An image quality evaluation method based on interest region multi-texture feature fusion

Abstract

The application discloses an image quality evaluation method based on a multi-texture feature fusion of an interest region and belongs to the field of computer vision. The application focuses on the texture information of local pollen grains as the quality score of the whole image. The process module of the model comprises an image preprocessing module, a local pollen grain extraction module, a multi-texture feature quantization index extraction module and a quality evaluation final score module. The results show that the process and solution proposed by the application focus on the effect of the local pollen grain interest region and can better perceive the quality of pollen data.

Description

Technical Field

[0001] This invention relates to an image quality evaluation method based on multi-texture feature fusion of regions of interest, belonging to the field of computer vision. Background Technology

[0002] Allergic diseases are a major public health problem, and pollen is one of the triggers for allergies. Experts often use manual counting methods to detect the concentration of pollen in the air to ensure the normal lives of pollen allergy sufferers. With the development of image digitization equipment, automated recognition methods have been introduced. The most important step in automated recognition is to generate digital data using instruments. However, due to different focal planes, samples may appear blurry, resulting in inconsistent pollen data quality. Data quality significantly affects the performance of classification networks, and manually selecting high-quality data is time-consuming and laborious. Therefore, there is an urgent need for an automated image quality assessment method to retain high-quality images. This paper proposes an image quality assessment method based on the fusion of multi-texture features in regions of interest to screen for high-quality data.

[0003] Cheng Deqiang, Shao Lirong [1] A method for evaluating the quality of reference-free blurred images based on saliency was proposed by Cheng Shan and Zeng Huanqiang. [2] Others proposed a feature fusion-based method for evaluating image quality in a no-reference screen. In summary, most existing image quality assessment methods are designed for general natural images and cannot assess the quality of specific, highly relevant local objects within a global image. References:

[0004] [1] Cheng Deqiang, Shao Lirong, Liu Weilong, Zhang Jianying, You Dalei. Quality assessment of reference-free blurred images based on saliency[J]. Television Technology, 2018, 42(07):1-4.DOI:10.16280 / j.videoe.2018.07.001.

[0005] [2] Cheng Shan, Zeng Huanqiang, Chen Jing, Tian Yu, Cai Canhui. Image quality assessment without reference screen based on feature fusion [J]. Signal Processing, 2019, 35(03):419-425. DOI:10.16798 / j.issn.1003-0530.2019.03.013.

[0006] As attached Figure 1 As shown, the pollen data has the following characteristics:

[0007] 1) The pollen image is very large. The area of ​​the pollen grains is very small, and they only appear in local areas.

[0008] 2) The quality information and texture information of pollen images are inseparable. The clearer the texture, the better the quality. Therefore, texture information is crucial for pollen grain quality analysis.

[0009] Texture information is a key focus, and it is also an organized regional phenomenon. It is related to brightness variations in an image, can be perceived by the human eye, and texture features have a good ability to characterize image quality levels. Given the characteristics of pollen images, people often focus more on the features of the regions of interest, that is, the image quality of local pollen grain areas. Therefore, multiple texture methods are fused to perform texture feature analysis on pollen grain areas to perceive the quality of pollen images, thereby improving the accuracy of pollen image quality assessment. Summary of the Invention

[0010] Most existing image quality assessment methods are only applicable to general natural images. However, pollen images are large in size, with large areas of background such as impurities and bubbles that are not of interest. Furthermore, the focus should be on the regions of interest (ROIs) of pollen grains within the image. The quality information of pollen grain images is inseparable from their texture information; the clearer the texture, the better the quality. This invention addresses this problem by designing an image quality assessment method based on multi-texture feature fusion of ROIs. Specifically, the predicted quality value of multiple pollen images is determined by a combination of multi-texture features from one or more local ROIs of pollen grains in each image. Multi-texture feature fusion avoids the limitations and randomness of single-texture features. Results show that the proposed process and solution, by focusing on the local ROIs of pollen grains, can better perceive the quality of pollen data.

[0011] This invention employs the following technical solution: it focuses on the texture information of local pollen grains as the quality score of the entire image. The model's workflow includes an image preprocessing module, a local pollen grain extraction module, a multi-texture feature quantification index extraction module, and a final quality evaluation scoring module.

[0012] The specific solution of the present invention is attached. Figure 2 As shown.

[0013] In the image preprocessing module, large-scale images are cropped and filtered to obtain multi-layered optimal-size pollen images. Then, image registration is performed to achieve spatial alignment of pollen grains in the multi-layered optimal-size pollen images.

[0014] In the local pollen grain extraction module, all regions of interest (ROIs) of local pollen grains in the optimal-sized pollen image are extracted. First, the image is denoised to achieve a smoothing effect. Next, obvious edges are detected through edge detection, and then circular contours are detected to identify near-circular targets. Then, color recognition is used to exclude near-circular bubbles and impurities. Finally, the pollen grain image is collected by calculating the minimum rectangular contour of the pollen grains.

