A physical-prior-based infrared small target adaptive mask generation method and product

Abstract

The application discloses an infrared small target adaptive mask generation method and product based on physical priori, and the method comprises the following steps: acquiring an infrared image to be processed and a single-point label located at any position in a small target effective response area; taking the single-point label as a seed point, performing polarity unification processing on the infrared image according to local background statistical information of the seed point neighborhood; establishing a candidate small target area with the seed point, constructing a priority queue of pixels to be expanded, and initializing the candidate area and a statistical quantity; modeling a target mask generation problem as a maximum posterior estimation problem, and constructing a posterior energy function; gradually expanding the candidate area based on the priority queue and a greedy search strategy, and recording the posterior energy in the expansion process to obtain an energy record sequence; selecting an energy peak value from the energy record sequence, and generating a target pixel-level mask according to a corresponding optimal state backtracking; and extracting geometric supervision information according to the target pixel-level mask.

Description

Technical Field

[0001] This invention relates to the fields of infrared small target detection, weakly supervised annotation, and image processing, specifically to a method and product for adaptive mask generation of infrared small targets based on physical priors. This method can be used for pixel-level pseudo-mask generation of infrared small targets, extraction of geometric supervision information, auxiliary annotation, and fine shape restoration of detection results. Background Technology

[0002] Infrared small target detection has significant application value in scenarios such as space surveillance, low-altitude early warning, maritime search and rescue, and border monitoring. Due to the long imaging distance, weak target radiation energy, and the point spread effect of the optical system, infrared small targets typically occupy only a small number of pixels, exhibiting characteristics such as weak texture, low signal-to-noise ratio, and blurred boundaries, and are easily submerged in complex background clutter and noise. Most existing mainstream infrared small target detection methods model this task as a pixel-level segmentation task, using neural networks to learn the spatial response characteristics of the target to distinguish it from the background. These methods usually rely on a large number of accurate pixel-level annotations, resulting in high annotation costs. Because infrared small target boundaries are blurred and transition regions are unclear, different annotators have different understandings of the target boundary range, leading to poor consistency in manual annotation results, thus limiting the widespread application of pixel-level supervised infrared small target detection methods in practical scenarios.

[0003] To reduce annotation costs, existing research has attempted to generate target pseudomasks using point-supervised methods. Point-supervised annotation refers to annotating only a single point within the effective response region of the target, and then reconstructing the target region or generating a supervision signal based on this sparse annotation. Compared to manually drawing all pixels of the target, point-supervised annotation can significantly reduce annotation costs and improve annotation efficiency. However, existing point-supervised methods for generating infrared small target pseudomasks generally suffer from the following problems: First, they are sensitive to the location of the annotation points, usually requiring the annotation points to be as close as possible to the target center; otherwise, mask offset, growth failure, or unstable boundary recovery can easily occur. Second, they rely on additional prior information such as target size, clipping windows, bounding boxes, or other auxiliary prior information, which is often difficult to obtain reliably in real-world applications. Third, they lack unified modeling of the local saliency, internal homogeneity, and spatial compactness of infrared small targets, which can easily lead to background leakage, overexpansion of boundaries, and undergrowth problems.

[0004] Therefore, there is an urgent need for an adaptive mask generation method that can stably recover the compact and effective support region of a small infrared target by relying solely on single-point annotation at any location within the effective response region of the target, without requiring additional size priors or center position priors. Summary of the Invention

[0005] To address the problems of existing point-supervised infrared small target pseudomask generation methods, such as sensitivity to initial point position, reliance on size priors, poor adaptability in complex backgrounds, and unstable target boundary recovery, this invention aims to overcome the shortcomings of existing technologies by proposing an infrared small target adaptive mask generation method and product based on physical priors. Under the condition of only inputting an infrared image and a single-point annotation at any position within the effective response area of ​​the target, it automatically generates a compact and reliable target pixel-level mask and simultaneously extracts geometric supervision information such as the target centroid, equivalent radius, and circumscribed bounding box for subsequent detection network training, auxiliary annotation, and fine target recovery.

