-scale contrast adaptive enhancement module comprises:

[0023] Feature segmentation, using multi-scale pyramid segmentation strategy, through three parallel paths and spatial attention mechanism to realize the scale separation of feature subspace;

[0024] Convolution feature extraction, through cascaded depth separable convolution group and residual connection, multi-scale features are extracted while ensuring the calculation efficiency;

[0025] Contrast enhancement, based on local contrast measure, through sliding window statistics, nonlinear gain adjustment and channel attention weighting, the feature enhancement of small target area is realized;

[0026] Feature reorganization and fusion, using cross-scale aggregation mechanism, through bilinear interpolation alignment, channel cascading fusion and spatial attention selection, the enhanced feature is output.

[0027] Further, in the S2, the multi-scale dynamic detector comprises:

[0028] Channel excitation branch, through average pooling, 1x1 convolution and activation function to realize channel level enhancement;

[0029] Channel space perception branch, through index operation, 3x3 convolution and bias adjustment to capture spatial context;

[0030] Channel re-calibration branch, through average pooling, fully connected layer and normalization to generate channel weight mapping.

[0031] Further, in the S4, the joint loss function is:

[0032] L PDSC =L PD +L SC ;

[0033] Where: L PDL is a physical size sensitive loss for small targets SC L is a spatial coordinate sensitive loss for small targets PD The weight is calculated by normalizing the target size difference, and the mathematical definition is as follows:

[0034]

[0035] Where P and G represent the pixel set of the prediction frame and the pixel set of the real frame respectively;

[0036] Spatial coordinate sensitive loss The center point distance and angle difference are calculated by polar coordinate transformation;

[0037] The focal point loss is:

[0038] L Focal =-α t (1-p t ) γ log(p t )。

[0039] Further, the evaluation index in S4 includes:

[0040] Pixel-level indicators: IoU, recall rate;

[0041] Target-level indicators: detection probability;

[0042] Comprehensive indicators: F1 score, the calculation formula is:

[0043]

[0044] Compared with the prior art, the present application has the following beneficial effects:

[0045] (1) The present application designs a new end-to-end infrared small target detection network model, which models the low-rank characteristics and spatio-temporal continuity of the background by designing a neural low-rank background modeling method, enhances the modeling ability of complex background, and develops neural regularization by applying three-dimensional total variation constraint, effectively suppresses background noise interference, and significantly improves the robustness of background modeling.

[0046] (2) The present application designs a new end-to-end infrared small target detection network model, which introduces a dual perception enhancement and resolution adaptive amplification module, which realizes adaptive enhancement of infrared images by parallel processing of channel and spatial dimension feature information.

[0047] (3)The application designs a new end-to-end infrared small target detection network model, and a multi-scale contrast adaptive enhancement module is introduced, which innovatively combines an adaptive feature segmentation mechanism and a differential convolution operation, strengthens the high-frequency details of the target region while suppressing the noise interference of the background region, thereby significantly enhancing the contrast between the target and the background while maintaining the consistency of the background.

[0048] (4)The application designs a new end-to-end infrared small target detection network model, and a multi-scale dynamic detector is introduced, which realizes adaptive weight distribution of different scale features based on an attention mechanism. The module first extracts multi-scale feature representations through a multi-branch structure, then calculates the importance weight of each scale feature using channel attention and spatial attention mechanisms, and finally generates enhanced feature representations through weighted fusion.

[0049] (5)The application designs a new end-to-end infrared small target detection network model, and a joint loss composed of a physical size sensitive loss, a spatial coordinate sensitive loss and a focal loss is introduced, which helps the detection network pay more attention to smaller target regions and provides stable gradients for the training process, making the training process more reliable.

[0050] (6)The detection model proposed in the application has good performance in NUDT-SIRST dataset and SIRST dataset, and the quantitative evaluation index has been greatly improved, which shows that the detection method proposed in the application has very strong generalization ability in single / multi-target and different complex backgrounds, and can adapt to most small target detection tasks and scenes. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description only some embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the structure shown in these drawings.

[0052] Figure 1 The step flow chart of the application is a double-branch infrared small target detection method based on neural low-rank background modeling.