[0015] In the multi-texture feature quantification index extraction module, various texture methods are used to extract texture feature indices from pollen grain images. First, image grayscale conversion is performed to prepare for texture extraction. Then, different texture features from various texture methods are used for texture extraction. Finally, the texture feature indices are normalized to eliminate the adverse effects of singular data.

[0016] In the final scoring module of the quality assessment, the normalized texture index is weighted and fused to serve as the quality score for a single pollen grain. Then, the quality scores of all pollen grains in the pollen image are weighted and integrated to output as the quality score for the entire pollen image.

[0017] Compared with existing technologies, the advantages of this invention are as follows:

[0018] I. The image quality assessment method based on multi-texture feature fusion of regions of interest adopted in this invention perceives the overall quality information through the texture details of local objects, making the quality assessment model more prominent in the regions of interest.

[0019] Second, the image quality assessment method based on multi-texture feature fusion of regions of interest adopted in this invention uses multiple texture extraction methods to select texture feature indicators with better performance in order to perceive quality information, thereby making the quality assessment model more accurate in prediction.

[0020] Third, the image quality evaluation method based on multi-texture feature fusion of regions of interest adopted in this invention uses image processing methods to extract all target information in the image. After collecting texture information for each local target, a normalization and weighted average strategy is used to perceive the quality of a single pollen grain image. The weighted average strategy is used again to perceive the quality of the entire pollen image, making the model's prediction value more accurate. Attached Figure Description

[0021] Figure 1 Example image of pollen at optimal size.

[0022] Figure 2 This is a model diagram of the method proposed in this invention. Detailed Implementation

[0023] The following detailed description of embodiments of the present invention is provided in conjunction with the accompanying drawings:

[0024] This invention is an image quality assessment method based on multi-texture feature fusion of regions of interest. This method uses weighted fusion of texture features from local pollen grain regions to perceive the quality score of the entire image. The specific process of image quality assessment is as follows: 1. Image preprocessing: Image registration is performed on the spatial locations of multiple pollen images. 2. Local pollen grain extraction: First, edge detection is performed on the edge features of the pollen image. Second, the circular or elliptical contour features of the pollen grains are detected and the circular contours are extracted. Finally, bubbles and impurities with similar contour features to the pollen grains are excluded based on color features. Morphological operations and other image processing methods are used to extract all pollen grain images in the pollen image. 3. Multi-texture feature quantification index extraction: First, the index with better performance is selected from the texture feature indices extracted by various texture methods. Then, for each pollen grain image in the pollen image, texture features are extracted using the selected texture index. Finally, normalization is performed on each extracted texture index. 4. Calculation of the final quality assessment score: First, a weighted strategy is used to calculate the total texture feature score of each pollen grain image, which is used as the quality score of a single pollen grain image. Then, the final quality score of the entire pollen image is obtained by weighting all the pollen grain image scores equally.

[0025] Specifically, the method includes the following steps:

[0026] Step 1: Image Preprocessing

[0027] Large-scale images are cropped and filtered to obtain multi-layered optimal-size pollen images. Then, image registration is performed to finally obtain multi-layered optimal-size pollen images with pollen grains aligned in spatial location.

[0028] Step 2: Localized pollen grain extraction

[0029] This step extracts regions of interest (ROIs) for all local pollen grains in the optimal-sized pollen image. The input for this step is the optimal-sized pollen image, and the output is an image of all pollen grains in the optimal-sized pollen image.

[0030] Step 2.1 Image Denoising: First, smooth the pollen grain image to reduce noise points. Then, apply an 11×11 Gaussian blur to remove background noise from the pollen image.

[0031] Step 2.2 Edge Detection: First, calculate the image gradient magnitude and direction to obtain a set of possible edges. Next, perform non-maximum suppression on the gradient magnitude, setting the grayscale values ​​corresponding to non-maximum values ​​to 0, while retaining edges with significant grayscale changes. Finally, use a double thresholding process to remove false positives. Specifically, pixels with grayscale values ​​less than 30 are considered false edges and set to 0, while pixels with grayscale values ​​greater than 100 are considered strong edges and set to 1. Pixels in between require further inspection. This process preserves fine-line edges and excludes non-edge pixels.

[0032] Step 2.3 Detecting Circular Contours: Based on the edge detection in the previous step, the Hough transform is used to adjust the threshold and detect circular contours in the pollen image, ultimately obtaining circular pollen grains, bubbles, and impurities. Following expert guidance, it was found that in optimally sized pollen images, the diameter of pollen grains is typically greater than 50 pixels. Therefore, circular contour information with a diameter less than 50 pixels is removed, further reducing bubble and impurity information.