[0006] In view of this, the present invention proposes an adaptive mask generation method for small infrared targets based on physical priors, comprising: Step 1: Acquire the infrared image to be processed and a single-point annotation located at any position within the effective response area of ​​the small target; Step 2: Using single-point annotations as seed points, perform polarity unification processing on the infrared image based on the local background statistics of the seed point's neighborhood; Step 3: Establish candidate small target regions using seed points, construct a priority queue of pixels to be expanded, and initialize candidate regions and statistics; Step 4: Model the target mask generation problem as a maximum a posteriori estimation problem and construct the posterior energy function; Step 5: Expand the candidate region step by step based on the priority queue and greedy search strategy, and record the posterior energy during the expansion process to obtain the energy recording sequence; Step 6: Select the energy peak from the energy recording sequence, and generate the target pixel-level mask by backtracking according to the corresponding optimal state; Step 7: Extract geometric supervision information based on the target pixel-level mask.

[0007] As an improvement to the above method, step 2 includes: Use single-point annotation as seed point The local background median is calculated using a window centered on the seed point. The window is square with an odd number of pixels on each side. When seed point grayscale value Less than The infrared image to be processed is subjected to grayscale inversion; otherwise, the original image is left unchanged, satisfying the following formula: , in, This represents the infrared image to be processed after normalization based on the maximum response grayscale of the infrared camera. This represents the infrared image after polarity unification.

[0008] As an improvement to the above method, step 3 includes: Seed point As initial candidate regions, priority queues are established by arranging pixels in descending order of their grayscale responses after polarity unification. and seed point Push them into the priority queue according to their priority; Initialize the runtime statistics of the candidate region according to the following formula:

[0009] in, This represents the mean value within the candidate region. Indicates the standard deviation within the candidate region. This indicates the number of pixels in the current candidate region; This represents the grayscale value of the infrared image at the seed point after polarity unification. Establish path recording sequences and energy recording sequences; among them, The path recording sequence is used to record the pixels that are incorporated into the candidate region at each step in chronological order. The energy recording sequence is used to record the posterior energy value of the corresponding growth step.

[0010] As an improvement to the above method, the posterior energy function constructed in step 4 is:

[0011] in, Indicates the path taken from the seed point. The candidate region formed after the secondary expansion express Total posterior energy, This represents the mean value within the candidate region. Indicates the standard deviation within the candidate region. This represents the background mean of the neighborhood outside the candidate region. This represents the maximum Euclidean distance from the boundary pixel of the current candidate region to the seed point. Indicates spatial support parameters, and This represents a small positive number used to ensure numerical stability.

[0012] As an improvement to the above method, the expansion of step 5 specifically includes: In each round, the pixel with the highest current priority is taken from the priority queue and merged into the current candidate region. Then, the internal mean, internal standard deviation, region area, background mean of the outer neighborhood, and maximum Euclidean distance of the candidate region are updated. Then, pixels in the unvisited neighborhood of the pixel are pushed into the priority queue. The expansion termination condition is: the priority queue is empty or the area of ​​the candidate region reaches the upper bound of the area, where the upper bound of the area is: ,in, Indicates the spatial support parameters.

[0013] As an improvement to the above method, step 5 also includes a statistical preheating process, when the candidate region pixels If the set preheating threshold is not reached, a preset minimum value is recorded in the energy recording sequence, and the candidate region pixels... When the set preheating threshold is reached, the posterior energy is calculated and recorded.

[0014] As an improvement to the above method, step 6 specifically includes: The optimal growth step is selected from the energy recording sequence corresponding to the posterior energy peak. ; Based on the energy recording sequence, the previous The selected pixels are used to construct a target pixel-level mask, and the construction relationship satisfies the following formula:

[0015] in, Indicates pixel position The target pixel-level mask at that location. Indicates the first in the path record sequence 1 pixel.

[0016] As an improvement to the above method, the geometric supervision information extracted in step 7 includes: the target centroid, equivalent radius, and bounding box; wherein, The target centroid for:

[0017] in, Represents the pixel-level mask. The coordinates of each small target pixel; This represents the number of pixels representing small targets in a pixel-level mask. The equivalent radius for:

[0018] in, The pixel area corresponding to the pixel-level mask; The outer bounding box for:

[0019] in, , , and These represent the minimum and maximum coordinates of the pixel-level mask in the horizontal and vertical directions, respectively.

[0020] On the other hand, the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0021] Compared with the prior art, the advantages of the present invention are: 1. Only a single point annotation at any location within the effective response area of ​​the target is needed to generate a pixel-level mask for the target, which helps to reduce the cost of annotating infrared small target data.