[0053] Figure 2 The structure diagram of the double-branch infrared detection model constructed by the application;

[0054] Figure 3 The structure diagram of the neural low-rank background modeling module of the application;

[0055] Figure 4A structural diagram of a dual perception enhancement module and a resolution adaptive magnification module of the application;

[0056] Figure 5 A structural diagram of a multi-scale contrast adaptive enhancement module of the application;

[0057] Figure 6 A structural diagram of a multi-scale dynamic detector of the application;

[0058] Figure 7 A qualitative comparison effect diagram of the infrared small target detection method of the application and an existing method;

[0059] Figure 8 An evaluation index comparison schematic diagram of the infrared small target detection method of the application and an existing method. DETAILED DESCRIPTION

[0060] The application provides a dual-branch infrared small target detection method based on neural low-rank background modeling, aiming to solve problems such as complex background interference, target feature loss and poor model generalization ability.

[0061] The application provides a dual-branch infrared small target detection method based on neural low-rank background modeling, aiming to solve problems such as complex background interference, target feature loss and poor model generalization ability.

[0062] Embodiment 1

[0063] A dual-branch infrared small target detection method based on neural low-rank background modeling, as shown in the figure, comprises the following steps: Figure 1

[0064] S1, preparing a data set: five kinds of public data sets are adopted, wherein data set one, data set two and data set three are used for network training and fine tuning; data set four and data set five are used for model testing;

[0065] S2, constructing a dual-branch detection model: including a neural low-rank background modeling module, a dual perception enhancement and resolution adaptive magnification module, a multi-scale contrast adaptive enhancement module and a multi-scale dynamic detector;

[0066] S3, training a network model: the training data of S1 are input into the neural low-rank background modeling module to extract background features, and the features are input into a fusion module after being enhanced by the dual perception enhancement module; meanwhile, the training data are input into the dual perception enhancement module for enhancement, and then input into the fusion module for multi-scale information fusion training after adaptive weight learning by the multi-scale contrast adaptive enhancement module;

[0067] S4, designing a loss function and an evaluation index: a joint loss function composed of a physical size sensitive loss, a spatial coordinate sensitive loss and a focal loss is adopted to optimize model parameters, and the performance of the model is evaluated through multi-dimensional indexes.​

[0068] S5, fine-tuning the model: optimizing the model parameters based on the second data set;

[0069] S6, save the model: solidify the fine-tuned network parameters to obtain the final detection model.

[0070] Further, in S1: the first data set is the MD vs FA data set; the second data set is the SIRST-Aug data set; the third data set is the IRSTD-1k data set; the fourth data set is the NUDT-SIRST data set; and the fifth data set is the SIRST data set.

[0071] Further, in S2, the neural low-rank background modeling module comprises:

[0072] The neural function realized by the multi-layer perception machine represents the background tensor through a parameterized factor function, which is defined in the height, width and time dimensions, respectively;

[0073] The neural regularization module applies local smoothness and temporal regularization constraints by calculating the gradient of the background tensor in three-dimensional space, and its mathematical expression is:

[0074]

[0075] wherein, and represent the height, width and time dimension gradients, respectively.

[0076] Further, in S2, the dual perception enhancement and resolution adaptive magnification module comprises:

[0077] Perception block: through the convolution to expand the channel number and then the SE module to dynamically adjust the channel weight, when the input and output channels are consistent, the residual connection is adopted;

[0078] Up-sampling module: through channel reorganization to improve the resolution, using bicubic interpolation method, the scale factor is 3;

[0079] Skip connection: fuse the traditional interpolation image and the output feature of the perception block to retain the spatial information.

[0080] Further, in S2, the multi-scale contrast adaptive enhancement module comprises:

[0081] Feature segmentation: using a multi-scale pyramid segmentation strategy, through a three-level parallel path and a spatial attention mechanism to realize the scale separation of the feature subspace;

[0082] Convolution feature extraction: through the cascaded deep separable convolution group and the residual connection, multi-scale features are extracted while ensuring the calculation efficiency;

[0083] Contrast enhancement, based on local contrast measure, through sliding window statistics, nonlinear gain adjustment and channel attention weighting, realizes feature enhancement of small target region;

[0084] Feature recombination fusion, using cross-scale aggregation mechanism, through bilinear interpolation alignment, channel cascade fusion and spatial attention selection, outputs enhanced features.

[0085] Further, in S2, the multi-scale dynamic detector comprises:

[0086] Channel excitation branch, through average pooling, 1x1 convolution and activation function to realize channel level enhancement;

[0087] Channel space perception branch, through index operation, 3x3 convolution and bias adjustment to capture spatial context;

[0088] Channel re-calibration branch, through average pooling, fully connected layer and normalization to generate channel weight mapping.

