Five-dimensional sample optimization method based on image data quality evaluation

Abstract

The invention relates to the field of image quality evaluation, in particular to a five-dimensional sample optimization method based on image data quality evaluation. According to the method, four different indexes are balanced according to index distribution characteristics: maximum-minimum normalization and logarithm normalization are firstly carried out, then an increasing scoring function, a decreasing scoring function and a Gaussian scoring function are selected according to different high-quality index characteristics for scoring, and finally, a final quality score is calculated according to weight distribution, so that sample optimization is completed. And the optimized samples are used to participate in the training process, so that the accuracy of deep learning is improved.

Description

Technical Field

[0001] The present invention relates to the field of image quality assessment, and in particular designs a five-dimensional sample optimization method based on image data quality assessment. Background Art

[0002] The lack of high-quality data has long plagued the improvement of the accuracy of computer vision models. In the current existing technology, the evaluation of the quality of sample set images is generally carried out using full-reference algorithms, semi-reference algorithms, and no-reference algorithms for distorted image quality evaluation. Because the samples used for training in the field of target detection have different focuses from those of image distortion algorithms, there are still small models with self-collected samples, which have problems such as difficulty in finding authoritative data sets as standard reference images, difficulty in balancing the quality and quantity of samples, and sample selection mainly relying on manual experience. Different deep learning fields have different requirements for training set samples. In the field of pedestrian detection in semi-outdoor garbage booths, it is necessary to focus on large and significant targets to avoid misjudgment as much as possible. Therefore, the sample training set needs to have clear annotations, high information abundance, clear contours and strong contrast, and sample images should minimize distortion or noise pollution.

[0003] In the field of image data quality assessment, the present invention introduces an image data quality assessment mechanism in the sample preprocessing stage, quantifies sample quality through suitable measurement, reduces the redundancy of image data, and increases the information density of feature or embedded representation. This improves the detection performance of the recognition model with less and more precise data and reduces the negative impact of low-quality samples on the target detection model. Summary of the invention

[0004] In order to overcome the deficiencies of the prior art, the present invention provides a five-dimensional sample optimization method based on image data quality assessment.

[0005] The present invention is achieved through the following technical solutions:

[0006] A five-dimensional sample selection method based on image data quality assessment, the method comprising the following steps:

[0007] Step 1: The research object is the data in the baseline sample space. A two-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio is used to calculate the information entropy measurement index IQA_ImageEntropy and the signal-to-noise ratio measurement index IQA_SNR for each sample.

[0008] Step 2: Use a no-reference image quality index calculation mechanism based on an improved version of NRLT. NRLT is a method for evaluating screen content image quality in the field of SCIs (Screen Content Images). The sample image quality evaluation field to which the present invention is applied has been improved in terms of the number of channels, gradient brightness calculation, and final quality score calculation. The implementation process of this mechanism is to first extract the edge features of the sample image, and calculate the frequency distribution histogram of the absolute value of the edge map E(x, y). E(x, y) represents the edge map of the sample image (x, y) coordinate position, which can be represented by a 10-dimensional vector {f1, f2, ..., f 10} is used as the feature vector representing the texture. The dimension can be freely selected. If it is too small, it cannot fully represent the edge features of the image, and if it is too large, it increases the difficulty of calculation. It is generally set to 10. Then extract the sample brightness feature and calculate the frequency distribution histogram of the absolute value of the normalized image I′(x, y). Use an i-dimensional vector {f1, f2, …, f i} as the feature vector representing brightness, where the dimension i is generally set to 4 times the dimension of the feature vector representing texture; finally, the quality score of the feature vector mean is used as the measurement indicator IQA_NRLT;

[0009] Step 3: Use the Meta-AI-based SAM semi-reference image quality objective evaluation index calculation mechanism to parse the annotations of the XML file corresponding to the jpg image through the SAM model to obtain the coordinates of the bounding box oxes , as the input prompt box for generating the mask, segment the specific object in the annotation box and generate a pixel-level mask M i , calculate the sum of the mask areas of all the annotation boxes of a single image as the measurement indicator IQA_SAM;

[0010] Step 4: Propose a mathematical balancing mechanism to balance the four different indicators into the final quality score according to the characteristics of the indicator distribution: first perform maximum-minimum normalization and logarithmic normalization, then select increasing scoring function, decreasing scoring function and Gaussian scoring function for scoring according to the characteristics of different high-quality indicators, and finally calculate the final quality score based on the weight distribution to complete the sample optimization.

