An infrared image small target tracking method and system based on image local enhancement and 3D sparse convolution network

Abstract

The application relates to an infrared image small target tracking method and system based on image local enhancement and a 3D sparse convolution network, and relates to the technical field of infrared small target tracking detection. In order to solve the problem that the target features of current infrared small targets are not obvious and the background is complex and difficult to detect, and to improve tracking precision and shorten tracking time, the application comprises the following steps: (1) enhancing the infrared target and suppressing the background by subjecting the infrared image to an image space filter; (2) generating sparse data by subjecting the enhanced infrared image to sparse sampling; (3) stacking the sparse data into a three-dimensional tensor according to the time dimension; (4) performing data enhancement on the three-dimensional tensor by cutting, superimposing and rotating; and (5) extracting a track by subjecting the input sparse data to segmentation by using a 3D sparse convolution network. The application can track the infrared small target with occluded track under the condition that there is no initial position marker and the background is complex, and the tracking precision is high.

Description

Technical Field

[0001] This invention relates to the fields of image processing and deep learning technology, and in particular to a method and system for tracking small targets in infrared images based on local image enhancement and 3D sparse convolutional networks. Background Technology

[0002] Infrared small target tracking is an important research direction in computer vision and image processing. With the continuous advancement of infrared imaging technology, infrared cameras have become essential equipment widely used in military, security, industrial, and medical fields. These infrared cameras can capture the infrared radiation emitted by targets, and are therefore widely used for target detection and tracking in low-light, nighttime, and harsh environments. Traditional visible light target tracking techniques face several challenges in infrared small target tracking, such as low contrast between the target and background, blurred target outlines, and dynamic occlusion. Therefore, research on infrared small target tracking algorithms has become particularly important. These algorithms need to combine the characteristics of infrared images, such as temperature distribution and radiation intensity, to design effective target detection and tracking methods to achieve accurate tracking of infrared small targets.

[0003] In recent years, with the development of deep learning technology, infrared small target tracking algorithms based on deep learning have made significant progress. Deep learning technology can automatically learn feature representations of targets from data, thereby improving the accuracy and robustness of target detection and tracking. At the same time, various novel neural network structures and training methods are constantly emerging, providing abundant possibilities for the improvement and optimization of infrared small target tracking algorithms. However, most existing tracking methods require marking the initial position of the infrared small target to obtain the subsequent trajectory, which presents many inconveniences in practical applications. Summary of the Invention

[0004] This invention primarily addresses the challenges of tracking unknown targets and detecting targets in complex, occluded backgrounds, presenting a method and system for tracking small targets in infrared images based on local image enhancement and 3D sparse convolutional networks. This invention can accurately extract small infrared targets and effectively suppress background noise.

[0005] The technical solution adopted in this invention is: a method for tracking small targets in infrared images based on local image enhancement and 3D sparse convolutional networks, comprising the following steps:

[0006] Step 1: The infrared image is processed through an image spatial filter to enhance the infrared target while suppressing the background;

[0007] Step 2: Sparsely sample the enhanced infrared image to generate sparse data;

[0008] Step 3: Stack the sparse data into a three-dimensional tensor according to the time dimension;

[0009] Step 4: Perform data augmentation by cropping, stacking, and rotating the 3D tensor;

[0010] Step 5: Use a 3D sparse convolutional network to segment and extract trajectories from the input sparse data.

[0011] Furthermore, the specific implementation method of enhancing the infrared target and suppressing the background by passing the infrared image through an image spatial filter in step 1 is as follows:

[0012] Step 1.1: The grayscale value of small infrared targets is high, often much higher than that of the background, but there are large areas of bright background. Local contrast suppression is adopted. An N×N sliding window is used to traverse the infrared image from left to right and from top to bottom according to the step size of step to obtain a set of slice data. step is usually 3 and N is usually 9.

[0013] Y1,Y2,…,Y n (1)

[0014] Where Y n The nth slice has a size of N×N

[0015] Step 1.2: Perform local target enhancement and background suppression on each slice of data, and restore the original image size:

[0016]

[0017] Where i∈1,…,n,δ is a local minimum, usually taken as 0.000001, α and θ are two hyperparameters, usually taken as 2 and 1, and finally the Y of the original image size is obtained through the resize operation.

