Method and system for evaluating dynamic camouflage effect of target
By adopting a dynamic camouflage evaluation method based on video segmentation models, the problem of difficulty in evaluating the dynamic camouflage effect in existing technologies is solved. This method achieves automated and objective quantification of the camouflage effect of moving targets, breaks through the limitations of traditional static image evaluation, and supports the analysis of the evolution of camouflage effect over time.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-07
AI Technical Summary
Existing technologies struggle to effectively assess the effectiveness of dynamic camouflage, especially in dynamic environments. Traditional methods rely on manual observation, which is inefficient, has limited accuracy, and is highly subjective. Furthermore, existing computer vision methods lack dynamic adjustment capabilities and cannot accurately identify the camouflage effectiveness of moving targets.
A video segmentation model-based approach is adopted, which calculates dynamic matching score, dynamic contour confusion score and dynamic camouflage comprehensive score by frame-by-frame annotation and feature extraction, combined with COCO2017 index, to achieve quantitative evaluation of the dynamic camouflage effect of moving targets.
It achieves automated and objective quantification of the camouflage effect of moving targets, breaking through the limitations of traditional static image evaluation. It can accurately capture the temporal continuous changes of camouflage effect, providing new technical means for military equipment and animal behavior research.
Smart Images
Figure CN121811292A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the field of video detection, and particularly relates to a method and system for evaluating dynamic camouflage effect of a target. BACKGROUND
[0002] Evaluating the dynamic camouflage effect of a target is particularly crucial in the research of animal ecology, behavior, etc. It not only relates to the analysis of biological survival strategies, but also has a profound impact on military camouflage technology, bionic material development, and ecological protection.
[0003] Animals can survive in the wild through multi-dimensional camouflage such as morphology, color, and behavior. For example, chameleons can dynamically adjust the color of their skin through pigment cells to match the ambient light and temperature, thereby achieving background fusion. The most common way to evaluate the camouflage effect of animals in the past is artificial observation, which mainly relies on the experience of observers to judge the camouflage effect through the "found / not found" two-level indicators. However, this method is inefficient, has limited accuracy, and is highly subjective.
[0004] In recent years, computer vision technology, especially AI technology, has been gradually introduced to empower and innovate the evaluation of animal camouflage effect in ecology. AI technology, with its powerful data processing capability, pattern recognition capability, and automation characteristics, has brought new ideas and methods to ecological research. For example, contour extraction method: using computer image processing technology to automatically identify the contour of the target, and evaluating the camouflage effect according to the recognition result. For another example, model analysis method: constructing a physical model of the camouflage scene, and quantifying the camouflage effect by simulating light, motion, and other parameters. However, in existing technologies, the objects identified are generally static pictures, and the evaluation models are mostly based on static scenes, lacking the ability to dynamically adjust parameters, which leads to a significant decline in recognition accuracy. For example, in rainy weather, the moisture permeability of the camouflage net may change its infrared radiation characteristics, but the evaluation system is difficult to correct in real time. Moreover, static pictures cannot store the motion information in nature, and a large amount of dynamic features are lost in this data modality, which theoretically cannot evaluate the dynamic camouflage effect of objects.
[0005] Although there are a few studies that have begun to evaluate dynamic videos, the existing methods are still based on picture frames and process each frame individually. In the actual information processing process, the algorithm / model still only processes 1 frame of picture at a time, and cannot handle the internal relative relationships such as motion, deformation, and color change of the camouflage target in the context of the entire time sequence. Therefore, it cannot effectively analyze and process the dynamic visual concealment and visual saliency of moving targets, and cannot effectively quantify and score the dynamic camouflage effect.
[0006] Therefore, if a dynamic camouflage effect recognition method for videos can be provided, it will have important research and application value in multiple disciplines and fields. SUMMARY
[0007] The purpose of the present application is to provide a method and system for evaluating the dynamic camouflage effect of a target.