[0089] Further, in S4, the joint loss function is:

[0090] L PDSC =L PD +L SC ;

[0091] Wherein: L PD is the physical size sensitive loss of small target, L SC is the spatial coordinate sensitive loss of small target; L PD The weight is calculated by normalizing the target size difference, which is mathematically defined as:

[0092]

[0093] Where, P and G represent the pixel set of the prediction box and the pixel set of the real box respectively;

[0094] Spatial coordinate sensitive loss The center point distance and angle difference are calculated by polar coordinate transformation;

[0095] The focal loss is:

[0096] L Focal =-α t (1-p t ) γ log(p t )。

[0097] Further, the evaluation index in S4 comprises:

[0098] Pixel-level index: IoU, recall rate;

[0099] Target level indicators: detection probability;

[0100] Overall indicator: F1 score, calculated as:

[0101]

[0102] Example 2:

[0103] A flowchart of an infrared small target detection method based on data augmentation and weak feature enhancement, the method specifically comprising the following steps:

[0104] S1, preparing a data set: MD vs FA data set one is the first public infrared small target data set, containing 11000 infrared images, part of which are synthetic images. SIRST-Aug data set two contains 9070 images, which is based on data set SIRST, and has undergone data augmentation (cropping, rotation, displacement, etc.). IRSTD-1k data set three contains 1001 images, covering multiple scenes and target distribution, and the target has high sparsity. NUDT-SIRST data set four contains 1,327 infrared small target images in complex background, and the target size and background contrast are low. SIRST data set five contains 427 infrared images and 480 targets, which is the first real infrared small target data set with high-quality images and labels.

[0105] S2, constructing a dual-branch infrared detection model: the detection model is as shown in Figure 2 The model introduces a dual-branch design: branch one uses a ResNeSt feature extraction module to extract deep features, and through a dual perception enhancement and resolution adaptive magnification module and a multi-scale contrast adaptive enhancement module, the recognition accuracy of small targets is improved; branch two uses a neural low-rank background modeling method for preprocessing to suppress background interference, and through a dual perception enhancement and resolution adaptive magnification module, the target details are further strengthened. Subsequently, the two path decoders are composed of a fusion module, which uses inverse wavelet transform to perform multi-scale feature fusion, integrates features from different network levels, and combines low-level details and high-level semantic information together.

[0106] Neural low-rank background modeling module, as shown in Figure 3As shown: the neural function of the multi-layer perceptron realizes the low-rank prior, the tensor of the image is obtained as the background, that is, the background is represented by the parameterized function of the multi-layer perceptron; then the neural three-dimensional total variation regularization uses the output derivative of the multi-layer perceptron to enforce the local smoothness and temporal regularization in the background component, so as to achieve the purpose of outputting the background; in the background modeling process, the low-rank characteristic is the core of capturing the background redundancy information; however, the traditional low-rank decomposition method (such as matrix decomposition and tensor decomposition) often performs weakly in dealing with nonlinear background changes, in order to overcome this deficiency, the embodiment introduces a neural low-rank tensor representation, which models the background through the nonlinear expression ability of the neural network. Specifically, the estimation of the background tensor can be represented as the following optimization objective:

[0107]

[0108] Where B represents the background tensor, and ||B|| is the nuclear norm. * is the nuclear norm, is the gradient regularization term in the time domain, which is used to capture the smoothness of the background in the time dimension. a is the regularization weight, which is used to balance the nuclear norm and the time gradient regularization term. In order to further enhance the expression ability of the background modeling, a neural function f θ is used to model the background tensor, which is in the form of:

[0109]

[0110] Where B(i,j,k) represents the dimension index of the background tensor, and are factor functions defined in the height, width and time dimensions, and each factor function is implemented by a multi-layer perceptron.

[0111] Although the low-rank tensor representation has significant advantages in capturing the global characteristics of the background, its modeling ability for local changes is relatively limited. In the background, local changes often have high complexity and uncertainty, and it is difficult to achieve accurate modeling of these details only by relying on low-rank constraints. Therefore, a neural three-dimensional total variation regularization method is proposed to explicitly enhance the local smoothness of the background in the spatial and temporal dimensions while preserving the target information. The mathematical definition of the neural three-dimensional total variation regularization is as follows:

[0112]

[0113] Where, and represent the gradients of the background tensor in the height, width and time dimensions, respectively. By constraining the local gradient of the background, the neural three-dimensional total variation regularization N3DTV can explicitly enhance the smoothness of the background while avoiding the loss of target signals caused by excessive smoothing.