[0011] Preferably, in step 1, the dual-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio comprises the following specific steps:

[0012] The grayscale values ​​of each pixel in the image are independent of each other, and the grayscale distribution of the image is p = {p1, p2, ..., p i , ..., p n}, where p iIt is the ratio of the number of pixels with gray value i to the total number of pixels in the image, and n is the total number of gray levels. The calculation formula is:

[0013]

[0014] Among them, p i is the probability of gray value i appearing in the sample; L is the gray level of the image; the gray level of the sample image of the present invention is 256, so L=255.

[0015] The signal amplitude is represented by the average value of the whole image or the selected area, and the noise amplitude is represented by the standard deviation. Then the ratio of the mean value to the standard deviation is calculated. The formula is:

[0016]

[0017] Among them, P signal Indicates the signal amplitude, P noise represents the noise amplitude; P mean Represents the average value of the entire image or the selected area, P std. Indicates the standard deviation of the entire image or selected area.

[0018] Preferably, in step 2, a non-reference image quality index calculation mechanism based on an improved version of NRLT is adopted, and the specific steps are:

[0019] First, extract the edge features of the sample image. The calculation formula for calculating the sample edge map is as follows:

[0020]

[0021] The edge map E(x, y) is obtained by combining the image I(x, y) with the Gaussian difference kernel G σ1，σ2 (x, y) is convolved and the absolute value of the result is taken; σ1, σ2 are the image standard deviations corresponding to the two Gaussian kernels respectively, and the Gaussian difference kernel G σ1，σ2 (x, y) is the difference between two Gaussian kernels with different standard deviations, and each Gaussian kernel G(x, y; σ) is calculated based on a given standard deviation σ, and the calculation formula is as follows:

[0022] G σ1.σ2 (x,y)=G(x,y;σ1)-G(x,y;σ2)

[0023] The Gaussian kernel calculation formula for the sample image standard deviation σ is:

[0024]

[0025] The edge features of the sample image are obtained by calculating the frequency distribution histogram of the absolute value of the edge map E(x, y), using a 10-dimensional vector {f1, f2, ..., f10} as the feature vector representing the texture, the formula is as follows:

[0026]

[0027] The histogram is divided into 10 groups. fk represents the edge features of the sample image. X and Y represent the length and width of the image, respectively. The value range of each group is represented by z(k), where k is the element index in the feature vector, k∈[1,10]. The calculation of each group of the histogram is completed by accumulating the number of times each pixel in the image belongs to the group. First, the absolute value of the edge map is calculated, and then a 10-dimensional feature vector is initialized according to the size of the image. Next, the value range of the histogram group is calculated, and its frequency is calculated for each group. Φ(|E(x,y)|,z(k)) is an indicator function used to determine whether the pixel value is within the current group range of the histogram. In addition, Φ(a,b) is an indicator function. If a belongs to b, 1 is output, otherwise 0 is output. a and b are abstract variable names for ease of understanding. Here, it is essentially Φ(|E(x,y)|,z(k)), which is used to determine whether the pixel value is within the current group range of the histogram. If it is, 1 is output, and if it is not, 0 is output.

[0028] Then, the brightness features of the sample image are extracted: the image I(x, y) cannot be used directly for brightness extraction, and needs to be normalized first. The process of local normalization of the sample image includes calculating the local mean μ and standard deviation σ of each pixel, and then using these values ​​to calculate the normalized image I′(x, y).