[0022] 2. By using posterior energy constraints, statistical preheating, and full path backtracking, we can help suppress background leakage and boundary overexpansion, thereby improving the stability of mask generation.

[0023] 3. It can simultaneously output geometric information such as the target centroid, equivalent radius, and bounding box, which is convenient for adapting to scenarios such as detection training, auxiliary annotation, and fine target recovery. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the overall process of the adaptive mask generation method for small infrared targets based on physical priors according to the present invention. Figure 2 This is a schematic diagram of the region growth and posterior energy backtracking process based on priority queues in this invention. Detailed Implementation

[0025] This invention provides an adaptive mask generation method for small infrared targets based on physical priors. The method takes an infrared image and a single-point annotation located at any position within the effective response region of the target as input. By constructing a posterior energy constraint that integrates local saliency, internal homogeneity, size gain, and spatial compactness, the method progressively expands the candidate target region and backtracks to the optimal state to obtain a pixel-level mask for the small infrared target. Furthermore, it extracts geometric supervision information such as the target centroid, equivalent radius, and bounding box.

[0026] Specifically, the method of the present invention includes at least the following steps: Step S1: Obtain input data.

[0027] Acquire the infrared image to be processed and a single-point annotation located at any position within the effective response area of ​​the target.

[0028] Step S2: Perform target polarity unification processing.

[0029] The single-point annotation is used as the seed point. Based on the local background statistics of the seed point's neighborhood, polarity unification processing is performed on the input image to unify bright and dark targets into the same type of local peak search problem. The local background statistics are preferably the median background value within a local window centered on the seed point, with the window being square, preferably having a side length of [value missing]. to An odd number of pixels.

[0030] Step S3: Initialize candidate regions and establish growth mechanisms.

[0031] Candidate target regions are established using seed points, a priority queue of pixels to be expanded is constructed, and the statistics, region growth paths, and energy recording information corresponding to the candidate regions are initialized. The pixel priority in the priority queue is preferably determined based on the pixel grayscale response after polarity unification, and the higher the grayscale response, the higher the priority.

[0032] Step S4: Construct posterior energy constraints.

[0033] Construct a posterior energy function that simultaneously considers the target-background difference of the candidate region, the homogeneity within the region, the size gain of the region, and the spatial geometric compactness; wherein, the target-background difference is preferably represented by the difference between the mean within the candidate region and the mean of the background in the neighboring area outside the candidate region, and the spatial geometric compactness is preferably represented by the maximum Euclidean distance from the boundary of the candidate region to the seed point.

[0034] Step S5: Perform region expansion and record energy trajectory.

[0035] Based on the aforementioned posterior energy constraint, a priority queue and greedy search strategy are used to progressively expand the candidate region. During the expansion process, the state and energy value of the candidate region corresponding to each growth step are recorded, and the expansion is terminated when the priority queue is empty or the area of ​​the candidate region reaches a preset upper bound. To avoid statistical instability caused by a small number of initial pixels, it is preferable to use the corresponding energy value as an effective peak candidate only after the number of pixels in the candidate region exceeds the preheating threshold.

[0036] Step S6: Backtrack to generate the mask and output the geometric supervision information.

[0037] The target pixel-level mask is generated by backtracking from the optimal state corresponding to the energy peak, and the target centroid, equivalent radius and circumscribed bounding box are extracted based on the target pixel-level mask.

[0038] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0039] Example 1 Embodiment 1 of the present invention provides an adaptive mask generation method for small infrared targets based on physical priors. This method takes an infrared image and a single-point annotation located at any position within the effective response region of the target as input. Without requiring priors for target center, target size, clipping window, or bounding box, it adaptively recovers the compact and effective support region of the small infrared target and further outputs the target mask and its geometric supervision information. The geometric supervision information includes at least the target centroid, equivalent radius, and circumscribed bounding box.

[0040] 1. Input image and single-point annotation.

[0041] See Figure 1 First, acquire the infrared image to be processed. and single-point annotations located at any position within the target's effective response area. This point is used as a seed point. The single-point annotation does not need to be strictly located at the center of the target's thermal response; it only needs to be within the target's effective response area. This setting differs from existing methods that rely on precise center points or additional prior dimensions, thus reducing the difficulty of manual annotation and increasing the degree of freedom in annotation.