[0114] Dual perceptual enhancement and resolution adaptive upscaling module, as shown in Figure 4 (1), which consists of three main components: perceptual block, up-sampling block, and skip connection. As shown in Figure 4 (2), the perceptual block aims to enhance feature perception and detail enhancement. It first expands the channel number through convolution to enhance the model's expressive power, and then uses the SE block to dynamically adjust the channel weight to highlight key features. If the input and output channel numbers match, a residual connection is used to maintain information flow. The up-sampling technique improves image resolution by reorganizing the input feature map channels. Its advantage is that it can preserve spatial relationships and reduce artifacts, which is important for improving infrared image resolution. The skip connection combines the adjusted image from the traditional method and the enhanced features from the perceptual block, helping the network better learn the target features while preserving spatial information and improving overall perceptual ability. The bicubic interpolation method is used in this embodiment, and the scale factor is 3.

[0115] Multi-scale contrast adaptive enhancement module, as shown in Figure 5 The multi-scale contrast adaptive enhancement module is the core component of the dual-branch infrared detection model, designed specifically to address the challenges of low contrast and variable target shape in infrared small target detection. This module effectively enhances and extracts small target information by segmenting and processing input features, significantly improving detection accuracy and robustness. Specifically, the module enhances small target features through three steps: first, segment the input features and enhance the contrast; second, extract convolutional features; and third, fuse contrast enhancement and feature reorganization. This process significantly improves detection accuracy and robustness, enabling the dual-branch infrared small target detection model to more accurately detect infrared small targets in complex backgrounds.

[0116] Multi-scale dynamic detector, as shown in Figure 6 The dynamic detector can adaptively focus on scale-space task information, better learn the relative importance of target semantic levels and spatial information, and adapt to different task forms. Specifically, the dynamic detector can be represented as:

[0117] W(F) = f T (f Sp (f Sc (F)·F)·F)·F;

[0118] where f Sc , f Sp , and f T represent attention functions on scale, space, and task, respectively; scale attention f Sc (F)·F is based on the importance of each layer of features to achieve dynamic feature fusion:

[0119]

[0120] where f(·) is a 1x1 convolution layer and σ(·) is a sigmoid function; spatial-aware attention f Sp (F) utilizes deformable convolution to fuse features from different levels at the same spatial location:

[0121]

[0122] where K is the number of sparse sampling locations, p k +Δp k and Δm k are learned from input features, p k +Δp k is a self-learned spatial offset, Δm k is an important scalar for self-learning at p k location. Task-aware attention f T (F) dynamically switches channels to support different tasks:

[0123] f T (F)·F = max(α 1 (F)·F c +β 1 (F), α 2 (F)·F c +β 2 (F));

[0124] where [α 1 ,α 2 ,β 1 ,β 2 ] are hyper functions to control thresholds, which are normalized by a sigmoid activation function after average pooling, two fully connected layers and a normalization layer.

[0125] S3, training the detection network model: input the data set prepared in step S1 into the detection network model constructed in step S2 for training. The above network architecture and training process are implemented using Python 3.10 and PyTorch 2.3, and efficient parallel computing is performed on NVIDIA Titan Xp GPU. During the training process, the Adam optimizer is selected, and the parameters λ1=0.5 and λ2=0.999 are set to accelerate the convergence process and reduce the shock. The number of training iterations is set to 500, and the batch size is set to 16. In the selection of optimization methods, the MultiStepLR learning rate decay strategy is adopted, which dynamically adjusts the learning rate to provide a larger learning rate in the early stage of training to accelerate convergence, and then gradually reduces it to avoid overfitting. The learning rate is set to 5x10 -4, the decay factor is set to 0.1. During the training process, validation is performed on the IRSTD-1k dataset after each training round.