[0029] The calculation formula of I′(x, y) is as follows:

[0030]

[0031]

[0032] Among them, C is a constant used to prevent the denominator from being 0; the normalized window size is controlled by M and L, and in this embodiment, M=L=7 is preset; ω(m, l) is a two-dimensional circularly symmetric Gaussian weighting function to perform weight distribution of local mean and standard deviation, and m and are variable names starting from the normalized window sizes M and L. Taking m as an example, it can indicate accumulation from -M to M. The brightness feature of the sample image is obtained by calculating the frequency distribution histogram of the absolute value of the normalized image I′(x, y), which can be expressed by an i-dimensional vector {f1, f2, ..., f i} is used as the feature vector representing brightness. The calculation process is basically the same as edge feature extraction, where Φ|I′(x, y)|, z(i)) is an indicator function used to determine whether the pixel value is within the current grouping range of the brightness map histogram. i represents the i-th element in the feature vector, and z(i) represents the value range of the i-th group of the histogram.

[0033] The formula is:

[0034]

[0035] Finally, the two dimensions of illumination and texture are combined into one to calculate the quality score. The mean of the feature vector is used as the quality score to reflect the high illumination and high texture of the sample image. The calculation formula is as follows:

[0036]

[0037] Where N is the total number of edge features of the sample image; M is the total number of brightness features of the sample image; fk is the edge feature value of the sample image; f i is the brightness eigenvalue of the sample image; IQA_NRLT is the signal-to-noise ratio measurement indicator.

[0038] Preferably, in step three, a SAM semi-reference image quality objective evaluation index calculation mechanism based on Meta-AI is adopted, and the specific steps are as follows:

[0039] Let, be the input JPG image, X be the corresponding XML file, B i is the coordinate of the i-th bounding box, M i corresponds to the bounding box B i The mask of A i The mask M i Area:

[0040] Parse the annotation of the XML file corresponding to the jpg image through the SAM model XML(X) , get the coordinates of the bounding box boxes :

[0041] Bounding boxes =parse XML(X)

[0042] Generate the input prompt box input_boxes of the mask, where create_input_box(B i ) indicates that the coordinates of multiple bounding boxes contained in a single sample are used to generate annotation prompt boxes:

[0043] input_boxes={create_input_box(B i)|B i ∈bounding boxes}

[0044] Segment the specific object in the annotation box and generate a pixel-level mask M i , the image is segmented into different regions and converted into a binary image, where pixels with a value of 1 represent the region belonging to the segmented object, and pixels with a value of 0 represent the background or other objects. Where I is the picture input to the SAM model:

[0045] {M i |M i is the mask for B i = SAM(I, input_boxes)

[0046] Calculate the sum of the mask areas of all the annotation boxes in a single image, where n is the number of annotation boxes, and A i is the mask area corresponding to the i-th annotation box:

[0047]

[0048] Measurement indicators for quantitative samples:

[0049] IQA_SAM=total_area

[0050] Preferably, in step 4, a mathematical balance mechanism is adopted, and the specific steps are:

[0051] The IQA_ImageEntropy and IQA_SNR metrics select min-max normalization:

[0052]

[0053] The IQA_NRLT and IQA_SAM indicators use log normalization:

[0054]

[0055] After normalization, the gaussian_score Gaussian score is used for the two indicators whose values ​​need to be balanced, the IQA_ImageEntropy indicator and the IQA_NRLT indicator:

[0056]

[0057] in: μ represents the mean and σ represents the variance.

[0058] For the IQA_SNR indicator, where the larger the value, the better, increasing_score is used:

[0059]

[0060] For the IQA_NRLT indicator, where the smaller the value, the better, the decreasing_score is used:

[0061]

[0062] After normalization and scoring function calculation, the four indicators are unified into the range of 0-1. The larger the indicator value, the higher the sample quality. The final quality score is obtained by allocating weights of 0.25, 0.25, 0.25, and 0.25. According to the final quality score results, high-quality samples are divided into A, B, and C categories, and low-quality samples are divided into D category.

[0063] The mathematical process to achieve the indicator balance mechanism is:

[0064] IQA_ImageEntrop indicator:

[0065]

[0066] in,

[0067] IQA_SNR indicator:

[0068]

[0069] IQA_NRLT indicator:

[0070]

[0071] IQA_SAM indicator:

[0072]

[0073] in,

[0074] Final quality score S:

[0075] S=w1·f1+w2·f2+w3·f3+W4·f4

[0076] Among them, the weight w1=w2=w3=W4=0.25.