[0018] Step 1.3: Obtain the feature data of the global image, and perform target enhancement and background suppression on the global image:

[0019]

[0020] Where Y global For the globally enhanced image, Laplacian(Y) is the Laplacian operator.

[0021] Furthermore, the specific implementation method for generating sparse data by sparse sampling of the enhanced infrared image in step 2 is as follows:

[0022] Step 2.1, calculate the average background intensity threshold of the image obtained in Step 1.3:

[0023]

[0024] Where λ is a hyperparameter, usually taken as 2;

[0025] Step 2.2: Using the average background intensity obtained in Step 2.1 as a threshold, the sparse image data Y is obtained. sparse :

[0026]

[0027] Furthermore, the specific implementation of stacking sparse data into a three-dimensional tensor according to the time dimension in step 3 is as follows:

[0028] Step 3.1: Stack all sparse data of the same sequence into a three-dimensional sparse tensor according to the time dimension. The points with values ​​in the tensor account for less than 10% of the total tensor.

[0029] Step 3.2: Resample the W and H dimensions of the sparse data to 256 as required;

[0030] Furthermore, the specific implementation of data augmentation through clipping, overlaying, and rotating the 3D tensor in step 4 is as follows:

[0031] Step 4.1: Cut the stacked three-dimensional sparse tensors of the same sequence in the time dimension according to the step size S to obtain a set of three-dimensional tensors of size W×H×S, where W is the width of the original image, H is the length of the original image, both are 256, and S is usually 50.

[0032] Step 4.2: Randomly select tensors obtained in step 4.1, rotate them, and extract two-thirds of the tensors, then superimpose them in pairs to obtain more complex sparse data.

[0033] Furthermore, the specific implementation method of using a 3D sparse convolutional network to segment and extract trajectories from the input sparse data in step 5 is as follows:

[0034] Step 5.1: Encode the tensors of the multiple trajectories obtained in Step 4.2, retaining only the intensity value and coordinates of the data location, and store them as a dictionary;

[0035] Step 5.2: The dictionary is input into the downsampling part of the 3D sparse convolutional network in batches of size `batch_size` for encoding. `batch_size` is 2. The encoding part consists of four convolutional modules, each of which is composed of several sparse convolutional modules. Each sparse convolutional module consists of one convolutional layer, one batch normalization (BN) layer, and one ReLU activation layer. The first convolutional module consists of a single sparse convolutional module with 1 input channel, 16 output channels, a kernel size of 3, and a stride of 1. The second convolutional module consists of a single sparse convolutional module with kernels of size 3, a stride of 1, padding of 1, and 16 input and output channels. The third convolutional module consists of three sparse convolutional modules. The first sparse convolutional module has a kernel side length of 3, a stride of 2, padding of 1, 16 input channels, and 32 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 32 input channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 32 input channels. The fourth convolutional module also consists of three sparse convolutional modules. The first sparse convolutional module has a kernel side length of 3, a stride of 2, padding of 1, 32 input channels, and 64 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 64 input channels, and 64 output channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 64 input channels. Obtain the advanced features after dimensional elevation;

[0036] Step 5.3: The upgraded high-level features are input into the upsampling modules, which consist of four modules. Each upsampling module includes concatenation and convolutional parts, and also contains three deconvolutions and three concatenation parts. The features output from each deconvolution are concatenated with the convolutional output features of the same size obtained in Step 5.2 and input into the next deconvolutional layer to obtain output feature values ​​of the same size. The first upsampling module contains three sparse convolutional modules: the first is a residual convolutional module, containing two convolutional layers with 64 input and output channels, a kernel size of 3, and padding of 1; the second is a 3D sparse convolutional module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1; the third is a 3D desparse convolutional module with 64 input and output channels, a kernel size of 3, and a stride of 1. The second upsampling module contains three sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 64 input and output channels each, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third is a 3D anti-sparse convolution module with 64 input channels, 32 output channels, a kernel size of 3, and a stride of 1. The third upsampling module contains three sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 32 input and output channels each, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 64 input channels, 32 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third is a 3D anti-sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, and a stride of 1. The fourth upsampling module contains two sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 16 input and 16 output channels, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, a stride of 1, and padding of 1. Finally, a dictionary is output.

