A camouflage effect automatic quantitative evaluation method based on multi-vision feature extraction and machine learning

Abstract

The application discloses a kind of based on multi-vision feature extraction and machine learning camouflage effect automation quantitative evaluation method, belong to machine vision and camouflage material detection technical field.This method first acquires camouflage sample and background image and completes pre-processing, generates mask image by artificial polygon annotation, and then extracts five big visual features of color, texture, brightness, contour, background;Utilize pearson correlation coefficient to complete feature collinearity check, then calculate each feature weight and camouflage comprehensive score by random forest regression model, finally output standardization evaluation report.The application discards traditional human eye subjective detection mode, realizes camouflage effect automation, full-dimension, objective quantitative evaluation, and the detection result can be reproduced, high efficiency, can provide accurate data support for the performance rating, formula and structure iteration of camouflage equipment and bionic camouflage material, suitable for equipment quality inspection, bionic material research and development and the like scene.

Description

Technical Field

[0001] This invention belongs to the fields of image recognition, machine vision, biomimetic material detection and camouflage equipment performance evaluation. Specifically, it relates to an automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning. It is mainly applied to the performance testing, effect rating and product iteration optimization of camouflage equipment and new camouflage materials. Background Technology

[0002] Camouflage equipment and materials are core products in the field of modern bionics, with camouflage adaptation effectiveness being their core performance indicator. Currently, the industry's methods for testing and evaluating camouflage products have significant shortcomings: traditional evaluation methods heavily rely on subjective human judgment, lacking unified, objective, and quantifiable evaluation standards. Different testers, due to visual differences and subjective judgment biases, will arrive at drastically different evaluation results, leading to extremely poor test repeatability. Furthermore, existing technologies often focus on color and pattern features, neglecting key visual elements such as texture, brightness, contours, and background, failing to comprehensively reflect the true concealment capabilities of camouflage equipment / materials.

[0003] Furthermore, manual inspection is inefficient and cannot meet the needs of large-scale camouflage product testing and multi-scenario simulation testing. It also struggles to output standardized data to support the iterative optimization of camouflage material formulations and equipment appearance structures. The industry urgently needs a multi-dimensional, highly accurate, and automated quantitative evaluation technology solution for camouflage effects. Summary of the Invention

[0004] 3.1 The technical problem to be solved by the present invention

[0005] To address the shortcomings of existing technologies in evaluating camouflage equipment and materials, such as subjective evaluation methods, single evaluation dimensions, low detection efficiency, and lack of quantitative data to support product iteration, this invention provides a five-dimensional visual quantitative evaluation method for camouflage equipment and materials, enabling automated, multi-dimensional, and objective quantitative detection of camouflage effects.

[0006] 3.2 Technical Solution of the Invention

[0007] To achieve the above objectives, the present invention adopts the following technical solution: an automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning, comprising the following steps:

[0008] Step S01: Image acquisition and preprocessing. Acquire the original images of the camouflaged sample to be tested and the corresponding background, and perform resolution unification and noise removal preprocessing on the images.

[0009] Step S02: Manual precise mask annotation. Using a multi-point polygon annotation method, the outline of the sample to be camouflaged is delineated in the preprocessed image to generate a binary mask image, which accurately distinguishes the sample area from the background area.

[0010] Step S03: Five-dimensional visual feature extraction. Based on the mask image, the region is divided, and five core camouflage features are calculated respectively: color similarity, texture similarity, brightness difference, contour complexity, and background complexity.

[0011] Step S04: Feature correlation verification. Pearson correlation coefficient is used to analyze the linear correlation between the five features, eliminate multicollinearity interference, and retain independent and effective features.

[0012] Step S05: Machine learning weight modeling. Using the five major visual features as input, a random forest regression model is constructed. Combining the feature weights output by the model, the overall camouflage score of the sample to be tested is calculated.

[0013] Step S06: Result output and analysis. Output the numerical values ​​of each individual feature, feature weights, and overall camouflage score, and generate a standardized evaluation report for product effectiveness rating and technical optimization.

[0014] Furthermore, in step S03, the specific calculation methods for color similarity, texture similarity, brightness difference, contour complexity, and background complexity are as follows:

[0015] Color similarity: The image is converted to the HSV color space, and the mean values ​​of the H, S, and V channels of the sample and background regions are calculated respectively. The color difference is calculated using Euclidean distance and normalized to obtain the color similarity in the range [0,1]. Texture similarity: The image is converted to a grayscale image, and the pixel variance is used to represent the texture roughness. The texture variance of the sample and background regions is calculated respectively, and the texture similarity is obtained by normalizing the ratio. Brightness difference: The mean grayscale values ​​of the sample and background regions are extracted, and the absolute difference between the two is calculated as the brightness difference index. Contour complexity: The contour complexity is calculated by using the ratio of the square of the contour perimeter to the area to eliminate the interference of the sample shape. Background complexity: The pixel variance of the grayscale image of the background region is calculated to represent the clutter of the simulated background.