[0077] Compared with the prior art, the present invention has the following advantages:

[0078] (1) The present invention processes a batch of data in the baseline sample space, calculates the information entropy measurement index IQA_ImageEntropy and the signal-to-noise ratio measurement index IQA_SNR for each sample; then extracts the edge features and brightness features of the sample image and uses the quality score of the mean of the feature vector as the measurement IQA_NRLT; parses the annotation of the XML file corresponding to the jpg image through the SAM model, obtains the coordinates of the bounding box as the input prompt box for generating the mask, segments the specific objects in the annotation box and generates a pixel-level mask, and evaluates the relative area of ​​the mask as the measurement IQA_SAM; so that the samples in the preferred sample space have five-dimensional characteristics of significance (moderate area of ​​the annotation area after segmentation), diversity (moderate sample information entropy), noise resistance (good signal-to-noise ratio index), high illumination and high texture (strong target contrast and detailed contour).

[0079] (2) The present invention balances the four different indicators obtained according to the characteristics of indicator distribution: first, maximum-minimum normalization and logarithmic normalization are performed, and then the increasing scoring function, decreasing scoring function and Gaussian scoring function are selected for scoring according to the characteristics of different high-quality indicators. Finally, the final quality score is calculated according to the weight distribution to complete the sample optimization. The optimized samples are used to participate in the training process, which improves the accuracy of deep learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] The present invention will be further described below in conjunction with the accompanying drawings.

[0081] Figure 1 The architecture diagram of the invented five-dimensional image data quality assessment mechanism;

[0082] Figure 2 A framework diagram for designing and implementing the invented five-dimensional sample selection method based on image data quality assessment;

[0083] Figure 3 It is the histogram of AP (evaluation index in machine vision) value of each category in the five-dimensional optimization mechanism;

[0084] Figure 4 Optimize the person class AP graph for five dimensions;

[0085] Figure 5 Optimize the garbage-bag AP graph in five dimensions;

[0086] Figure 6 Optimize the worker class AP graph for five dimensions;

[0087] Figure 7 It is the five-dimensional optimized suliaozhipin class AP graph;

[0088] Figure 8Optimize the AP graph of Zhizhipin class in five dimensions;

[0089] Fig. 9 It is the AP graph of each category of the five-dimensional optimization mechanism. DETAILED DESCRIPTION

[0090] The technical solution of the present invention is further specifically described below through embodiments and in conjunction with the accompanying drawings.

[0091] The method is used to select samples for the pedestrian detection field (PD field) of the semi-outdoor garbage booth, the garbage detection field (GD field) and the logo detection field (LD field) supplemented by generalization performance detection. The method includes the following steps:

[0092] Step 1: The research object is the jpg files of the baseline sample space in various fields. A two-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio is adopted to calculate the information entropy measurement index IQA_ImageEntropy and the signal-to-noise ratio measurement index IQA_SNR for each sample.

[0093] Step 2: Using the non-reference image quality index calculation mechanism based on the improved version of NRLT, first extract the edge features of the sample image and calculate the frequency distribution histogram of the absolute value of the edge map E(x, y), which can be obtained by a 10-dimensional vector {f1, f2, ..., f 10} as the feature vector representing the texture; then extract the sample brightness feature and calculate the frequency distribution histogram of the absolute value of the normalized image I′(x, y), which can be obtained by an i-dimensional vector {f1, f2, …, f i} as the feature vector representing brightness; finally, the quality score of the feature vector mean is used as the measurement indicator IQA_NRLT.

[0094] Step 3: Use the Meta-AI-based SAM semi-reference image quality objective evaluation index calculation mechanism to parse the annotations of the XML file corresponding to the jpg image through the SAM model to obtain the coordinates of the bounding box. boxes , as the input prompt box for generating the mask, segment the specific object in the annotation box and generate a pixel-level mask M i , calculate the sum of the mask areas of all annotated boxes in a single image as the measurement indicator IQA_SAM.