[0126] S4, select a suitable loss function and determine the optimal evaluation index of the method: the loss function is designed for the physical size and spatial coordinates of the infrared small target, which solves the traditional L IoU and L Dice loss shows certain limitations in detecting small targets, and its mathematical definition is:

[0127] L PDSC =L PD +L SC ;

[0128] Where L PD is the physical size sensitive loss of the small target, and L SC is the spatial coordinate sensitive loss of the small target. Then, for L PD By weighting L Dice loss:

[0129]

[0130] Where P and G represent the predicted frame pixel set and the real frame pixel set respectively, and the mathematical definition of the weight ω is:

[0131]

[0132] Where Var is a scalar variance function. The greater the difference between P and G, the smaller ω will be, resulting in a greater physical size sensitive loss. The basic principle behind the design of ω is that if there is a significant difference between P and G, the detector should allocate more attention to the loss of the target, that is, when the physical size difference between P and G is large, the weight ω will increase, thus imposing a higher penalty on the target with a larger physical size deviation.

[0133] For L SC , the corresponding center points of P and G are obtained by averaging the coordinates of all pixels. Then, the coordinates of these two center points are transformed into the polar coordinate system. For the corresponding center points, the corresponding distance and angle in the polar coordinate system are:

[0134]

[0135] The spatial coordinate sensitive loss L SC is calculated as follows:

[0136]

[0137] To solve the problem of class imbalance in the target-sparse infrared small target dataset, the focal loss is adopted. The focal loss reduces the loss contribution of easy-to-classify samples and increases the loss contribution of difficult-to-classify samples, so that the model pays more attention to small targets that are difficult to detect.

[0138] L Focal =-α t (1-p t ) γ log(p t );

[0139]

[0140] where p t is the probability of being predicted as positive, α t is the weight parameter for balancing positive and negative samples, which has different values for positive samples (targets) and negative samples (backgrounds). γ is the focusing parameter for adjusting the loss of difficult-to-classify samples. α t and γ are set to default values of 0.25 and 2, respectively.

[0141] In S4, the appropriate loss function is selected and the optimal evaluation index of the method is determined. The intersection over union (IoU) and recall rate of the predicted region and the real region are used as pixel-level evaluation indicators, and the detection probability is used to evaluate the target-level performance. These different indicators provide insights into the performance of the detector from different angles. Recall rate and detection probability focus on recall and false alarm, while IoU considers both aspects. The mathematical definition of IoU is:

[0142]

[0143] The mathematical definition of recall rate is:

[0144]

[0145] where P ture is the count of true positive pixels, and P all is the total number of pixels in the image.

[0146] The detection probability is given by:

[0147]

[0148] where N ture is the number of correctly predicted targets, and N all is the total number of targets. In addition, the harmonic mean of precision and recall, F1-score, is also considered in the evaluation. Through the joint analysis of these indicators, the detection performance of the dual-branch infrared small target detection model and its performance in different scenarios can be comprehensively analyzed.

[0149] S5, Fine-tuning the model: The model was retrained and fine-tuned using three pairs of IRSTD-1k datasets with a learning rate of 0.005 and a total of 500 iterations. Other parameters remained unchanged to further improve the performance of the detection network.

[0150] S6, Save the model: After training is completed in step S3, solidify the fine-tuned network parameters. In step S5, fine-tune the model to determine the final detection network model. If performing infrared small target detection tasks, the infrared small target image can be directly input into the trained end-to-end network model to obtain the final prediction result.

[0151] Qualitative and quantitative comparison of existing technologies and the method proposed in this invention, such as... Figure 7 , Figure 8 As shown, where Figure 8 The bold white boxes represent accurately detected small infrared targets, the dashed white boxes represent false detections, and the white boxes represent missed detections. As shown in the figure, the method proposed in this invention has a higher normalized intersection-union ratio and detection rate than existing methods, achieving a better balance between precision and recall. These metrics further demonstrate that the method proposed in this invention achieves superior performance in detecting small infrared targets and meets the expected results. Method 1 (ASTTV-NTLA) employs an adaptive spatiotemporal total variational and nonlocal attention mechanism, achieving good background suppression at 87.00% accuracy. Method 2 (U-Net) is based on a classic encoder-decoder architecture, with balanced but relatively mediocre performance across metrics. Method 3 (ALCNet) leads with 96.93% accuracy through attention-guided local contrast enhancement, but its recall fluctuates. Method 4 (ISNet) utilizes an iterative saliency detection mechanism, exhibiting stable segmentation performance at an F1-score of 88.82%. In comparison, the method of this invention significantly outperforms the comparative method in both key metrics of recall (89.20%) and IoU (84.55%). In particular, the IoU metric is nearly 7 percentage points higher than the best comparative method (ALCNet), demonstrating a more accurate target localization capability.

