An infrared salient target detection method based on improved YOLO11

Abstract

This invention provides an infrared salient target detection method based on an improved YOLO11, belonging to the field of computer vision and target detection technology, and applicable to infrared target recognition in low-light and complex backgrounds. Addressing the inherent defects of infrared images, the incompatibility of YOLO11 with infrared thermal radiation characteristics, and the scarcity of labeled data, this invention achieves accurate and rapid detection through multi-stage optimization: It constructs a dedicated infrared dataset, preprocessing it through denoising and multi-scale enhancement to adapt it to network input; it improves YOLO11 by embedding a channel-spatial attention mechanism, constructing a cross-scale feature pyramid, and employing adaptive anchor boxes; it introduces infrared target feature cosine similarity loss to construct a composite loss function, combined with a preprocessing-detection linkage mechanism and lightweight optimization. This invention effectively solves the problems of data scarcity and model adaptation, improves the model's feature extraction and recognition capabilities for infrared targets, and achieves a balance between detection accuracy and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to an infrared salient target detection method based on an improved YOLO11, which is suitable for infrared image target recognition in low illumination and complex backgrounds. Background Technology

[0002] Infrared target detection technology, with its stable imaging capabilities in all weather and low-light environments, has demonstrated significant advantages in scenarios such as security monitoring, nighttime navigation, and autonomous driving. However, infrared images themselves have inherent defects such as generally low contrast, blurred target edges, and a lack of texture information, which pose a serious challenge to high-precision target identification and localization. Especially under complex background interference, the difference in radiation between the target and the environment is small, further increasing the difficulty of detection and often leading to missed detections or false detections using traditional detection methods.

[0003] The YOLO11 model, widely used in the visible light domain, achieves a good balance between detection speed and accuracy. However, its feature extraction mechanism, the scale distribution of pre-defined anchor boxes, and the design of its loss function are primarily optimized for the color, texture, and contour information of visible light images, without fully considering the thermal radiation characteristics and multi-scale distribution patterns of infrared targets. Directly transferring this model to infrared scenes makes it difficult to effectively capture the key features of infrared targets, significantly limiting detection accuracy. Furthermore, the relative scarcity of high-quality, large-scale infrared labeled data severely restricts the model's training effectiveness and generalization ability. Therefore, there is an urgent need to make targeted improvements to YOLO11 for infrared sensing characteristics. Through architectural adjustments, loss function optimization, and data augmentation, its detection accuracy and real-time performance for salient infrared targets in complex environments can be comprehensively improved. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by providing an infrared salient target detection method based on an improved YOLO11, which solves the difficulties of infrared data scarcity and the difficulty of balancing detection accuracy and speed, and enables accurate and rapid identification of infrared targets in complex environments.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0006] A method for detecting salient infrared targets based on an improved YOLO11, the specific steps of which are as follows:

[0007] S1: Construct an image dataset;

[0008] Based on publicly available infrared image datasets, an image set covering various typical infrared scenes is constructed. All images in the constructed image set are formatted in a unified manner, and fine-grained ground truth annotations are performed based on salient targets. The annotation information includes the target category and the normalized coordinates of its bounding box. Then, the dataset is divided into training set, validation set, and test set to provide a data foundation for subsequent model training.

[0009] S2: Image preprocessing;

[0010] Based on the feature extraction method of the C2f module in the YOLO11 model, preprocessing operations are performed on the images in the image dataset constructed in step 1, including noise suppression, enhancement processing, and adaptation adjustment, to meet the input requirements of the improved YOLO11 model;

[0011] S3: Construct an infrared target detection model; make multi-dimensional improvements to the YOLO11 basic network architecture based on the physical characteristics of infrared targets;

[0012] The infrared target detection model is based on the YOLO11 infrastructure configuration file provided by the Ultralytics framework. Specifically, it includes: embedding a cascaded channel-spatial attention mechanism in the feature extraction path of the backbone network; constructing a cross-scale feature pyramid structure in the feature fusion module of the neck; removing the original fixed anchor boxes of the detection head and using the K-Means++ clustering algorithm to generate adaptive anchor boxes according to the target scale distribution of the infrared dataset.