[0095] Step 4: Propose a mathematical balancing mechanism to balance the four different indicators into the final quality score according to the characteristics of the indicator distribution: first perform maximum-minimum normalization and logarithmic normalization, then select increasing scoring function, decreasing scoring function and Gaussian scoring function for scoring according to the characteristics of different high-quality indicators, and finally calculate the final quality score based on the weight distribution to complete the sample optimization.

[0096] In step 1, the two-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio has the following specific steps:

[0097] The grayscale values ​​of each pixel in the image are independent of each other, and the grayscale distribution of the image is p = {p1, p2, ..., p i , ..., p n}, where p i It is the ratio of the number of pixels with gray value i to the total number of pixels in the image, and n is the total number of gray levels. The calculation formula is:

[0098]

[0099] Among them, p i is the probability of gray value i appearing in the sample; L is the gray level of the image. In this embodiment, the gray level of the sample image is 256, so L=255.

[0100] The average value of the entire image (or the selected area) is used to represent the signal amplitude, the standard deviation is used to represent the noise amplitude, and then the ratio of the mean to the standard deviation is calculated. The formula is:

[0101]

[0102] In step 2, a no-reference image quality index calculation mechanism based on the improved version of NRLT is adopted, and the specific steps are as follows:

[0103] First, extract the edge features of the sample image. The calculation formula for calculating the sample edge map is as follows:

[0104]

[0105] The edge map E(x, y) is obtained by combining the image I(x, y) with the Gaussian difference kernel G σ1，σ2 (x, y) is convolved and the absolute value of the result is taken. Gaussian difference kernel G σ1，σ2 (x, y) is the difference between two Gaussian kernels with different standard deviations, and each Gaussian kernel G(x, y; σ) is calculated based on a given standard deviation σ, and the calculation formula is as follows:

[0106] G σ1.σ2 (x,y)=G(x,y;σ1)-G(x,y;σ2)

[0107] The Gaussian kernel with standard deviation σ is calculated as:

[0108]

[0109] The edge features of the sample graph are obtained by calculating the frequency distribution histogram of the absolute value of the edge graph E(x, y), which can be expressed as a 10-dimensional vector {f1, f2, ..., f 10} as the feature vector representing the texture, the formula is as follows:

[0110]

[0111] The histogram is divided into 10 groups, and the range of each group is represented by z(k), where k is the element index in the feature vector, k∈[1,10]. The calculation of each group of the histogram is completed by accumulating the number of times each pixel in the image belongs to this group. First, the absolute value of the edge map is calculated, and then a 10-dimensional feature vector is initialized according to the size of the image. Next, the range of the histogram group is calculated, and its frequency is calculated for each group. Φ(|E(x,y)|,z(k)) is an indicator function used to determine whether the pixel value is within the current group range of the histogram.

[0112] Then, the brightness feature of the sample image is extracted: the image I (x, y) cannot be used directly for brightness extraction, and needs to be normalized first. The process of local normalization of the sample image includes calculating the local mean μ and standard deviation σ of each pixel, and then using these values ​​to calculate the normalized image I′ (x, y). The normalization window size is controlled by m and l. In this embodiment, m=l=7 is preset, and a two-dimensional circularly symmetric Gaussian weighting function ω (m, l) is used to weight the local mean and standard deviation.

[0113] The calculation formula of I′(x, y) is as follows:

[0114]

[0115] The brightness feature of the sample image is obtained by calculating the frequency distribution histogram of the absolute value of the normalized image I′(x, y), which can be expressed as an i-dimensional vector {f1, f2, …, f i} is used as the feature vector representing brightness, and the calculation process is basically the same as edge feature extraction. The formula is:

[0116]

[0117] Finally, the two dimensions of illumination and texture are combined to calculate the quality score, and the mean of the feature vector is used as the quality score to reflect the high illumination and high texture of the sample image. The calculation formula is as follows:

[0118]

[0119] In step 3, the Meta-AI-based SAM semi-reference image quality objective evaluation index calculation mechanism is adopted, and the specific steps are as follows:

[0120] Let, be the input JPG image, X be the corresponding XML file, B i is the coordinate of the i-th bounding box, M i corresponds to the bounding box B i The mask of A i The mask M i Area:

[0121] Use the SAM model to parse the annotations of the XML file corresponding to the jpg image and obtain the coordinates of the bounding box:

[0122] Bounding boxes =parsexML(X)

[0123] Generate a mask input prompt box:

[0124] input_boxes={create_input_box(B i )|B i ∈bounding_boxes}

[0125] The specific object in the annotation box is segmented and a pixel-level mask is generated. The image is segmented into different regions and converted into a binary image, where pixels with a value of 1 represent the area belonging to the segmented object and pixels with a value of 0 represent the background or other objects:

[0126] {M i |M i is the mask for B i} = SAM(I, input boxes)

[0127] Calculate the sum of the mask areas of all the annotation boxes in a single image, where n is the number of annotation boxes, and A i is the mask area corresponding to the i-th annotation box:

[0128]

[0129] Measurement indicators for quantitative samples:

[0130] IQA_SAM=total_area

[0131] In step 4, a mathematical balance mechanism is used, and the specific steps are as follows:

[0132] The IQA_ImageEntropy and IQA_SNR metrics select min-max normalization:

[0133]

[0134] The IQA_NRLT and IQA_SAM indicators use log normalization:

[0135]

[0136] After normalization, the gaussian_score Gaussian score is used for the two indicators whose values ​​need to be balanced, the IQA_ImageEntropy indicator and the IQA_NRLT indicator:

[0137]

[0138] in:

[0139] For the IQA_SNR indicator, where the larger the value, the better, increasing_score is used:

[0140]

[0141] For the IQA_NRLT indicator, where the smaller the value, the better, the decreasing_score is used:

[0142]

[0143] After normalization and scoring function calculation, the four indicators are unified into the range of 0-1. The larger the indicator value, the higher the sample quality. The final quality score is obtained by allocating the weights of 0.25, 0.25, 0.25, and 0.25. The high-quality samples (A, B, and C categories) and low-quality samples (D category) are divided according to the final quality score results.

[0144] The mathematical process to achieve the indicator balance mechanism is:

[0145] IQA_ImageEntrop indicator:

[0146]

[0147] in,

[0148] IQA_SNR indicator:

[0149]

[0150] IQA_NRLT indicator:

[0151]

[0152] IQA_SAM indicator:

[0153]

[0154] in,

[0155] Final quality score S:

[0156] S=w1·f1+w2·f2+w3·f3+w4·f4

[0157] Among them, the weight w1=w2=w3=w4=0.25.

[0158] The optimized samples are used in the training process. The experimental results of the PD field dataset are shown in the figure:

[0159] The total number of PD samples is 1212. Based on the baseline model, the top 75% of the samples with the final quality score S were selected, and 404 D-class samples (accounting for 25% of the total number of samples) were eliminated. The training set: test set = 9:1 ratio was also used for the division. The final mAP obtained after training and evaluation was 92.15%, where mAP is the average of the AP values ​​of each category, and the larger the value, the better. This indicator can more intuitively and concisely reflect the accuracy of the training results in target detection.

[0160] Table 1 AP value table of each category of five-dimensional image data quality evaluation optimization mechanism

[0161]

[0162] The AP index values ​​of each major category are as follows Figure 3-9 , where the dotted line is the baseline model.

[0163] Figure 3 The results are the AP values ​​selected after five-dimensional image data quality assessment compared with the baseline. The results show that each category has been improved to a certain extent. Figure 4-9The following is a line chart comparing the AP value improvements of each category after five-dimensional optimization. The person category increased to 95.67%, an increase of 5.67 percentage points; the garbage-bag category increased to 88.76%, an increase of 7.83 percentage points; the worker category increased to 99.15%, an increase of 7.75 percentage points; the suliaozhipin category increased to 92.04%, an increase of 14.74 percentage points; and the zhizhipin category increased to 85.14%, an increase of 0.8 percentage points. This shows that the five-dimensional image data quality assessment optimization mechanism is basically effective as a whole, and the effect is quite significant.