[0037] Step 5.4: Decode the obtained dictionary into sparse data of the original size, slice it to obtain each image containing only the target, and obtain the target trajectory.

[0038] Step 5.5: Set the network learning rate to 0.0001, with exponential decay after each iteration. Initialize the parameters: alpha = 0.25 and gamma = 2.0. Optimize the model using the Focal loss function. Perform 200 iterations of training on the training set.

[0039] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0040] The infrared image small target tracking method proposed in this invention, based on local image enhancement and a 3D sparse convolutional network, first decomposes the original image into image blocks. Within each block, local background suppression is performed, and then global image features are calculated for target enhancement and background suppression. The enhanced image undergoes background removal to obtain sparse data. This sparse data is then stacked along the time dimension to obtain a 3D sparse tensor. The tensor is then truncated along the time dimension to obtain a series of tensor blocks of equal length, and a portion of these blocks are rotated and superimposed for data enhancement. The enhanced tensors are then output in batches to a 3D sparse network for encoding to obtain high-level features, followed by decoding. Finally, a sparse tensor containing only the target is output. Slicing this tensor achieves target detection, yielding the infrared small target trajectory. This invention enhances the target through local and global filtering and samples the data to sparseness, utilizing a neural network for detection. This effectively leverages the sparsity of the data, improving detection speed and accuracy. Furthermore, this invention directly detects time series data without requiring prior initial target position marking, significantly improving practicality. This invention can accurately extract small infrared targets and effectively suppress background, enabling high-precision tracking of small infrared targets with trajectory obstruction even without initial position markers and under complex backgrounds.

[0041] This invention employs advanced infrared small target tracking technology, eliminating the need to determine the target's initial position. Traditional target tracking techniques typically require prior knowledge of the target's location before tracking can begin. This invention, however, utilizes temporal information to track the target immediately upon its appearance, significantly improving tracking efficiency and accuracy. Secondly, this invention employs 3D sparse convolution, enabling rapid and accurate capture of the target's position information while reducing computational requirements. Traditional target tracking techniques are often affected by ambient light and background interference, leading to low tracking accuracy. This invention, through the use of deep learning algorithms, effectively reduces the impact of these interference factors, improving tracking accuracy and stability. Furthermore, this invention exhibits strong anti-interference capabilities and adaptability. Due to the strong penetrating power and anti-interference capabilities of infrared technology, this invention can reliably and stably track targets even in complex environments, such as at night, in adverse weather conditions, or against complex backgrounds, demonstrating strong applicability and practicality. In summary, this invention not only overcomes the limitation of traditional target tracking technology requiring prior determination of the target position, but also possesses high precision, high stability, and strong anti-interference capabilities. It can achieve rapid and accurate tracking of small targets in various complex environments, and has broad application prospects and economic benefits. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a schematic diagram of a 3D sparse convolutional network structure;

[0044] Figure 3 This is a detection result image of small infrared targets;

[0045] Figure 4 This is an image showing the results of infrared target tracking. Detailed Implementation

[0046] The following is a description of the embodiments and appendices. Figure 1-3 The present invention will be described in further detail, but the embodiments of the present invention are not limited thereto.

[0047] This embodiment uses a publicly available infrared target detection dataset for air-to-ground applications—an infrared time-sensitive target detection and tracking dataset—as an example to detail the process of using this invention for infrared small target detection. This dataset contains 87 infrared sequences, each with 250 images. The dataset comprises a total of 21,750 frames and 89,174 targets, forming 87 video sequences and 393 target trajectories, making it quite large. Each infrared image in the dataset has a resolution of 640×480 pixels, an 8-bit bit depth, and a size of approximately 300KB.

[0048] according to Figure 1 The flowchart shown illustrates a method for tracking small targets in infrared images based on local image enhancement and 3D sparse convolutional networks, comprising the following steps:

[0049] Step 1: Local contrast suppression is adopted. An N×N sliding window is used to traverse the infrared image from left to right and from top to bottom according to the step size of step to obtain a set of slice data. N is usually 9 and step is usually 3.