[0013] S4: Model training and testing;

[0014] The model is trained using a cosine annealing learning rate scheduling strategy based on the training and test sets, and an early stopping mechanism is set to prevent overfitting. During the training process, based on the infrared target classification and localization error analysis, a composite loss function is constructed by introducing infrared target feature cosine similarity loss on the basis of YOLO11 default loss CIoU localization loss and cross-entropy classification loss, and the model performance is monitored.

[0015] The trained model is lightweighted and a structured pruning strategy is adopted to remove redundant convolutional kernels and channels with absolute weight values ​​below the threshold and low feature contribution. Short-term fine-tuning is performed after pruning.

[0016] Finally, the preprocessed infrared image test set is input into the trained and optimized model to obtain the final detected target, thus realizing real-time target detection of infrared images.

[0017] Furthermore, the preprocessing in step 2 specifically includes:

[0018] Adaptive weighted noise suppression is performed; by analyzing the variance and gradient features of local image regions, higher filtering weights are dynamically calculated and assigned to noisy pixels, and Gaussian noise and salt-and-pepper noise are filtered out using an adaptive filtering algorithm, preserving the edge and texture information of the target while denoising.

[0019] The Retinex multi-scale enhancement algorithm is used for enhancement processing. The multi-scale Retinex algorithm is used with scale factors of 15 as small scale, 80 as medium scale and 250 as large scale. It separates the illumination component and reflectance component of the image. By adjusting the reflectance component, the contrast between the target and the background is enhanced and the texture details are strengthened. For scenes with different characteristics, the enhancement results of different scales are adaptively selected or fused.

[0020] Adaptability adjustments are made; the size of the preprocessed image is uniformly scaled to the standard input size of the YOLO11 network, and the pixel grayscale values ​​are normalized and mapped to the [0, 1] interval.

[0021] Furthermore, the channel-spatial attention module in step 3 first analyzes the importance of each channel in the feature map through channel attention branches, filters and enhances useful feature channels that are strongly correlated with the thermal radiation characteristics of the infrared target, and suppresses information from redundant background channels; by using large-kernel deep convolution to focus on the spatial dimension of the feature map, it strengthens the response of the region where the infrared target is located and further suppresses background interference; the attention in the two stages works together to make the network more focused on the infrared salient target.

[0022] The cross-scale feature pyramid structure adds a direct interaction channel between features of non-adjacent scales to the original adjacent scale feature fusion path, and achieves adaptive fusion of feature maps of different resolutions through lightweight convolution operations.

[0023] Furthermore, in step 4, the CIoU localization loss and the cross-entropy classification loss are combined and denoted as... The format is as follows:

[0024] ;

[0025] in, It is the CIoU localization loss, used to optimize the position and size of the bounding box; It is the cross-entropy classification loss, used to optimize target category prediction; and These are the weighting coefficients for the two losses; and The calculation formulas are as follows:

[0026] ;

[0027] ;

[0028] in, This is the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box. The Euclidean distance between the center of the predicted bounding box and the center of the ground truth bounding box; The center coordinates of the prediction box are in the form of ( , ); The center coordinates of the true bounding box are in the form of ( , ); The diagonal length of the smallest closure rectangle that can simultaneously cover both the predicted and ground truth boxes; As a measure of aspect ratio consistency, The square of the difference between two angles is used to measure the consistency of the predicted bounding box and the ground truth bounding box in terms of aspect ratio. The smaller the value, the closer the aspect ratios are. The width of the actual bounding box. The actual height of the bounding box. The width of the prediction box. The height of the predicted bounding box; It is a balance parameter. ; Total number of categories; One-hot encoding for the real label; The predicted class probabilities by the model;

[0029] The cosine similarity loss of infrared target features is denoted as This is used to constrain the model to learn feature representations that conform to infrared physical characteristics, and the formula is:

[0030] ;

[0031] in, This represents the number of samples predicted as the target in the current batch. The first extracted from the model The feature vector of the prediction target This is the corresponding prior infrared target standard feature vector; The L2 norm of a vector;

[0032] The composite loss function is:

[0033] ;

[0034] in, and The weighting coefficients are used to balance the two losses.

[0035] The preprocessing-model detection linkage mechanism dynamically feeds back and adjusts the parameters of noise suppression intensity, Retinex enhancement weights, and adaptive anchor boxes in the image preprocessing stage based on the confidence distribution, false detection rate, and regression error index of the model's detection results on the validation set, thereby achieving collaborative optimization of preprocessing and model.