[0008] To achieve the above-mentioned purpose of the application, the technical solution adopted by the present application is as follows: a method for evaluating the camouflage effect of a target, comprising the following steps:
[0009] S1, information collection: collecting video information of the target to be tested, and obtaining visual image data under different time instants and background conditions;
[0010] S2, data cleaning: editing and removing video clips with unqualified recording quality; the unqualified recording video clips include frames in which the target is visually obscured, frames in which part of the target is outside the lens, frames in which the target is not in a camouflage state conforming to the experimental purpose, and frames in which the camera is not successfully focused or has jitter and other poor shooting states;
[0011] S3, reference data set annotation: loading the obtained video clips, and manually annotating the real mask or detection box of the target in the video frame by frame; specifically, the video is converted into a video frame sequence, the mask or detection box annotation of each video frame is completed, and an artificially annotated reference data set for target camouflage evaluation is constructed;
[0012] S4, selection of data to be tested: loading the obtained video clips, jumping to the first frame in which the target to be analyzed appears, selecting and annotating the target to be analyzed, and selecting and annotating the start frame number and end frame number of the target to be analyzed;
[0013] S5, identification of data to be tested: reading, feature extraction, and feature processing of the entire target video segment, outputting the result features and converting them into predicted masks of the target;
[0014] S6, data evaluation processing: calling a visual detection index evaluation tool, loading the reference data set of the target, processing high-dimensional data information, and calculating a camouflage quantitative score. The specific steps include:
[0015] S6-1, using the model original output and the real annotation information output by the video segmentation model to calculate the original COCO2017 index; the original COCO2017 index includes:
[0016] AP@[.50:.95]: average precision with an IoU threshold from 0.5 to 0.95;
[0017] AP@0.50: precision when the IoU threshold is 0.5;
[0018] AP@0.75: precision when the IoU threshold is 0.75;
[0019] AR@[.50:.95]: average recall rate with IoU threshold from 0.5 to 0.95;
[0020] AR@0.50: recall rate with IoU threshold of 0.5;
[0021] AR@0.75: recall rate with IoU threshold of 0.75;
[0022] S6-2, based on the original COCO2017 index, the dynamic matching score DMS, the dynamic contour confusion score DCAS and the dynamic camouflage comprehensive score DCCS are calculated.
[0023] DMS = 1 / (1+exp(k x (AR_composite-x0)));
[0024] AR_composite = w1 x AR@[.50:.95] + w2 x AR@0.50 + w3 x AR@0.75;
[0025] DCAS = 1 / (1+exp(k x (AP_composite-x0)));
[0026] AP_composite = v1 x AP@[.50:.95] + v2 x AP@0.50 + v3 x AP@0.75;
[0027] DCCS = (DMS x DCAS) 0.5 ;
[0028] K and x0 are constants, and w1, w2, w3, v1, v2 and v3 are weights.
[0029] A computer device / apparatus / system, comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to perform the steps of the method.
[0030] A computer readable storage medium having stored thereon a computer program / instruction, which, when executed by a processor, implements the steps of the method.
[0031] A computer program product comprising a computer program / instruction, which, when executed by a processor, implements the steps of the method.
[0032] The present application has the following beneficial effects: the present application is a camouflage evaluation process based on the overall processing of continuous picture frames, can read multiple frames of video for information processing at one time, and with the powerful adaptive modeling capability of the space-time Transformer model, can model and realize the visual concealment and visual saliency of the quantized camouflage target in the whole time sequence context, thereby truly effectively quantizing the dynamic camouflage effect of the moving target.
[0033] The present application can automatically and objectively quantize the algorithm and score for the dynamic or static camouflage of moving or static targets, thereby realizing the low device usage requirement, fast, convenient and relatively accurate quantization of the camouflage effect of the target, and performing objective scoring.
[0034] The present application can continuously and dynamically quantize and analyze the camouflage effect of the moving target based on the dynamic evaluation framework of the streaming video segmentation and memory mechanism, break through the limitations of the traditional static image evaluation, can accurately capture the camouflage effect change of the target in the moving state, realize the quantization description of the time continuity in the dynamic camouflage process, and provide technical support for the study of the time-space characteristics of the camouflage strategy.