[0042] 2. The polarity of the objectives is unified.

[0043] At the seed point Surrounding estimated local background median Preferably, the local background median is calculated by a square window centered on the seed point, and the window side length is an odd number. to Pixels. When equation (1) is satisfied, perform grayscale inversion processing as shown in equation (2) on the input image; otherwise, keep the original image unchanged. Through this step, bright and dark targets can be unified into a local peak search problem.

[0044]

[0045]

[0046] in, This represents the grayscale value at the seed point. This represents the original infrared image obtained by normalizing the maximum response grayscale of the infrared camera. This represents the infrared image after polarity unification. This represents the local background median in the neighborhood of the seed point.

[0047] 3. Candidate region initialization.

[0048] After completing polarity unification, a priority queue is established using the seed point as the initial candidate region. Seed points are then pushed into a priority queue according to their priority. Preferably, the priority queue is sorted from high to low according to the pixel grayscale response after polarity unification, so that the priority queue can preferentially expand high-response pixels. At the same time, the running statistics of the candidate region are initialized according to equation (3), and a path record sequence is established. and energy recording sequences .in, Used to record the pixels that are incorporated into the candidate region at each step in chronological order. Used to record the posterior energy value of the corresponding growth step.

[0049]

[0050] in, This represents the mean value within the candidate region. Indicates the standard deviation within the candidate region. This indicates the number of pixels in the current candidate region. This represents the priority queue of pixels to be expanded. This represents the grayscale value at the seed point in the infrared image after polarity unification.

[0051] 4. Posterior energy modeling.

[0052] This implementation models the target mask generation problem as a maximum a posteriori estimation problem. Let... This represents the observed infrared image data. Indicates the path taken from the seed point. The candidate regions formed after the expansion are then the optimal target region satisfies equation (4). Further, the problem is transformed into a logarithmic posterior energy maximization problem through equation (5), and the data term, size gain term and spatial geometric constraint term are constructed using equations (8), (9) and (10) respectively, thus obtaining the total posterior energy function shown in equation (11).

[0053]

[0054] in, This represents the optimal target region. This represents the candidate region formed by the gradual expansion of the seed point. This represents the observed infrared image data. Indicates in the candidate region Observational likelihood under given conditions This represents the prior probability of a candidate region. This represents the evidence item for a given image.

[0055]

[0056] in, Indicates candidate region Total posterior energy, Represents a data item. This indicates a priori terms.

[0057]

[0058]

[0059]

[0060]

[0061]

[0062]

[0063] in, This represents the set of background pixels that are directly adjacent to the current candidate region but have not yet been incorporated into the candidate region, preferably consisting of a ring of unvisited pixels outside the current candidate region. This represents the set of pixels representing the boundary of the current candidate region. and This represents a small positive number used to ensure numerical stability. The data term is used to highlight the effective contrast of the target relative to the background and suppress non-target areas with excessive texture undulations within the region; the size gain term is used to provide expansion momentum in the early stages of region growth, preventing candidate regions from staying at local spurious peaks; the spatial geometric constraint term is used to suppress unconstrained expansion of the region to locations far from the seed point, reducing background leakage.

[0064] in, Indicates spatial support parameters, This represents the background mean of the neighborhood outside the candidate region. This represents the maximum Euclidean distance from the boundary pixel of the current candidate region to the seed point.

[0065] 5. Candidate region growth based on priority queue.

[0066] See Figure 2 This implementation uses a priority queue and a greedy strategy to gradually expand the candidate region. In each round, the pixel with the highest priority is taken from the priority queue and merged into the current candidate region. Then, the mean, standard deviation, area, mean of the surrounding background, and maximum distance of the region are updated. Then, pixels in the unvisited neighborhood of the pixel are pushed into the priority queue. Preferably, the candidate region neighborhood is an eight-neighborhood, or a four-neighborhood when computational resources are limited. To limit the unconstrained expansion of the candidate region, the area of ​​the candidate region satisfies equation (12), and the region expansion is terminated when the priority queue is empty or the upper bound of the area is reached.

[0067]

[0068] 6. Statistical preheating mechanism.