[0036] Furthermore, in step 4, the structured pruning strategy is specifically channel-kernel level structured pruning, with the following steps:

[0037] Step 4.1: Preparations before pruning;

[0038] Based on the trained infrared target detection model, the importance metrics of each convolutional kernel and feature channel in the model network are calculated. The core metrics include:

[0039] Absolute weight value; the smaller the absolute value of the weights of the convolution kernel, the lower its contribution to feature extraction.

[0040] Feature contribution, which is the proportion of the response value of the output feature map of a channel in the infrared target region. The lower the proportion, the smaller the contribution of the corresponding channel to the extraction of infrared target features.

[0041] The correlation of infrared thermal radiation characteristics is the correlation coefficient between the channel output and the mean and gradient of the grayscale of the infrared target's thermal radiation characteristics. The lower the correlation, the more likely it is to be pruned.

[0042] Step 4.2: Determine the pruning threshold and perform pruning;

[0043] Set the pruning thresholds: set the absolute value of the weight threshold to 0.03, the feature contribution threshold to 5%, and the infrared feature correlation threshold to 0.1.

[0044] Perform pruning: Remove redundant convolutional kernels and their corresponding feature channels that simultaneously satisfy the following conditions: absolute weight value less than 0.03, feature contribution less than 5%, and infrared feature correlation less than 0.1; the pruning scope includes:

[0045] Backbone network: Remove convolutional kernels that extract redundant background features, and retain kernels that extract infrared target thermal radiation and edge features;

[0046] Channel-Spatial Attention Module: Only channels strongly correlated with the thermal radiation characteristics of infrared targets are retained, while redundant channels for background suppression are removed;

[0047] Detection head: Removes redundant convolution kernels after simplifying the structure, further reducing the computational load of inference;

[0048] The pruning principle is: do not prune the core channels of the cross-scale feature pyramid to ensure the model's ability to detect infrared multi-scale targets after pruning;

[0049] Step 4.3: Short-term fine-tuning after pruning, specifically:

[0050] Fine-tuning dataset: Using the validation set of the infrared dataset eliminates the need to reuse the entire training set, improving fine-tuning efficiency;

[0051] Fine-tuning parameters: Set the number of fine-tuning rounds to 20, the learning rate to 1e-5, and the optimizer to AdamW;

[0052] Fine-tuning objective: Only update the weights of the remaining convolutional kernels and channels after pruning, allowing the model to re-adapt to infrared target features and restore the detection accuracy lost due to pruning.

[0053] On the other hand, this application proposes an electronic device, including: one or more processors, and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the infrared salient target detection method based on the improved YOLO11.

[0054] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the described infrared salient target detection method based on the improved YOLO11.

[0055] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the aforementioned infrared salient target detection method based on the improved YOLO11.

[0056] The beneficial effects of adopting the above technical solution are as follows: The infrared salient target detection method based on the improved YOLO11 provided by this invention solves the pain point of scarce infrared datasets through high-quality image generation technology from visible light to infrared; the scene-specific improvement of the YOLO11 architecture enhances the model's feature extraction and recognition capabilities for infrared targets; and the preprocessing-detection linkage mechanism and lightweight optimization achieve a balance between detection accuracy and real-time performance. The 200-round training setting ensures sufficient model convergence while avoiding overfitting caused by overtraining, effectively adapting to the deployment requirements of embedded devices and possessing broad application prospects. Attached Figure Description

[0057] Figure 1 The flowchart illustrates the infrared salient target detection method based on the improved YOLO11 provided in the first embodiment of the present invention. Detailed Implementation

[0058] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0059] Example 1:

[0060] An infrared salient target detection method based on an improved YOLO11, such as... Figure 1 As shown, the method of this embodiment is described below.

[0061] S1: Dataset construction.

[0062] Based on publicly available infrared image datasets, such as FLIR and LSOTB-TIR, an image set covering various typical infrared scenes is constructed. All images in this set are formatted uniformly, and fine-grained ground truth annotations are performed based on the location of salient targets such as pedestrians and vehicles. The annotation information includes the target category and the normalized coordinates of its bounding box. This results in a dedicated dataset for infrared salient target detection that is of moderate size, diverse in scenes, and standardized in annotation, providing a reliable data foundation for subsequent model training and performance evaluation.