Claims

1. A five-dimensional sample selection method based on image data quality assessment, characterized by: The method comprises the following steps: Step 1: The research object is the data in the baseline sample space. A two-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio is used to calculate the information entropy measurement index IQA_ImageEntropy and the signal-to-noise ratio measurement index IQA_SNR for each sample. Step 2: Using the non-reference image quality index calculation mechanism based on the improved version of NRLT, first extract the edge features of the sample image and calculate the frequency distribution histogram of the absolute value of the edge map E(x, y), which can be obtained by a 10-dimensional vector {f1, f2, ..., f 10 } as the feature vector representing the texture; then extract the sample brightness feature, calculate the frequency distribution histogram of the absolute value of the normalized image I′(x, y), and use an i-dimensional vector {f1, f2, …, f i } as the feature vector representing brightness; finally, the quality score of the feature vector mean is used as the measurement indicator IQA_NRLT; Step 3: Use the Meta-AI-based SAM semi-reference image quality objective evaluation index calculation mechanism to parse the annotations of the XML file corresponding to the jpg image through the SAM model to obtain the coordinates of the bounding box. boxes , as the input prompt box for generating the mask, segment the specific object in the annotation box and generate a pixel-level mask M i , calculate the sum of the mask areas of all the annotation boxes of a single image as the measurement indicator IQA_SAM; Step 4: Propose a mathematical balancing mechanism to balance the four different indicators into the final quality score according to the characteristics of the indicator distribution: first perform maximum-minimum normalization and logarithmic normalization, then select increasing scoring function, decreasing scoring function and Gaussian scoring function for scoring according to the characteristics of different high-quality indicators, and finally calculate the final quality score based on the weight distribution to complete the sample optimization.

2. The five-dimensional sample selection method based on image data quality assessment according to claim 1, characterized in that: In step 1, the dual-dimensional image data quality assessment mechanism based on information entropy and signal-to-noise ratio has the following specific steps: The grayscale values ​​of each pixel in the image are independent of each other, and the grayscale distribution of the image is p = {p1, p2, ..., p i ,…,p n }, where p i It is the ratio of the number of pixels with gray value i to the total number of pixels in the image, and n is the total number of gray levels. The calculation formula is: Among them, p i is the probability of gray value i appearing in the sample; L is the gray level of the image; in this embodiment, the gray level of the sample image is 256, so L=255. The signal amplitude is represented by the average value of the whole image or the selected area, and the noise amplitude is represented by the standard deviation. Then the ratio of the mean value to the standard deviation is calculated. The formula is: Among them, P signal Indicates the signal amplitude, P noise represents the noise amplitude; P mean Represents the average value of the entire image or the selected area, P std. Indicates the standard deviation of the entire image or selected area.

3. The five-dimensional sample selection method based on image data quality assessment according to claim 1, characterized in that: In step 2, a no-reference image quality index calculation mechanism based on the improved version of NRLT is adopted, and the specific steps are as follows: First, extract the edge features of the sample image. The calculation formula for calculating the sample edge map is as follows: The edge map E(x, y) is obtained by combining the image I(x, y) with the Gaussian difference kernel G σ1，σ2 (x, y) is convolved and the absolute value of the result is taken; σ1, σ2 are the image standard deviations corresponding to the two Gaussian kernels respectively, and the Gaussian difference kernel G σ1，σ2 (x, y) is the difference between two Gaussian kernels with different standard deviations, and each Gaussian kernel G(x, y; σ) is calculated based on a given standard deviation σ, and the calculation formula is as follows: G σ1.σ2 (x,y)=G(x,y;σ1)-G(x,y;σ2) The Gaussian kernel calculation formula for the sample image standard deviation σ is: The edge features of the sample graph are obtained by calculating the frequency distribution histogram of the absolute value of the edge graph E(x, y), using a 10-dimensional vector {f1, f2, ..., f 10 } as the feature vector representing the texture, the formula is as follows: The histogram is divided into 10 groups, X and Y represent the length and width of the image respectively. The value range of each group is represented by z(k), where k is the element index in the feature vector, k∈[1,10]. The calculation of each group of the histogram is completed by accumulating the number of times each pixel in the image belongs to the group. First, the absolute value of the edge map is calculated, and then a 10-dimensional feature vector is initialized according to the size of the image. Next, the value range of the histogram group is calculated, and the frequency of each group is calculated. Φ(|E(x, y)|, z(k)) is an indicator function used to determine whether the pixel value is within the current group range of the histogram. Then, the brightness features of the sample image are extracted: the image I(x, y) cannot be used directly for brightness extraction, and needs to be normalized first. The process of local normalization of the sample image includes calculating the local mean μ and standard deviation σ of each pixel, and then using these values ​​to calculate the normalized image I′(x, y). The calculation formula of I′(x, y) is as follows: Among them, C is a constant used to prevent the denominator from being 0; the normalized window size is controlled by M and L, and in this embodiment, M=L=7 is preset; ω(m,l) is a two-dimensional circularly symmetric Gaussian weighting function to perform weight distribution of local mean and standard deviation. The brightness feature of the sample image is obtained by calculating the frequency distribution histogram of the absolute value of the normalized image I'(x,y), which can be expressed as an i-dimensional vector {f1,f2,…,f i } is used as the feature vector representing brightness, and the calculation process is basically the same as edge feature extraction. The formula is: Finally, the two dimensions of illumination and texture are combined into one to calculate the quality score. The mean of the feature vector is used as the quality score to reflect the high illumination and high texture of the sample image. The calculation formula is as follows: Where N is the total number of edge features of the sample image; M is the total number of brightness features of the sample image; f k is the edge feature value of the sample image; f i is the brightness feature value of the sample image.