[0050] Y1,Y2,…,Y n (1)

[0051] Where Y n For the nth slice, with a size of N×N, perform local target enhancement and background suppression on each slice, and restore the original image size:

[0052]

[0053] Where i∈1,…,n, δ is a local minimum, typically 0.000001, and α and θ are two hyperparameters, typically 2 and 1. Finally, the Y-size of the original image is obtained through a resize operation. The feature data of the global image is obtained, and target enhancement and background suppression are performed on the global image:

[0054]

[0055] Where Yglobal For the globally enhanced image, Laplacian(Y) is the Laplacian operator.

[0056] Step 2, calculate the average background intensity threshold:

[0057]

[0058] Where λ is set to 2, the average background intensity obtained in step 2.1 is used as the threshold to obtain the sparse image data Y. sparse :

[0059]

[0060] Step 3: Stack all sparse data of the same sequence into a three-dimensional sparse tensor according to the time dimension. The points with values ​​in the tensor account for less than 10% of the total tensor. Resample the W and H dimensions of the sparse data to 256 as required.

[0061] Step 4: Cut the stacked three-dimensional sparse tensors of the same sequence in the time dimension according to the step size S to obtain a set of three-dimensional tensors of size W×H×S, where W is the width of the original image, H is the length of the original image, both are 256, and S is usually 50; randomly select the tensors obtained in step 4.1 for rotation, and extract two-thirds of the tensors, and stack them in pairs to obtain more complex sparse data.

[0062] Step 5: Encode the tensors of the multiple trajectories obtained in Step 4.2, retaining only the intensity values ​​and coordinates of the data locations, and store them as a dictionary. Input the dictionary into the downsampling part of the 3D sparse convolutional network in batches of size 2. The encoding part consists of four convolutional modules, each of which is composed of several sparse convolutional modules. A sparse convolutional module consists of one convolutional layer, one batch normalization (BN) layer, and one ReLU activation layer. The first convolutional module consists of a sparse convolutional module with 1 input channel, 16 output channels, a kernel size of 3, and a stride of 1. The second convolutional module consists of a sparse convolutional module with kernels of side length 3, a stride of 1, padding of 1, and 16 input and output channels. The third convolutional module consists of three sparse convolutional modules. The first sparse convolutional module has a kernel side length of 3, a stride of 2, padding of 1, 16 input channels, and 32 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 32 input channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 32 input channels. The fourth convolutional module also consists of three sparse convolutional modules. The first sparse convolutional module has a kernel side length of 3, a stride of 2, padding of 1, 32 input channels, and 64 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 64 input channels, and 64 output channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, and 64 input channels. The high-level features obtained after dimensionality upscaling are then input into the upsampling modules. There are four upsampling modules in total. Each upsampling module includes concatenation and convolutional parts, and also contains three deconvolutions and three concatenation parts. The features output from each deconvolution are concatenated with the convolutional output features of the same size obtained in step 5.2 and input into the next deconvolutional layer to obtain output feature values ​​of the same size. The first upsampling module contains three sparse convolutional modules: the first is a residual convolutional module, containing two convolutional layers with 64 input and output channels, a kernel size of 3, and padding of 1; the second is a 3D sparse convolutional module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1; the third is a 3D desparse convolutional module with 64 input and output channels, a kernel size of 3, and a stride of 1.The second upsampling module contains three sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 64 input and output channels each, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third is a 3D anti-sparse convolution module with 64 input channels, 32 output channels, a kernel size of 3, and a stride of 1. The third upsampling module contains three sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 32 input and output channels each, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 64 input channels, 32 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third is a 3D anti-sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, and a stride of 1. The fourth upsampling module contains two sparse convolution modules. The first is a residual convolution module, consisting of two convolutional layers with 16 input and output channels, a kernel size of 3, and padding of 1. The second is a 3D sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, a stride of 1, and padding of 1. Finally, a dictionary is output. The resulting dictionary is decoded into sparse data of the original size, sliced, and each slice contains only the target image, thus obtaining the target trajectory.