[0063] This embodiment collected a total of 2,000 visible light images covering various scenes and lighting conditions. To increase data diversity, 2,000 real infrared images from the FLIR and KAIST public datasets were added.

[0064] All images were labeled according to the YOLO11 standard format, with annotation information including the normalized center coordinates (x_center, y_center), width (w), height (h), and class label of the target bounding box. Finally, hybrid data augmentation techniques were applied, including random scaling, grayscale perturbation, Gaussian noise simulation, horizontal flipping, and brightness / contrast adjustment. The total dataset of 4000 images was then divided into training, validation, and test sets in an 8:1:1 ratio.

[0065] S2: Image preprocessing. To address the characteristics of infrared images, such as low contrast, high noise levels, and blurred target edges, a collaborative preprocessing workflow is designed to improve image quality and adapt to network input.

[0066] The following preprocessing operations are performed sequentially on the images in the constructed infrared dataset:

[0067] Adaptive weighted noise suppression is performed. By analyzing the variance and gradient features of local image regions, higher filtering weights are dynamically calculated and assigned to noisy pixels. An adaptive filtering algorithm is used to effectively filter out Gaussian noise and salt-and-pepper noise, preserving the edge and texture information of the target as much as possible while denoising.

[0068] Multi-scale Retinex enhancement. A multi-scale Retinex algorithm with scale factors of 15 (small scale), 80 (medium scale), and 250 (large scale) is employed to separate the illumination and reflectance components of the image. By adjusting the reflectance component, the contrast between the target and the background is significantly enhanced, and texture details are strengthened. Enhancement results from different scales can be adaptively selected or fused for scenes with varying characteristics.

[0069] Adaptability adjustments were made. The preprocessed image size was uniformly scaled to the standard input size of the YOLO11 network, 640×640 pixels, and the pixel grayscale values ​​were normalized and mapped to the [0, 1] interval.

[0070] After the above process, a high-quality infrared image is output, so that its feature distribution is accurately matched with the C2f module of the YOLO11 backbone network.

[0071] The specific workflow of the C2f module is as follows:

[0072] After receiving shallow details such as the target thermal radiation edge and local high temperature points from the infrared image output by the preceding network, the number of channels is doubled through 1×1 convolution to broaden the dimension of feature expression and adapt to the extraction requirements of multi-dimensional thermal radiation features of infrared targets.

[0073] The feature map is divided into two parallel branches. One is a direct path, which directly preserves the original infrared features to avoid the loss of weak thermal radiation details and target edge information of small infrared targets. The other is a processing path, which gradually extracts the thermal radiation distribution pattern, overall target outline, and category-specific thermal features of infrared targets through n cascaded bottleneck modules. The output features of each bottleneck module are retained, realizing the gradient extraction of multi-level infrared features.

[0074] The original shallow features of the direct pathway are concatenated and fused with the deep semantic features output by all Bottleneck modules in the channel dimension. Then, the multi-channel features are integrated through 1×1 convolution and restored to the target output dimension, resulting in a high-purity feature map that combines the integrity of infrared shallow details with the recognizability of deep semantics.

[0075] S3: Model Construction. Based on the YOLO11 basic network architecture, and considering the physical characteristics of infrared targets, including the thermal radiation law which can be characterized as the gray-level mean, variance, and gradient of the image, multi-dimensional improvements are made to construct a dedicated infrared target detection model.

[0076] Modifications were made based on the YOLO11 infrastructure configuration file provided by the Ultralytics framework. Specifically, these included: embedding a channel-spatial attention module after the C3k2 module in the backbone network; adding a cross-scale feature pyramid structure to the feature fusion module in the neck region; and removing the original fixed anchor boxes from the detection head and using the K-Means++ clustering algorithm to generate adaptive anchor boxes based on the target scale distribution of the infrared dataset.

[0077] A cascaded channel-spatial attention mechanism is embedded in the feature extraction path of the backbone network. This mechanism first analyzes the importance of each channel in the feature map through a channel attention branch, filtering and enhancing useful channels strongly correlated with the thermal radiation characteristics of the infrared target, while suppressing information from redundant channels such as the background. By employing a large-kernel depthwise convolution to focus on the spatial dimension of the feature map, the response of the region containing the infrared target is strengthened, further suppressing background interference. The two-stage attention works collaboratively, enabling the network to focus more intently on salient infrared targets.