[0035] Compared with the traditional machine vision algorithms based on gray similarity, color similarity, texture similarity, patch size similarity, edge similarity and edge fusion degree, the present application uses a deep neural network, can adaptively process different visual information through model training, and utilizes the pre-training technology, without retraining the artificial neural network in each evaluation process, thereby being able to more stably, reliably, conveniently and economically complete the camouflage evaluation task.
[0036] Moreover, compared with the traditional classic machine vision algorithm or CNN, the method of the present application can recognize and process the global visual feature information in the lens, is closer to the recognition mode of the human eye, and is not only based on some local visual information to make the evaluation of the camouflage ability, but is more stable and reliable.
[0037] Compared with the traditional single frame or single frame independent processing mode, the present application provides a camouflage target evaluation based on streaming processing and a video segmentation model with a memory mechanism, the model is based on the visual information of the current frame t, the historical input frame features, i.e., the visual information of the 0th to (t-1)th frames, and the historical output mask features, i.e., the visual information of the 0th to (t-1)th frames, for comprehensive analysis, thereby giving the inference result of the target mask in the tth frame, combining the COCO17 video segmentation model output evaluation index calculation method, to perform more objective score calculation of the camouflage target effect evaluation.
[0038] In conclusion, the method and device provided by the application can effectively identify the dynamic camouflage effect of a target, and have practical application value in the fields of computer vision, military countermeasure technology development, ecological protection of animals with camouflage, and behavior research of animals with camouflage. BRIEF DESCRIPTION OF DRAWINGS
[0039] Figure 1 It is a flowchart of the application;
[0040] Figure 2 It is a software interface diagram based on the method of the application;
[0041] Figure 3 It is a dynamic identification result diagram of Qinghai sand lizard based on the software constructed based on the method of the application. DETAILED DESCRIPTION
[0042] The application provides a method for evaluating the dynamic camouflage effect of a target based on a video instance segmentation base neural network model, and the method flow is as shown in the figure. Figure 1 The method specifically comprises the following steps:
[0043] S1, information collection
[0044] Video information of a target to be measured is collected to obtain visual image data at different time points and under different background conditions. For example, a single or multiple high-resolution video recorders can be used to record the target to be measured.
[0045] S2, data cleaning
[0046] Video segments with unqualified recording quality are edited out, including frames in which the target to be measured is visually blocked (except for special evaluation purposes), frames in which the target to be measured is partially out of the lens, frames in which the target to be measured is not in a camouflage state in line with the experimental purpose, and frames in which the camera is not successfully focused or has shaking, etc.
[0047] S3, reference data set labeling: the obtained video is converted into a video frame sequence using the ffmpeg open source script, and the open source tool Labelme is used to complete the video frame mask or detection box labeling (selected according to the specific experimental purpose), and the open source script x2coco.py is used to make a COCO2017 standard data set, thereby constructing an artificial labeling reference data set for target camouflage evaluation, and completing an objective, quantitative and highly universal animal dynamic and static camouflage function quantitative labeling process.
[0048] S4, loading the video segments after data cleaning in step S2, jumping to the first frame in which the target to be analyzed appears, selecting and labeling the target to be analyzed, and selecting and labeling the start frame number and the end frame number of the target to be analyzed.
[0049] S5, call the video segmentation model pre-trained on large-scale standard datasets in advance. The specific training process is mature prior art, for example, refer to the literature [[1] Ravi N, Gabeur V, Hu Y T, et al. SAM 2: Segment Anything in Images and Videos [J]. 2024.]; using the trained video segmentation model, reading, feature extraction, feature processing of the entire target segment video, finally outputting the result feature and converting it into the prediction mask of the target.