[0016] Furthermore, in step S04, if the absolute value of the Pearson correlation coefficient between the two sets of features is greater than 0.7, they are determined to be strongly collinear, and redundant features are eliminated; if the absolute value of the correlation coefficient is less than 0.7, the features are determined to be independent of each other and are retained.

[0017] Furthermore, in step S05, the dataset is divided into a training set and a test set in an 8:2 ratio. After the random forest regression model is trained, the contribution weight of each feature to the camouflage effect is output, and the camouflage comprehensive score of 0 to 100 points is obtained by weighted calculation.

[0018] 3.3 Beneficial Effects of the Invention

[0019] Compared with the prior art, the present invention has the following outstanding advantages:

[0020] Comprehensive evaluation dimensions: A five-dimensional evaluation system is constructed, covering five major visual elements: color, texture, brightness, outline, and background. It closely matches real camouflage scenarios in complex environments, and the evaluation results are more in line with actual usage conditions.

[0021] The evaluation results are objectively quantified: abandoning the subjective scoring mode of human eyes, relying on image recognition and machine learning algorithms to complete the entire process calculation, outputting standardized values ​​and scores, without human bias, and the detection results are reproducible and comparable.

[0022] Significantly improved detection efficiency: The entire process is automated, supporting batch detection of large quantities of disguised samples, solving the problem of low efficiency in manual detection, and adapting to industrial batch detection scenarios.

[0023] Supporting product iteration and optimization: It can output the influence weight and individual data of each feature, accurately locate the performance shortcomings of camouflage equipment and biomimetic materials, provide quantitative data support for material formula adjustment, appearance structure design and texture optimization, and accelerate product technology iteration.

[0024] Highly practical: It is compatible with a variety of products such as camouflage equipment, camouflage materials, and field concealment equipment, and can be deployed in laboratories, quality inspection workshops, and testing institutions, with a wide range of applications. Attached Figure Description

[0025] This invention is accompanied by four accompanying drawings.

[0026] Figure 1 is a flowchart of the overall method of the present invention; Figure 2 is a schematic diagram of image mask annotation; Figure 3 is a heat map of the correlation of five major visual features; Figure 4 is a schematic diagram of feature weight distribution and camouflage comprehensive score output. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0028] 5.1 Overall Process

[0029] As shown in Figure 1, the complete process of this invention is as follows: image acquisition and preprocessing → manual mask annotation → five-dimensional visual feature extraction → feature correlation verification → random forest weight modeling and comprehensive scoring → evaluation result output.

[0030] 5.2 Step S01: Image Acquisition and Preprocessing

[0031] Acquire images of the camouflaged sample to be tested and their corresponding background images. Gaussian denoising and distortion correction are performed on the original images to remove noise and image distortion generated during shooting, ensuring image data consistency. In this embodiment, the image resolution is uniformly set to 1920×1080, and the storage format is PNG.

[0032] 5.3 Step S02: Manual Precise Mask Labeling

[0033] The polygon annotation program is invoked, and the operator uses the mouse to outline the outer contour of the camouflaged sample point by point. The program automatically generates a closed polygon and fills it with a binary mask image of the sample (white) and the background (black) (as shown in Figure 2). The mask image corresponds one-to-one with the original image, achieving pixel-level region segmentation, completely avoiding edge errors of automatic segmentation algorithms, and ensuring the accuracy of subsequent feature extraction.

[0034] 5.4 Step S03: Five-dimensional visual feature extraction.

[0035] Based on the mask image, the sample region and background region are divided, and five major features are calculated sequentially:

[0036] Color similarity: The image is converted from RGB format to HSV color space, and the hue, saturation, and lightness channels are separated. The channel mean values ​​of all pixels in the sample area and the background area are calculated separately. The Euclidean distance between the two sets of means is calculated, and after normalization, the color similarity is obtained. The closer the value is to 1, the better the color matching effect.

[0037] Texture similarity: Convert the image to grayscale and calculate the pixel variance of the sample region and the background region separately. Variance represents the texture roughness. The texture similarity is obtained by normalizing the ratio of the two sets of variances. The higher the value, the stronger the fit between the sample and the background texture.

[0038] Brightness difference: Calculate the absolute difference between the mean gray values ​​of the sample area and the background area. The smaller the difference, the more natural the transition between light and dark, and the better the camouflage effect.