[0078] In the neck feature fusion module, a cross-scale feature pyramid structure is constructed. Based on the existing adjacent-scale feature fusion path, a direct interaction channel between features of non-adjacent scales is added. Adaptive fusion of feature maps at different resolutions is achieved through lightweight convolutional operations, effectively mitigating the problem of feature information attenuation for small infrared targets in deep networks and improving the model's detection capability for multi-scale targets, especially small targets.

[0079] The default fixed anchor box size of YOLO11, based on a general dataset, was removed. For the constructed infrared dataset, the K-Means++ clustering algorithm was used to analyze the true scale distribution of target bounding boxes, generating multiple sets of anchor boxes adapted to typical infrared target scales. An adaptive anchor box adjustment mechanism was introduced, which can dynamically fine-tune anchor box parameters based on regression errors during training. Simultaneously, the detector head structure was simplified, removing redundant convolutional layers to improve the model's inference speed.

[0080] S4: Model training and prediction.

[0081] The improved model was trained using the dataset constructed in step S1, with a total of 200 training epochs. A cosine annealing learning rate scheduling strategy was used for model training, and an early stopping mechanism was set to prevent overfitting.

[0082] The cosine annealing learning rate scheduling strategy takes the following form:

[0083] Initialization parameters: Set the initial learning rate, minimum learning rate, and total training epochs to 200; do not set a restart mechanism.

[0084] Cosine decay calculation: During training, the learning rate for each round is dynamically calculated using the cosine function formula, which is:

[0085] ;

[0086] in: Let be the learning rate in round t. The initial maximum learning rate, To minimize the learning rate, For the current training round, Total training rounds (200 rounds).

[0087] During training, model performance was monitored using validation set loss. Based on infrared target classification and localization error analysis, the model was trained using the default YOLO11 loss CIoU localization loss and cross-entropy classification loss (combined as CIoU and cross-entropy classification loss). Based on this, cosine similarity loss of infrared target features is introduced. We construct a composite loss function to jointly optimize the target's localization, classification accuracy, and the physical rationality of feature representation.

[0088] It includes CIoU localization loss and cross-entropy classification loss, in the following form:

[0089] .

[0090] in, It is the CIoU localization loss, used to optimize the position and size of the bounding box; It is the cross-entropy classification loss, used to optimize target category prediction; and These are the weighting coefficients for the two losses. and The calculation formulas are as follows:

[0091] ;

[0092] .

[0093] in, This is the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box. The Euclidean distance between the center of the predicted bounding box and the center of the ground truth bounding box; The center coordinates of the prediction box (a two-dimensional vector, in the form of ( , )); The center coordinates of the ground truth (a two-dimensional vector, in the form of ( , )); The diagonal length of the smallest closure rectangle that can simultaneously cover both the predicted and ground truth boxes; As a measure of aspect ratio consistency, ,in, The width of the actual bounding box. The actual height of the bounding box. The width of the prediction box. The height of the predicted bounding box. v measures the similarity in aspect ratio between the predicted bounding box and the ground truth bounding box by the square of the difference between the two angles; the smaller the value, the closer the aspect ratio. It is a balance parameter. . Total number of categories; One-hot encoding for the real label; This represents the class probabilities predicted by the model.

[0094] Cosine similarity loss of infrared target features The cosine similarity between the predicted target features and the prior infrared target standard features is calculated to constrain the model to learn feature representations that conform to infrared physical characteristics. The formula is as follows:

[0095] ;

[0096] in, This represents the number of samples predicted as the target in the current batch. The first extracted from the model The feature vector of the prediction target The corresponding prior infrared target standard feature vector is usually derived from a pre-trained infrared feature extraction network or cluster center; This represents the L2 norm of a vector.

[0097] Finally, a composite loss function is constructed:

[0098] ;

[0099] in, and To balance the weighting coefficients of the two losses, the value ranges in this embodiment are as follows: , .

[0100] The improved model was trained end-to-end using the constructed infrared dataset. During training, a preprocessing-model detection linkage mechanism was built: based on the confidence distribution and false detection rate of the model's detection results on the validation set, the parameters of the image preprocessing stage, such as noise suppression intensity, Retinex enhancement weights, and adaptive anchor boxes, were dynamically fed back and adjusted to achieve collaborative optimization of the preprocessing and detection models and improve the overall adaptability of the system.