[0050] S6, call the visual detection index evaluation tool (such as the COCO17 evaluation index calculation tool), if the format of the reference dataset manually annotated in step S3 is a mask, load the real mask of the target obtained; if the annotation type of the reference dataset in step S3 is "detection box", the model output video frame mask needs to be converted into bbox format. Then compare and analyze this data with the reference dataset in step S3, and calculate the camouflage quantitative score. That is, if the format of the reference dataset manually annotated is a mask, load the reference mask, and then compare the mask predicted by the model with it; if the reference dataset is a detection box, load the detection box, and then compare the detection box with the detection box. The specific operation steps include:
[0051] S6-1, using the Logits (model original output) and Ground Truth annotation information (real annotation information) output by the video segmentation model, calculate the original COCO2017 index, including:
[0052] AP@[.50:.95]:IoU threshold from 0.5 to 0.95 average precision;
[0053] AP@0.50: precision when IoU threshold is 0.5;
[0054] AP@0.75: precision when IoU threshold is 0.75;
[0055] AR@[.50:.95]:IoU threshold from 0.5 to 0.95 average recall rate;
[0056] AR@0.50: recall rate when IoU threshold is 0.5;
[0057] AR@0.75: recall rate when IoU threshold is 0.75.
[0058] Among them, the IoU threshold (Intersection over Union Threshold) is a key parameter for measuring the degree of overlap between Logits and Ground Truth in object detection and image segmentation tasks, and its core role is to define the standard of "correct prediction". Specifically, IoU is the ratio of the overlapping area of the predicted region and the true region to the union area, and the calculation formula is: IoU = (predicted region U true region) / (predicted region intersection true region); The value is 0 to 1.
[0059] S6-2, based on the original COCO2017 index to calculate Dynamic Matching Score (DMS, Dynamic Matching Score), Dynamic Contour Ambiguity Score (DCAS, Dynamic Contour Ambiguity Score), Dynamic Comprehensive Camouflage Score (DCCS, Dynamic Comprehensive Camouflage Score) Three core scores:
[0060] (1) Dynamic Matching Score (DMS, Dynamic Matching Score):
[0061] DMS = 1 / (1 + exp (k x (AR composite - x0))).
[0062] AR composite = w1 x AR@[.50:.95] + w2 x AR@0.50 + w3 x AR@0.75.
[0063] Where k, x0 are empirical constants for regulating the function form and response characteristics, which can be selected as k = 15.0, x0 = 0.5; w1, w2 and w3 are weights, which can be selected as w1 = 0.6, w2 = 0.25, w3 = 0.15.
[0064] The DMS function has a middle slow transition interval, which will change approximately in a step at AR composite = 0.50. This is because 0.5 has special significance as the classification discriminant value of the machine learning model of most detection functions. For example, when Logits is lower than 0.5, 0.4 and 0.3 have no decisive influence because they will all be classified as negative results. When the score is 0.5, a fuzzy and difficult-to-judge conclusion of "good and bad camouflage effect" appears. The method provided by the present application is specially designed for this purpose, so that when the AR index decreases, the DMS takes about 0.5 as the dividing point, and in the way of first accelerating growth and then deceleration, the value is close to 1; when the AR index increases, the DMS takes about 0.5 as the dividing point, and in the way of first accelerating reduction and then deceleration, the value is close to 0, so that the score of DMS can be as far away from 0.5 as possible, so as to avoid the invalid conclusion of "good and bad camouflage effect" as much as possible, so as to accurately reflect the similarity degree of the target and the background, and quantitatively reflect the visual matching degree of the target. Specifically includes:
[0065] 1) Color matching
[0066] Spectral reflectance: The spectral reflectance curves of the animal body surface (such as fur, scales) and the background (such as leaves, soil) are measured by a spectrophotometer, and the color difference value (such as the Euclidean distance in CIE Lab* color space) is calculated. For example, the body color reflectance of the desert lizard needs to be highly overlapped with the spectrum of the sand particles (especially in the ultraviolet band, many reptiles can perceive ultraviolet).
[0067] Hue, saturation and brightness: Artificial scoring or image analysis software (such as ImageJ) extracts RGB values, compares the consistency of animal main color tone (such as green vs. green leaves), color purity (saturation) and light and dark (brightness) with the background.