[0069] Since the candidate region initially contains only a very small number of pixels, and statistical quantities such as the internal standard deviation are unstable, this implementation method sets a warm-up stage. The posterior energy is only formally calculated and recorded when the number of pixels in the candidate region satisfies equation (13); otherwise, the calculation is stopped. At this time, only the statistics of the candidate region are updated and the growth path is recorded, without considering the energy of that stage as a valid peak candidate. Preferably, for growth steps that do not reach the preheating threshold, the energy recording sequence can be updated. The system records a preset minimum value to prevent early unstable states from being mistakenly selected as the optimal peak value. This warm-up mechanism, in conjunction with seed initialization, can effectively alleviate the instability caused by insufficient samples during the cold start phase.

[0070]

[0071] 7. Optimal mask backtracking generation.

[0072] Unlike traditional region growing methods that rely on online thresholds or fixed stopping conditions, this implementation does not directly determine the final result during the growing process. Instead, it records the complete candidate region sequence and posterior energy sequence throughout the expansion process. The optimal growth step is determined according to equation (14). Then, based on the path record sequence, the first... The selected pixels construct the final target binary mask. The construction relationship preferably satisfies equation (15). By using this full-path recording and peak backtracking method, early stopping errors and local optima problems caused by target boundary ambiguity, local energy fluctuations or background interference can be avoided.

[0073]

[0074]

[0075] in, Indicates pixel position The final binary mask value at that point, Indicates the first in the path record sequence 1 pixel, This represents the optimal growth step when the posterior energy reaches its maximum value.

[0076] 8. Geometric supervision information extraction.

[0077] Obtain the final target mask Then, its geometric supervision information can be further extracted. Let the number of target pixels in the final mask be... The corresponding pixel area is denoted as In the pixel-based measurement implementation, and The values ​​are the same.

[0078] The target centroid can be calculated according to equation (16), the target equivalent radius can be calculated according to equation (17), and the circumscribed bounding box can be constructed according to equation (18). The target centroid and equivalent radius can be used as compact geometric supervision information for subsequent detection network training, scale estimation, and target recovery; the circumscribed bounding box can be used as auxiliary localization information, annotation display information, or supplementary supervision information compatible with other detection paradigms.

[0079]

[0080]

[0081]

[0082] in, Indicates the first digit in the final mask. The coordinates of each target pixel. and These represent the x and y coordinates of the target's centroid, respectively. Indicates the target's equivalent radius. Indicates the outer bounding box. , , and These represent the minimum and maximum coordinates of the final mask in the horizontal and vertical directions, respectively.

[0083] 9. Spatial support parameter settings.

[0084] In one implementation, when only single-point annotations are available without instance scale information, fixed spatial support parameters can be used. Preferably, a fixed value can be used during the annotation stage or in a blind mode without scale prior. As a default implementation, in another implementation, when target scale estimation information is available, the spatial support parameters can be adaptively set according to equation (19). Preferably, the scaling factor... Greater than , better And can to The spatial support parameter is adjusted within a certain range. Since small infrared targets typically have a blurred transition region outside their compact core, the spatial support parameter is generally larger than the equivalent radius. This parameter setting is used to constrain the spatial expansion range of the candidate region and can be adjusted according to different imaging conditions.

[0085]

[0086] in, This represents the target scale estimate. This represents the proportionality coefficient.

[0087] 10. Explanation of implementation results.

[0088] This implementation method integrates local saliency, internal homogeneity, size gain, and spatial geometric constraints into posterior energy modeling. Combined with target polarity unification, statistical warm-up, priority queue growth, and a full-path backtracking mechanism, it achieves stable generation of a compact mask from a single point within the target's effective response region. This method reduces annotation costs while simultaneously outputting information such as the target centroid, equivalent radius, and bounding box, thus supporting subsequent detection training, assisted annotation, and fine-grained target recovery.

[0089] Example 2 Embodiments of the present invention may also provide a computer program product, including a computer program. When the computer program is executed by a processor, it can implement the various steps in the above method embodiments.

[0090] The innovations of this invention are mainly reflected in the following aspects: 1. An adaptive mask generation method for infrared small targets is proposed, which relies solely on single-point annotation at any location within the effective response area of ​​the target to recover the effective support area of ​​the target without requiring precise center point and strict size prior.

[0091] 2. Local saliency, internal homogeneity, size gain, and spatial compactness are uniformly modeled as a posterior energy function, and the optimal target region is determined by full path recording and peak backtracking.