[0101] The trained model is lightweighted by using a structured pruning strategy to remove redundant convolutional kernels and channels with absolute weight values ​​below a threshold (e.g., 0.03) and low feature contribution. Short-term fine-tuning is then performed after pruning to restore accuracy.

[0102] The specific implementation of the structured pruning strategy is channel-kernel level structured pruning, and the steps are as follows:

[0103] Step 4.1: Preparations before pruning.

[0104] Based on the trained improved YOLO11 model, the importance metrics for each convolutional kernel and feature channel in the network are calculated. The core metrics include:

[0105] Absolute weight value: The smaller the absolute value of the weights of the convolution kernel, the lower its contribution to feature extraction;

[0106] Feature contribution, which is the proportion of the response value of the output feature map of a channel in the infrared target region. The lower the proportion, the smaller the contribution of the channel to the extraction of infrared target features.

[0107] The correlation of infrared thermal radiation characteristics is the correlation coefficient between the channel output and the mean and gradient of the grayscale characteristics of infrared target thermal radiation. The lower the correlation, the more likely it is to be pruned.

[0108] Step 4.2: Determine the pruning threshold and perform pruning.

[0109] Set the pruning thresholds: set the absolute value of the weight threshold to 0.03, the feature contribution threshold to 5%, and the infrared feature correlation threshold to 0.1.

[0110] Pruning is performed by removing redundant convolutional kernels and their corresponding feature channels that simultaneously satisfy the following criteria: absolute weight value less than 0.03, feature contribution less than 5%, and infrared feature correlation less than 0.1. The pruning scope is as follows:

[0111] Backbone network: Remove convolutional kernels that extract redundant background features, and retain kernels that extract infrared target thermal radiation and edge features;

[0112] Channel-Spatial Attention Module: Only channels strongly correlated with the thermal radiation characteristics of infrared targets are retained, while redundant channels for background suppression are removed;

[0113] Detection head: Removes redundant convolution kernels after simplifying the structure, further reducing the computational load of inference.

[0114] The pruning principle is to not prune the core channels of the cross-scale feature pyramid, so as to ensure the model's ability to detect infrared multi-scale targets after pruning.

[0115] Step 4.3: Short-term fine-tuning after pruning.

[0116] After pruning, the model parameters are removed, resulting in a slight decrease in accuracy. Short-term fine-tuning is required to restore accuracy. The fine-tuning operation is fully adapted to infrared detection scenarios, specifically as follows:

[0117] Fine-tuning dataset: Using the validation set of the infrared dataset eliminates the need to reuse the entire training set, improving fine-tuning efficiency;

[0118] Fine-tuning parameters: Set the number of fine-tuning rounds to 20, the learning rate to 1e-5, the optimizer to AdamW, and the loss function to still use the composite loss function;

[0119] Fine-tuning objective: Only update the weights of the remaining convolutional kernels and channels after pruning, allowing the model to re-adapt to infrared target features and restore the detection accuracy lost due to pruning.

[0120] By inputting the preprocessed infrared image test set into the trained and optimized improved YOLO11 model for forward inference, the detection results such as target category, confidence level, and bounding box coordinates can be output, thus achieving accurate detection of infrared salient targets.

[0121] Example 2:

[0122] This embodiment proposes an electronic device, including: one or more processors, and a memory, the memory being used to store instructions, which, when executed by the one or more processors, cause the one or more processors to perform the infrared salient target detection method based on the improved YOLO11.

[0123] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, implements the infrared salient target detection method based on the improved YOLO11 as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.

[0124] The processor is used to execute all or part of the steps in the infrared salient target detection method based on the improved YOLO11 as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.

[0125] The processor can be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the infrared salient target detection method based on the improved YOLO11 described in the above embodiments.

[0126] Example 3:

[0127] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0128] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the infrared salient target detection method based on the improved YOLO11 described in the various embodiments of this application.

[0129] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes. These media store computer programs, which, when executed by a processor, can implement the various steps of the improved YOLO11-based infrared salient target detection method described above.

[0130] Example 4:

[0131] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the infrared salient target detection method based on the improved YOLO11.

[0132] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.

[0133] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the present invention.