[0068] 2) Pattern matching:
[0069] Edge complexity: Fourier transform or wavelet analysis is used to quantify the spatial frequency distribution of the animal body surface pattern (such as leopard pattern, zebra pattern) and the background (such as tree shadow, rock texture). For example, the wing texture of the dead leaf butterfly needs to be consistent with the leaf vein direction and the curvature of the damaged edge of the fallen leaves.
[0070] Patch size and distribution: By counting the patch area, density and arrangement (such as random / regular distribution) of the pattern, and matching the patch characteristics (such as the size of the light spot in the forest is about 5-10 cm) of the background. For example, the patch size on the back of the forest frog needs to be close to the coverage unit of the surrounding moss.
[0071] That is, DMS changes from 1 to 0, the closer to 0, the lower the score, the worse the target camouflage effect, the easier to be found, the lower the color matching degree and the pattern matching degree.
[0072] (2) Dynamic Contour Ambiguity Score (DCAS, dynamic contour ambiguity score):
[0073] DCAS = 1 / (1 + exp(k x (AP composite - x0))).
[0074] AP composite = v1 x AP@[.50:.95] + v2 x AP@0.50 + v3 x AP@0.75.
[0075] wherein k, x0 are empirical constants for regulating the function form and response characteristics, which can be selected as k = 15.0, x0 = 0.5; v1, v2 and v3 are weights, which can be selected as v1 = 0.5, v2 = 0.3, v3 = 0.2.
[0076] The calculation of DCAS adopts the same mathematical structure as DMS but different weight distribution, because in the field of computer vision detection, it is generally believed that AP@[.50:.95] has more important role than AP@0.50 and AP@0.75 in evaluating model performance, while the importance of AP@0.50 and AP@0.75 is generally considered to be similar, which is specially evaluating the ambiguity and uncertainty of target contour. In addition, compared to AR@[.50:.95] to AR, AR pays more attention to recall, and does not pay as much attention to IoU as AP, so the weight is adjusted to reduce the proportion of weight, w1 = 0.6, and v1 = 0.5. Finally, when the AP index decreases, DCAS increases, reflecting that the target contour is difficult to be accurately identified, and the function curve designed in this way can quantitatively reflect the contour ambiguity of the target. Contour ambiguity and the following factors are related:
[0077] 1) Edge fusion degree:
[0078] Contour continuity: measure the overlap ratio of animal body edge (such as head, limbs) and background object (such as branches, stones). For example, the body segment edge of stick insect has the effect of visual coincidence with the segmented part of tree branch, and the contour continuity rate can reach more than 90%. The higher the overlap ratio of animal body edge and background object, the better the contour continuity, and it is more difficult to clearly distinguish the contour boundary of the animal from the visual, so that the contour presents a fuzzy state.
[0079] Pseudo-edge exists: statistics whether there are interference features such as "false eyes" and "false damage" (such as eye-shaped spots at the end of butterfly wings) on the surface of animal body, which can disperse the attention of observer to the real contour. These pseudo-edges can create visual illusions, making it difficult for observers to accurately judge the real contour of the animal, thereby increasing the ambiguity of the contour.
[0080] 2) Three-dimensional stereoscopic adaptation:
[0081] For animals living in complex terrain (such as coral reefs, canopy layer), the body shape (such as flat / cylindrical) needs to be evaluated for its fit with the background structure. For example, the flat body shape of a flatfish matches the two-dimensional distribution of the seabed sand, while the convex skin of a stonefish matches the three-dimensional concave and convex structure of the coral. When the body shape of the animal matches the background structure, the animal and the background are more integrated in three-dimensional space. When viewed from different angles, the outline of the animal will be obscured or interfered with by the background structure, making it difficult to clearly outline its complete outline, resulting in increased contour ambiguity.
[0082] That is, the higher the DCAS score, the more difficult it is to identify the target contour, the higher the edge fusion and three-dimensional stereoscopic adaptation, the better the target camouflage effect; the lower the DCAS score, the worse the target camouflage effect.