[0152] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A dual-branch infrared small target detection method based on neural low-rank background modeling, characterized in that, The method comprises the following steps: S1, preparing a data set: five public data sets are adopted, wherein data set one, data set two and data set three are used for network training and fine tuning; Data set four and data set five are used for model testing; S2, constructing a double-branch detection model: including a neural low-rank background modeling module, a dual perception enhancement and resolution adaptive magnification module, a fusion module, a multi-scale contrast adaptive enhancement module and a multi-scale dynamic detector; S3, training the network model: the training data of S1 is input into the neural low-rank background modeling module to extract background features, and the features are enhanced through the dual perception enhancement and resolution adaptive magnification module and then input into the fusion module; meanwhile, the training data is input into the dual perception enhancement and resolution adaptive magnification module for enhancement, then adaptive weight learning is performed through the multi-scale contrast adaptive enhancement module, and then multi-scale information fusion training is performed through the fusion module; S4, designing a loss function and an evaluation index: a joint loss function composed of a physical size sensitive loss, a spatial coordinate sensitive loss and a focal loss is adopted to optimize the model parameters, and the performance of the model is evaluated through multi-dimensional indexes; S5, fine tuning the model: based on data set two, the model parameters are optimized; S6, saving the model: the network parameters after fine tuning are solidified to obtain a final detection model; In S4, the joint loss function is: L PDSC = L PD + L SC ; where: L PD is the physical dimension sensitive loss for small targets, L SC is the spatial coordinate sensitive loss for small targets; L PD The weight is calculated by normalizing the target size difference, mathematically defined as: Wherein, P and G represent the pixel set of the prediction box and the pixel set of the real box respectively; Spatial coordinate sensitive loss The center point distance and angle difference are calculated by polar coordinate transformation; The focal loss is: L Focal = -a t (1 - p t ) γ log(p t ).

2. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, In S1: data set one is MDvsFA data set; data set two is SIRST-Aug data set; data set three is IRSTD-1k data set; data set four is NUDT-SIRST; and data set five is SIRST data set.

3. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, In S2, the neural low-rank background modeling module comprises: A neural function realized by a multi-layer perception machine, which represents a background tensor through a parameterized factor function, and the factor function is defined in height, width and time dimensions respectively; A neural regularization module, which applies local smoothness and time regularization constraints by calculating the gradient of the background tensor in three-dimensional space, and its mathematical expression is: wherein and respectively represent height, width and time dimension gradients.

4. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, In S2, the dual perception enhancement and resolution adaptive magnification module comprises: A perception block: after expanding the channel number through convolution, the channel weight is dynamically adjusted through an SE module, and when the input and output channels are consistent, residual connection is adopted; An up-sampling module: the resolution is improved through channel reorganization, and a double cubic interpolation method is adopted with a scale factor of 3; A skip connection: the traditional interpolation image and the output feature of the perception block are fused to retain spatial information.

5. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, In S2, the multi-scale contrast adaptive enhancement module comprises: Feature segmentation: the input feature is processed through 3*3 convolution and double 1*1 convolution branches in parallel, and after splicing and fusion, the dimension is reduced to output, realizing multi-scale feature extraction; Convolution feature extraction: containing three branch convolution structures and introducing residual connection, the original information is retained while deep layer features are extracted; Contrast enhancement: based on a local contrast measure meter, through sliding window statistics, nonlinear gain adjustment and channel attention weighting, the feature enhancement of small target regions is realized; Feature reorganization and fusion: after splicing the features at all levels, the features are compressed and integrated through 1*1 convolution, and finally the target dimension features are output.

6. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, In the S2, the multi-scale dynamic detector comprises: a channel excitation branch, which realizes channel-level enhancement through average pooling, 1*1 convolution and an activation function; a channel space perception branch, which captures spatial context through index operation, 3*3 convolution and bias adjustment; a channel re-calibration branch, which generates channel weight mapping through average pooling, a fully connected layer and normalization.

7. The dual-branch infrared small target detection method based on neural low-rank background modeling according to claim 1, characterized in that, The evaluation indexes in the S4 comprise: pixel-level indexes: IoU, recall rate; target-level indexes: detection probability; comprehensive indexes: F1 score, and the calculation formula is:

Images