Claims

1. A method for detecting salient infrared targets based on an improved YOLOv11, characterized in that: The specific steps are as follows: S1: Construct an image dataset; Based on publicly available infrared image datasets, an image set covering various typical infrared scenes is constructed. All images in the constructed image set are formatted in a unified manner, and fine-grained ground truth annotations are performed based on salient targets. The annotation information includes the target category and the normalized coordinates of its bounding box. Then, the dataset is divided into training set, validation set, and test set to provide a data foundation for subsequent model training. S2: Image preprocessing; Based on the feature extraction method of the C2f module in the YOLO11 model, preprocessing operations are performed on the images in the image dataset constructed in step 1, including noise suppression, enhancement processing, and adaptation adjustment, to meet the input requirements of the improved YOLO11 model; S3: Construct an infrared target detection model; make multi-dimensional improvements to the YOLO11 basic network architecture based on the physical characteristics of infrared targets; The infrared target detection model is based on the YOLO11 infrastructure configuration file provided by the Ultralytics framework. Specifically, it includes: embedding a cascaded channel-spatial attention mechanism in the feature extraction path of the backbone network; constructing a cross-scale feature pyramid structure in the feature fusion module of the neck; removing the original fixed anchor boxes of the detection head and using the K-Means++ clustering algorithm to generate adaptive anchor boxes according to the target scale distribution of the infrared dataset. S4: Model training and testing; The model is trained using a cosine annealing learning rate scheduling strategy based on the training and test sets, and an early stopping mechanism is set to prevent overfitting. During the training process, based on the infrared target classification and localization error analysis, a composite loss function is constructed by introducing infrared target feature cosine similarity loss on the basis of YOLO11 default loss CIoU localization loss and cross-entropy classification loss, and the model performance is monitored. The trained model is lightweighted and a structured pruning strategy is adopted to remove redundant convolutional kernels and channels with absolute weight values ​​below the threshold and low feature contribution. Short-term fine-tuning is performed after pruning. Finally, the preprocessed infrared image test set is input into the trained and optimized model to obtain the final detected target, thus realizing real-time target detection of infrared images.

2. The infrared salient target detection method based on improved YOLO11 according to claim 1, characterized in that: The preprocessing in step 2 specifically includes: Adaptive weighted noise suppression is performed; by analyzing the variance and gradient features of local image regions, higher filtering weights are dynamically calculated and assigned to noisy pixels, and Gaussian noise and salt-and-pepper noise are filtered out using an adaptive filtering algorithm, preserving the edge and texture information of the target while denoising. A multi-scale retinal enhancement algorithm is used for enhancement processing. The algorithm employs a scale factor of 15 for small scale, 80 for medium scale, and 250 for large scale to separate the illumination component and reflectance component of the image. By adjusting the reflectance component, the contrast between the target and the background is enhanced, and texture details are strengthened. For scenes with different characteristics, enhancement results of different scales are adaptively selected or fused. Adaptability adjustments are made; the size of the preprocessed image is uniformly scaled to the standard input size of the YOLO11 network, and the pixel grayscale values ​​are normalized and mapped to the [0, 1] interval.

3. The infrared salient target detection method based on improved YOLO11 according to claim 1, characterized in that: The channel-space attention module in step 3 first analyzes the importance of each channel in the feature map through channel attention branches, filters and enhances useful feature channels that are strongly correlated with the thermal radiation characteristics of the infrared target, and suppresses information from redundant background channels. By using large-kernel deep convolution to focus on the spatial dimension of the feature map, the response of the region where the infrared target is located is enhanced, and background interference is further suppressed. The attention in the two stages works together to make the network more focused on the infrared salient target. The cross-scale feature pyramid structure adds a direct interaction channel between features of non-adjacent scales to the original adjacent scale feature fusion path, and achieves adaptive fusion of feature maps of different resolutions through lightweight convolution operations.