[0083] (3) Dynamic Comprehensive Camouflage Score (DCCS, Dynamic Camouflage Score):
[0084] DCCS=(DMS×DCAS) 0.5 .
[0085] The geometric mean is used to integrate the information of the two dimensions of DMS and DCAS, ensuring that only when the target performs well in both visual matching and contour ambiguity can a higher comprehensive score be obtained. This multiplication structure avoids the scoring bias caused by the advantage of a single dimension.
[0086] According to the adjustment of the specific task setting parameters (w1, w2, w3, v1, v2, v3), DMS and DCAS are calculated, and finally DMS and DCAS are used to obtain the dynamic or static camouflage effect score of the target.
[0087] In practice, according to different experimental purposes, different emphasis is placed on DMS, DCAS and DCCS, and the corresponding relationship between the scores of the three indicators and the camouflage effect is also different. Generally speaking, if only the score of DCCS is referred to, it can be considered that: DCCS∈[0.9, 1.0] is excellent in camouflage effect, DCCS∈[0.8, 0.9) is good in camouflage effect, DCCS∈[0.7, 0.8) is medium in camouflage effect, and DCCS∈[0.5, 0.7) is qualified in camouflage effect.
[0088] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. If not specifically indicated, the technical means used in the embodiments are conventional means familiar to those skilled in the art, and the obtained data are all average values obtained after at least three repetitions, and the data obtained in each repetition are all valid data.
[0089] Embodiment: Quantitative evaluation of the dynamic camouflage function effect of Phrynocephalus przewalskii in the sand particle background
[0090] According to the method provided in the present application, a software capable of quantitatively evaluating the dynamic camouflage effect is designed and constructed. The dynamic camouflage function effect of Phrynocephalus przewalskii in different sand particle backgrounds is quantitatively evaluated by using the method and the software. Specifically, the following steps are included:
[0091] S1, data collection: Phrynocephalus przewalskii belonging to Reptilia, Squamata, Sauria, Varanidae and Phrynocephalus was collected from the Ruoergai Wetland Nature Reserve in Aba Tibetan and Qiang Autonomous Prefecture, Sichuan Province; desert sand particles were collected from the original habitat of the species and other places in Sichuan, Xinjiang, Qinghai and Gansu. The collected animals were taken back to the laboratory, and different target camouflage function test environments were constructed using the collected sand particles and special incubators for lizards and other equipment. A professional camera with specifications and parameters meeting the experimental needs was used to record 16 video clips of the natural movement of the lizards in the set experimental environment, ensuring that the camera was fixed and focused well, the recording angle was random, and the width and height of the animal area in the frame accounted for 0.125-0.5 of the frame size.
[0092] S2, data cleaning and preprocessing: the high-quality videos obtained above were cropped to ensure that the video contained about 1:1 of the motion and static behaviors of the camouflage target, and each video was cropped into a 36-frame sub-video. Videos with abnormal photography quality such as blocked camouflage target, failed focusing and serious blur were discarded. Finally, data auditing was performed to ensure that the cleaned videos met the above requirements.
[0093] S3, real label annotation: Labelme open source software was used to annotate the animal real detection box or mask of the video cleaned and processed in step S2, and the x2coco.py script was used to convert the reference dataset annotations-clipped-ground_truth.json file in COCO2017 format.
[0094] S4, artificial intelligence software processing: the video cleaned and processed above was opened and read by using the specially developed Bio-Camouflage Quant (BCQ) software. The software interface is as shown in Figure 2 .
[0095] In the BCQ software, select and add the camouflaged animal target to be quantified through the graphical interface, run the inference process of the video instance segmentation model, obtain the original output result of the video model, preview and confirm in the BCQ software whether the camouflaged target to be quantified is correct, and then click to export the model output result as an annotations-clipped-prediction.json file. Click the Analysis->COCO Analysis button in the upper left corner of the menu bar in the BCQ software, then click the Browse button to load annotations-clipped-ground_truth.json, annotations-clipped-prediction.json respectively, and select the target folder for saving the quantification result, click the Run Analysis button, and the BCQ software will automatically calculate several evaluation indexes of COCO2017, as shown in the following table. Figure 3
[0096] According to the DMS and DCAS definitions and formulas, it is calculated that DMS = 0.9781, DCAS = 0.8169, DCCS = 0.8939, AR composite = 0.2466, and AP composite = 0.4003.