[0063] Step Six: Divide segments 1 to 76 of the infrared time-sensitive target detection and tracking dataset for air-to-ground applications into training and validation sets, and segments 77 to 87 into a test set. Set the network learning rate to 0.0001, with exponential decay after each iteration. Initialize the parameters: alpha = 0.25 and gamma = 2.0. Optimize the model using the Focal loss function. Perform 200 iterations of training on the training set.

[0064] The experimental environment used is as follows:

[0065] Hardware environment: Intel Xeon CPU, Quadro GV100 GPU

[0066] Software environment: Python 3.8, PyTorch 1.10.0

[0067] The model was validated using a validation set after 200 iterations of training, and the model with the lowest loss function value was selected as the optimal network parameter model. The detection results from the test set were then used to verify the model's actual detection performance. The test set images were fed into the optimal network parameter model to obtain a dictionary containing only the targets. The dictionary was decoded according to the original tensor structure to restore the original image size, resulting in the detected small target images, such as... Figure 3 , Figure 4As shown, this invention has excellent performance in tracking and detecting weak infrared targets.

[0068] The above description represents a preferred embodiment of the present invention. However, it is important to emphasize that the implementation of the present invention is not limited to the above examples. Any adjustments, modifications, substitutions, combinations, or simplifications that do not depart from the core essence and principles of the present invention should be considered equivalent alternatives and included within the scope of protection of the present invention.

Claims

1. A method for tracking small targets in infrared images based on local image enhancement and 3D sparse convolutional networks, characterized in that, Includes the following steps: Step 1: The infrared image is processed through an image spatial filter to enhance the infrared target while suppressing the background; Step 2: Sparsely sample the enhanced infrared image to generate sparse data; Step 3: Stack the sparse data into a three-dimensional tensor according to the time dimension; Step 4: Perform data augmentation by cropping, stacking, and rotating the 3D tensor; Step 5: Use a 3D sparse convolutional network to segment and extract trajectories from the input sparse data; The implementation process of step 4 is as follows: Step 4.1: Cut the stacked three-dimensional sparse tensors of the same sequence in the time dimension according to the step size S to obtain a set of three-dimensional tensors of size W×H×S, where W is the width of the original image, H is the length of the original image, both are 256, and S is 50. Step 4.2: Randomly select tensors obtained in step 4.1, rotate them, and extract two-thirds of the tensors, then superimpose them in pairs to obtain more complex sparse data. The implementation process of step 5 is as follows: Step 5.1: Encode the tensors of the multiple trajectories obtained in Step 4.2, retaining only the intensity value and coordinates of the data location, and store them as a dictionary; Step 5.2: Input the dictionary into the downsampling part of the 3D sparse convolutional network in batches of size `batch_size` for encoding. `batch_size` is 2. The encoding part consists of four convolutional modules, each composed of several sparse convolutional modules. Each sparse convolutional module consists of one convolutional layer, one batch normalization (BN) layer, and one ReLU activation layer. The first convolutional module consists of one sparse convolutional module with 1 input channel, 16 output channels, a kernel size of 3, and a stride of 1. The second convolutional module consists of one sparse convolutional module with kernels of 3 sides, a stride of 1, padding of 1, and 16 input and output channels. The third convolutional module consists of three sparse convolutional modules, with the first sparse convolutional module having kernels of 3 sides and a stride of 2. The first sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 16 input channels, and 32 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 32 input channels, and 32 output channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 32 input channels, and 32 output channels. The fourth convolutional module consists of three sparse convolutional modules. The first sparse convolutional module has a kernel side length of 3, a stride of 2, padding of 1, 32 input channels, and 64 output channels. The second sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 64 input channels, and 64 output channels. The third sparse convolutional module has a kernel side length of 3, a stride of 1, padding of 1, 64 input channels, and 64 output channels. This yields the higher-dimensional features after dimensionality enhancement. Step 5.3: The high-level features after dimensionality upsampling are input into the upsampling modules, which consist of four modules. Each upsampling module includes concatenation and convolution parts, and also contains three deconvolutions and three concatenation parts. The features output from each deconvolution are concatenated with the convolution output features of the same size obtained in Step 5.2 and input into the next deconvolution layer to obtain output feature values ​​of the same size. The first upsampling module contains three sparse convolution modules, the first of which is a residual convolution module containing two input and output layers. The first layer is a convolutional layer with 64 input channels, a kernel size of 3, and padding of 1. The second layer is a 3D sparse convolutional module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The third layer is a 3D anti-sparse convolutional module with 64 input channels, 64 output channels, a kernel size of 3, and a stride of 1. The second upsampling module contains three sparse convolutional modules. The first is a residual convolutional module containing two convolutional layers with 64 input and output channels, a kernel size of 3, and padding of 1. The first layer is a 3D sparse convolution module with 128 input channels, 64 output channels, a kernel size of 3, a stride of 1, and padding of 1. The second layer is a 3D anti-sparse convolution module with 64 input channels, 32 output channels, a kernel size of 3, and a stride of 1. The third upsampling module contains three sparse convolution modules: the first is a residual convolution module containing two convolutional layers with 32 input and output channels, a kernel size of 3, and padding of 1; the second is a 3D sparse convolution module with 64 input channels and 32 output channels. The first layer has a kernel size of 3, a stride of 1, and padding of 1. The second layer is a 3D inverse sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, and a stride of 1. The third layer is an upsampling module containing two sparse convolution modules: the first is a residual convolution module with two convolutional layers, each with 16 input and output channels, a kernel size of 3, and padding of 1; the second is a 3D sparse convolution module with 32 input channels, 16 output channels, a kernel size of 3, a stride of 1, and padding of 1. Finally, a dictionary is output. Step 5.4: Decode the obtained dictionary into sparse data of the original size, slice it to obtain each image containing only the target, and obtain the target trajectory; Step 5.5: Set the learning rate of the network to 0.0001 and decay exponentially after each iteration; The parameters were initialized with alpha set to 0.25 and gamma set to 2.