4. The infrared salient target detection method based on improved YOLO11 according to claim 3, characterized in that: In step 4, the CIoU localization loss and the cross-entropy classification loss are combined and denoted as... The format is as follows: ； in, It is the CIoU localization loss, used to optimize the position and size of the bounding box; It is the cross-entropy classification loss, used to optimize target category prediction; and These are the weighting coefficients for the two losses; and The calculation formulas are as follows: ； ； in, This is the intersection-union ratio (IU) between the predicted bounding box and the ground truth bounding box. The Euclidean distance between the center of the predicted bounding box and the center of the ground truth bounding box; The center coordinates of the prediction box are in the form of ( , ); The center coordinates of the true bounding box are in the form of ( , ); The diagonal length of the smallest closure rectangle that can simultaneously cover both the predicted and ground truth boxes; As a measure of aspect ratio consistency, The square of the difference between two angles is used to measure the consistency of the predicted bounding box and the ground truth bounding box in terms of aspect ratio. The smaller the value, the closer the aspect ratios are. The width of the actual bounding box. The actual height of the bounding box. The width of the prediction box. The height of the predicted bounding box; It is a balance parameter. ; Total number of categories; One-hot encoding for the real label; The predicted class probabilities by the model; The cosine similarity loss of infrared target features is denoted as This is used to constrain the model to learn feature representations that conform to infrared physical characteristics, and the formula is: ； in, This represents the number of samples predicted as the target in the current batch. The first extracted from the model The feature vector of the prediction target This is the corresponding prior infrared target standard feature vector; The L2 norm of a vector; The composite loss function is: ； in, and The weighting coefficients are used to balance the two losses; The preprocessing-model detection linkage mechanism dynamically feeds back and adjusts the parameters of noise suppression intensity, Retinex enhancement weights, and adaptive anchor boxes in the image preprocessing stage based on the confidence distribution, false detection rate, and regression error index of the model's detection results on the validation set, thereby achieving collaborative optimization of preprocessing and model.

5. The infrared salient target detection method based on improved YOLO11 according to claim 4, characterized in that: In step 4, the structured pruning strategy is specifically channel-kernel level structured pruning, and the steps are as follows: Step 4.1: Preparations before pruning; Based on the trained infrared target detection model, the importance metrics of each convolutional kernel and feature channel in the model network are calculated. The core metrics include: Absolute weight value; the smaller the absolute value of the weights of the convolution kernel, the lower its contribution to feature extraction. Feature contribution, which is the proportion of the response value of the output feature map of a channel in the infrared target region. The lower the proportion, the smaller the contribution of the corresponding channel to the extraction of infrared target features. The correlation of infrared thermal radiation characteristics is the correlation coefficient between the channel output and the mean and gradient of the grayscale of the infrared target's thermal radiation characteristics. The lower the correlation, the more likely it is to be pruned. Step 4.2: Determine the pruning threshold and perform pruning; Set the pruning thresholds: set the absolute value of the weight threshold to 0.03, the feature contribution threshold to 5%, and the infrared feature correlation threshold to 0.

1. Perform pruning: Remove redundant convolutional kernels and their corresponding feature channels that simultaneously satisfy the following conditions: absolute weight value less than 0.03, feature contribution less than 5%, and infrared feature correlation less than 0.1; the pruning scope includes: Backbone network: Remove convolutional kernels that extract redundant background features, and retain kernels that extract infrared target thermal radiation and edge features; Channel-Spatial Attention Module: Only channels strongly correlated with the thermal radiation characteristics of infrared targets are retained, while redundant channels for background suppression are removed; Detection head: Removes redundant convolution kernels after simplifying the structure, further reducing the computational load of inference; The pruning principle is: do not prune the core channels of the cross-scale feature pyramid to ensure the model's ability to detect infrared multi-scale targets after pruning; Step 4.3: Short-term fine-tuning after pruning, specifically: Fine-tuning dataset: Using the validation set of the infrared dataset eliminates the need to reuse the entire training set, improving fine-tuning efficiency; Fine-tuning parameters: Set the number of fine-tuning rounds to 20, the learning rate to 1e-5, and the optimizer to AdamW; Fine-tuning objective: Only update the weights of the remaining convolutional kernels and channels after pruning, allowing the model to re-adapt to infrared target features and restore the detection accuracy lost due to pruning.

6. An electronic device, characterized in that: The method includes one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the infrared salient target detection method based on the improved YOLO11 as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that: The device stores executable instructions that, when executed, cause the processor to perform the infrared salient target detection method based on the improved YOLO11 as described in any one of claims 1-5.

8. A computer program product, characterized in that: Includes a computer program or instructions that, when executed by a processor, implement the infrared salient target detection method based on the improved YOLO11 as described in any one of claims 1-5.

Images