[0097] According to the qualitative quantification grading rule of camouflage: DCCS ∈ [0.9, 1.0] is excellent, DCCS ∈ [0.8, 0.9) is good, DCCS ∈ [0.7, 0.8) is medium, and DCCS ∈ [0.5, 0.7) is qualified. Therefore, the corresponding sand particle background of the embodiment is good in the sand creek of Qinghai. In addition, it can be known from the analysis result that since DMS is greater than DCAS, the camouflage function of the lizard in this environment is dominated by low dynamic visual matching degree rather than dynamic contour confusion degree.
[0098] The above-described embodiments are only preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications, variations, modifications and replacements of the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.
Claims
1. A method for evaluating the camouflage effectiveness of a target, characterized in that, Includes the following steps: S1. Information Acquisition: Acquire video recording information of the target under test and obtain visual image data under different time and background conditions; S2. Data Cleaning: Editing and removing unqualified video clips; S3. Reference dataset annotation: Load the obtained video clips and annotate the real mask or detection box of the target in the video frame by frame; S4. Selection of data to be tested: Load the obtained video clip, jump to the first frame in which the target to be analyzed appears, select and mark the target to be analyzed, and select and mark the start frame number and end frame number of the target to be analyzed. S5. Identification of the data to be tested: Read the entire target segment video, extract features, process features, output the resulting features and convert them into the target's prediction mask; S6. Data Evaluation and Processing: Call the visual inspection index evaluation tool, load the target's reference dataset, process high-dimensional data information, and calculate the camouflage quantitative score.
2. The method according to claim 1, characterized in that: In S2, the unqualified video segments include: frames where the camouflaged target is visually obscured, frames where part of the camouflaged target extends beyond the lens, frames where the camouflaged target is not in a camouflaged state that meets the experimental purpose, and frames where the camera fails to focus or is shaking.
3. The method according to claim 1, characterized in that: In S3, the video is converted into a sequence of video frames, and frame-by-frame masking or bounding box annotation is performed to construct a reference dataset for target camouflage evaluation.
4. The method according to claim 1, characterized in that: Step S6 includes: S6-1. Using the original output of the video segmentation model and the ground truth annotation information, calculate the original COCO2017 index; the original COCO2017 index includes: AP@[.50:.95]: Average accuracy of IoU threshold from 0.5 to 0.95; AP@0.50: Accuracy when the IoU threshold is 0.5; AP@0.75: Accuracy when the IoU threshold is 0.75; AR@[.50:.95]: Average recall rate with IoU thresholds from 0.5 to 0.95; AR@0.50: Recall rate when the IoU threshold is 0.5; AR@0.75: Recall rate when the IoU threshold is 0.75; S6-2. Based on the original COCO2017 index, the Dynamic Matching Score (DMS), Dynamic Contour Confusion Score (DCAS), and Dynamic Disguise Score (DCCS) are calculated.
5. The method according to claim 4, characterized in that: DMS=1 / (1+exp(k×(AR_composite-x0))); AR_composite=w1×AR@[.50:.95]+w2×AR@0.50+w3×AR@0.75; K and x0 are constants, and w1, w2 and w3 are weights.
6. The method according to claim 4, characterized in that: DCAS=1 / (1+exp(k×(AP_composite-x0))); AP_composite=v1×AP@[.50:.95]+v2×AP@0.50+v3×AP@0.75; K and x0 are constants, and v1, v2 and v3 are weights.
7. The method according to any one of claims 4 to 6, characterized in that: DCCS = (DMS × DCAS) 0.5 .
8. A computer device / equipment / system, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.
10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 7.