[0092] 3. Based on the mask generation, the target centroid, equivalent radius, and bounding box are further output, thus enabling compatibility with various application forms such as pixel-level supervision, geometric supervision, and box-level localization.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for adaptive mask generation of small infrared targets based on physical priors, comprising: Step 1: Acquire the infrared image to be processed and a single-point annotation located at any position within the effective response area of ​​the small target; Step 2: Using single-point annotations as seed points, perform polarity unification processing on the infrared image based on the local background statistics of the seed point's neighborhood; Step 3: Establish candidate small target regions using seed points, construct a priority queue of pixels to be expanded, and initialize candidate regions and statistics; Step 4: Model the target mask generation problem as a maximum a posteriori estimation problem and construct the posterior energy function; Step 5: Expand the candidate region step by step based on the priority queue and greedy search strategy, and record the posterior energy during the expansion process to obtain the energy recording sequence; Step 6: Select the energy peak from the energy recording sequence, and generate the target pixel-level mask by backtracking according to the corresponding optimal state; Step 7: Extract geometric supervision information based on the target pixel-level mask.

2. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, Step 2 includes: Use single-point annotation as seed point The local background median is calculated using a window centered on the seed point. The window is square with an odd number of pixels on each side. When seed point grayscale value Less than The infrared image to be processed is subjected to grayscale inversion; otherwise, the original image is left unchanged, satisfying the following formula: ， in, This represents the infrared image to be processed after normalization based on the maximum response grayscale of the infrared camera. This represents the infrared image after polarity unification.

3. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, Step 3 includes: Seed point As initial candidate regions, priority queues are established by arranging pixels in descending order of their grayscale responses after polarity unification. and seed point Push them into the priority queue according to their priority; Initialize the runtime statistics of the candidate region according to the following formula: in, This represents the mean value within the candidate region. Indicates the standard deviation within the candidate region. This indicates the number of pixels in the current candidate region; This represents the grayscale value of the infrared image at the seed point after polarity unification. Establish path recording sequences and energy recording sequences; among them, The path recording sequence is used to record the pixels that are incorporated into the candidate region at each step in chronological order. The energy recording sequence is used to record the posterior energy value of the corresponding growth step.

4. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, The posterior energy function constructed in step 4 is: in, Indicates the path taken from the seed point. The candidate region formed after the secondary expansion express Total posterior energy, This represents the mean value within the candidate region. Indicates the standard deviation within the candidate region. This represents the background mean of the neighborhood outside the candidate region. This represents the maximum Euclidean distance from the boundary pixel of the current candidate region to the seed point. Indicates spatial support parameters, and This represents a small positive number used to ensure numerical stability.

5. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, The expansion of step 5 specifically includes: In each round, the pixel with the highest current priority is taken from the priority queue and merged into the current candidate region. Then, the internal mean, internal standard deviation, region area, background mean of the outer neighborhood, and maximum Euclidean distance of the candidate region are updated. Then, pixels in the unvisited neighborhood of the pixel are pushed into the priority queue. The expansion termination condition is: the priority queue is empty or the area of ​​the candidate region reaches the upper bound of the area, where the upper bound of the area is: ,in, Indicates the spatial support parameters.

6. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, Step 5 also includes a statistical preheating process, when candidate region pixels If the set preheating threshold is not reached, a preset minimum value is recorded in the energy recording sequence, and the candidate region pixels... When the set preheating threshold is reached, the posterior energy is calculated and recorded.

7. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, Step 6 specifically includes: The optimal growth step is selected from the energy recording sequence corresponding to the posterior energy peak. ; Based on the energy recording sequence, the previous The selected pixels are used to construct a target pixel-level mask, and the construction relationship satisfies the following formula: in, Indicates pixel position The target pixel-level mask at that location. Indicates the first in the path record sequence 1 pixel.

8. The method for adaptive mask generation of small infrared targets based on physical priors according to claim 1, characterized in that, The geometric supervision information extracted in step 7 includes: the target centroid, equivalent radius, and bounding box; wherein, The target centroid for: in, Represents the pixel-level mask. The coordinates of each small target pixel; This represents the number of pixels representing small targets in a pixel-level mask. The equivalent radius for: in, The pixel area corresponding to the pixel-level mask; The outer bounding box for: in, , , and These represent the minimum and maximum coordinates of the pixel-level mask in the horizontal and vertical directions, respectively.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method of claim 1.

Images