0. The model was then optimized using the Focal loss function and trained on the training set for 200 iterations.

2. The infrared image small target tracking method based on local image enhancement and 3D sparse convolutional network as described in claim 1, characterized in that: The specific implementation method of step 1 is as follows: Step 1.1: Local contrast suppression is adopted. An N×N sliding window is used to traverse the infrared image from left to right and from top to bottom according to the step size of step. Step is 3, and a set of slice data is obtained. (1) in This is the nth slice, with a size of N×N, where N is 9; Step 1.2: Perform local target enhancement and background suppression on each slice of data, and restore the original image size: (2) (3) in This is a very small value, taken as 0.000001. We have two hyperparameters, taking values ​​of 2 and 1, and finally obtain the Y size of the original image through a resize operation; Step 1.3: Obtain the feature data of the global image, and perform target enhancement and background suppression on the global image: (4) (5) in The image after global enhancement. For the Laplace operator.

3. The infrared image small target tracking method based on local image enhancement and 3D sparse convolutional network as described in claim 1, characterized in that: The specific implementation method of step 2 is as follows: Step 2.1, calculate the average background intensity threshold of the image obtained in Step 1.3: (6) in Let be a hyperparameter, and let 2 be its value. Step 2.2: Using the average background intensity threshold obtained in Step 2.1 as a threshold, sparse image data is obtained. : (7)。 4. The infrared image small target tracking method based on local image enhancement and 3D sparse convolutional network as described in claim 1, characterized in that: The specific implementation method of step 3 is as follows: Step 3.1: Stack all sparse data of the same sequence into a three-dimensional sparse tensor according to the time dimension. The points with values ​​in the tensor account for less than 10% of the total tensor. Step 3.2: Resample the W and H dimensions of the sparse data to 256 as required.

5. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program configured to, when invoked by a processor, implement the steps of the infrared image small target tracking method based on image local enhancement and 3D sparse convolutional networks as described in any one of claims 1-4.

6. A small target tracking system for infrared images based on local image enhancement and 3D sparse convolutional networks, characterized in that: The system has a program module corresponding to the steps of any one of the claims 1-4 above, and executes the steps in the infrared image small target tracking method based on local image enhancement and 3D sparse convolutional network when it runs.

Images