4. The five-dimensional sample selection method based on image data quality assessment according to claim 1, characterized in that: In step 3, the Meta-AI-based SAM semi-reference image quality objective evaluation index calculation mechanism is adopted, and the specific steps are as follows: Let I be the input JPG image, X be the corresponding XML file, and B i is the coordinate of the i-th bounding box, M i corresponds to the bounding box B i The mask of A i The mask M i Area: Parse the XML file corresponding to the jpg image through the SAM model XML(X) , get the coordinates of the bounding box biunding boxes : biunding boxes =parse XML(X) Generate the input prompt box input_boxes of the mask, where create_input_box(B i ) indicates that the coordinates of multiple bounding boxes contained in a single sample are used to generate annotation prompt boxes: input_boxes={create_input_box(B i )|B i ∈bounding boxes } Segment the specific object in the annotation box and generate a pixel-level mask M i , the image is segmented into different regions and converted into a binary image, where pixels with a value of 1 represent the region belonging to the segmented object, and pixels with a value of 0 represent the background or other objects; where I is the picture input to the SAM model: {M i |M i is the mask for B i }=SAM(I,input_boxes) Calculate the sum of the mask areas of all the annotation boxes in a single image, where n is the number of annotation boxes, and A i is the mask area corresponding to the i-th annotation box: Measurement indicators for quantitative samples: IQA_SAM=total_area 5. The five-dimensional sample selection method based on image data quality assessment according to claim 1, characterized in that: In step 4, a mathematical balance mechanism is used, and the specific steps are as follows: The IQA_ImageEntropy metric and the IQA_SNR metric select the minimum-maximum normalization (min_max_normalization): The IQA_NRLT and IQA_SAM indicators use log normalization: After normalization, the gaussian_score Gaussian score is used for the two indicators whose values ​​need to be balanced, the IQA_ImageEntropy indicator and the IQA_NRLT indicator: in: For the IQA_SNR indicator, where the larger the value, the better, increasing_score is used: For the IQA_NRLT indicator, where the smaller the value, the better, the decreasing_score is used: After normalization and scoring function calculation, the four indicators are unified into the range of 0-1. The larger the indicator value, the higher the sample quality. The final quality score is obtained by allocating weights of 0.25, 0.25, 0.25, and 0.

25. According to the final quality score results, high-quality samples are divided into A, B, and C categories, and low-quality samples are divided into D category. The mathematical process to achieve the indicator balance mechanism is: IQA_ImageEntrop indicator: in, IQA_SNR indicator: IQA_NRLT indicator: IQA_SAM indicator: in, Final quality score S: S=w1·f1+w2·f2+w3·f3+w4·f4 Among them, the weight w1=w2=w3=w4=0.25.

Images