[0039] Contour complexity: The contour detection algorithm is called to extract the outer contour of the sample and the contour perimeter² / contour area is calculated. This index eliminates the influence of sample size. The higher the value, the more complex the contour and the lower the visual recognition.

[0040] Background complexity: Calculate the variance of grayscale pixels in the background region. The larger the variance, the more cluttered the simulated background, and the stronger the tolerance for sample camouflage.

[0041] Step S04: Feature Correlation Verification

[0042] The Pearson correlation coefficient was used to calculate the degree of linear correlation between each pair of the five features, generating a correlation matrix and a heatmap (as shown in Figure 3). The criteria were: if the absolute value of the correlation coefficient was >0.7, it was considered strong collinearity, and redundant features were removed; if the absolute value was ≤0.7, the features were considered independent.

[0043] In this embodiment, only color similarity and brightness difference have a strong negative correlation (r=-0.73), while the other features are weakly correlated or uncorrelated. All five features are retained for modeling.

[0044] 5.6 Step S05: Machine Learning Weight Modeling and Comprehensive Scoring

[0045] Dataset partitioning: All sample feature data are randomly divided into training and test sets in an 8:2 ratio; Model training: A random forest regression model is constructed, using five major visual features as input variables and manually labeled camouflage effects as labels to complete model training; Weight output: After model training, the contribution weights of the five major features to the camouflage effect are automatically output; In this embodiment, texture similarity has the highest weight at 66.78%, followed by color similarity and brightness difference, while contour complexity and background complexity have weights of less than 4%; Comprehensive score: Based on the weighted calculation of feature weights, a comprehensive camouflage score of 0-100 points is output.

[0046] 5.7 Step S06: Result Output and Analysis

[0047] The system automatically generates visual charts (feature weight map, score distribution map, as shown in Figure 4), feature data tables, and evaluation reports. The reports include individual feature values, feature weights, overall scores, and performance ratings. Based on the reports, testing personnel determine the performance of the disguised product and optimize the design of biomimetic material textures, colors, and appearance contours for the dimensions corresponding to low-scoring features, thus completing product iteration.

Claims

1. An automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning, characterized in that, Includes the following steps: Step S01: Image acquisition and preprocessing. Acquire the original images of the camouflaged sample to be tested and the corresponding background, and perform noise reduction and resolution unification preprocessing on the original images. Step S02: Manual precise mask annotation. The outline of the sample to be camouflaged is delineated using a multi-point polygon annotation method to generate a binary mask image and distinguish the sample area from the background area. Step S03: Five-dimensional visual feature extraction. Based on the binary mask image, calculate five camouflage features: color similarity, texture similarity, brightness difference, contour complexity, and background complexity. Step S04: Feature correlation verification. Pearson correlation coefficient is used to analyze the linear correlation of the five features and eliminate redundant features with multicollinearity. Step S05: Machine learning weight modeling, using the selected features as input to construct a random forest regression model, and calculating the overall camouflage score of the sample to be tested based on the feature weights output by the model; Step S06: Output the results. Output the numerical values ​​of each individual feature, feature weights, and camouflage comprehensive score, and generate a standardized evaluation report.

2. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, The color similarity calculation method in step S03 is as follows: the image is converted into the HSV color space, the mean values ​​of the H, S, and V channels of the sample area and the background area are calculated respectively, the color difference is calculated by Euclidean distance and normalized, and the color similarity with a value range of [0,1] is obtained.

3. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, The texture similarity calculation method in step S03 is as follows: convert the image into a grayscale image, calculate the pixel variance of the sample area and the background area respectively, and obtain the texture similarity by normalizing the variance ratio.

4. The automated quantitative evaluation method for camouflage effect based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S03, the brightness difference is the absolute difference between the average gray level of the sample area and the average gray level of the background area.

5. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S03, the contour complexity is calculated using the ratio of the square of the contour perimeter to the contour area.

6. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S03, the background complexity is the pixel variance of the grayscale image of the background region.

7. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S04, when the absolute value of the Pearson correlation coefficient of two sets of features is greater than 0.7, they are determined to be strongly collinear and are removed; when the absolute value is less than or equal to 0.7, they are determined to be independent of each other and are retained.

8. The automated quantitative evaluation method for camouflage effects based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S05, the entire feature dataset is divided into a training set and a test set in an 8:2 ratio for training and validation of the random forest regression model.

9. The automated quantitative evaluation method for camouflage effect based on multi-visual feature extraction and machine learning according to claim 1, characterized in that, In step S01, the preprocessing also includes image distortion correction to ensure that all image data are of uniform specifications.

Images