[0033] Step 2.4 Eliminating Bubbles and Impurities: In order to eliminate bubbles and impurities in the circular outlines of the image, based on the characteristic that pollen grains turn pink after staining, the RGB color space is converted to the HSV color space, and a threshold range of (0,0,100) to (180,30,255) is set (different threshold ranges exist for images with different color characteristics; other colors can be adjusted by consulting the HSV color table) to filter out pollen grains.

[0034] Step 2.5 Collect pollen grains: First, calculate the smallest rectangle containing the pollen outline, then collect the smallest rectangles in the pollen image, finally obtaining all pollen grain images. Let L_sum be the total number of all local pollen grain images in a single pollen image.

[0035] Step 3: Extraction of multi-texture feature quantization indicators

[0036] This module employs various texturing methods to extract texture feature indices. The input to this module is the pollen grain image obtained in the previous stage, and the output is the normalized texture feature indices extracted from the pollen grain image.

[0037] Step 3.1 Image grayscale conversion: First, the pollen grain image is converted to grayscale to prepare for the next step of texture feature extraction.

[0038] Step 3.2 Multi-texture feature extraction: Since different texture extraction methods extract different texture features and achieve different results, indicators that better highlight the texture effect were selected as texture feature extraction indicators. These indicators include roughness and contrast in Tamura texture features, high grayscale running emphasis, short-term high grayscale emphasis, and long-term low grayscale emphasis in GLRLM, and contrast, dissimilarity, and homogeneity in second-order statistics based on the gray-level co-occurrence matrix (GLCM).

[0039] Positively correlated metrics include: contrast in Tamura texture features, high grayscale running emphasis and short-term high grayscale emphasis in GLRLM, and contrast and dissimilarity in second-order statistics based on the gray-level co-occurrence matrix (GLCM).

[0040] Negatively correlated metrics include: roughness in Tamura texture features, long-term low gray-level emphasis in GLRLM, and homogeneity in second-order statistics based on the gray-level co-occurrence matrix (GLCM).

[0041] Furthermore, the average values ​​of the indicators selected in GLRLM and GLCM were taken from different directions. Because the texture information extracted from different directions is different and reflects different situations, four different directions were selected to extract texture information, comprehensively reflecting the texture feature information of pollen grains.

[0042] In Tamura texture features, roughness describes the degree of coarseness in a pollen grain image. The fewer the number of repetitions of a primitive, the greater the roughness. The specific calculation method is as follows:

[0043]

[0044] In equation (1), A k (x, y) represents a value of 2. k x2 k The value of g(i,j) represents the average brightness of each pixel within the rectangular window, (x,y) represents the position of the selected window in the entire image, g(i,j) represents the brightness value of the pixel at point (i,j) within the selected window, and the value of k determines the pixel range.

[0045] E k,h =|A k (x+2 k-1 ,y)-A k (x-2 k-1 ,y)| (2)

[0046] E k,v =|A k (x,y+2 k-1 )-A k (x,y-2 k-1 (3)

[0047] S best (x,y)=2 k (4)

[0048] E k =E max =max(E1,E2,…,E h (5)

[0049]

[0050] In equation (2), E k,h E represents the difference in the horizontal direction corresponding to the pixel. In equation (3), E k,v This represents the difference in the vertical direction corresponding to the pixel. In equation (4), S best This represents the optimal size that maximizes the E value. k Let k represent the value that maximizes E in any direction, as shown in equation (5). In equation (6), the roughness F of the pollen image... crs By calculating S in the entire image best The average value is obtained. Where m and n are the length and width of the image.

[0051] In Tamura texture features, contrast describes the brightness difference between light and dark areas in a pollen grain image. The greater the range of pixel difference, the greater the contrast. The specific calculation method is as follows:

[0052]

[0053] In equation (7), σ represents the standard deviation of the image grayscale values, and α4 represents the kurtosis of the image grayscale values. μ4 is the mean of the fourth moment, σ 2 denoted as the variance of the image grayscale values.

[0054] High Gray Scale Run Emphasis (HGRE) in GLRLM describes the distribution of grayscale values ​​in a pollen grain image. The more high grayscale values ​​present, the larger the value. The specific formula is as follows:

[0055]

[0056] In equation (8), p(i,j|θ) represents the gray-level run-length matrix, where θ is 1, 2, 3, or 4, representing 0°, 45°, 90°, and 135° respectively. L represents the number of gray-level run-length matrices in different directions. The point (i,j|θ) represents the count of a pixel with gray level i appearing j times consecutively in the θ direction of the image, and N... g N represents the number of gray levels in an image. r This indicates the number of different runs on the graph.

[0057] Short-Term High Gray-Level Emphasis (SRHGE) in GLRLM describes the characteristics of gray levels and run length in pollen grain images. A shorter run length results in a greater distribution of high gray-level values ​​and a larger SRHGE value. The specific formula is as follows:

[0058]

[0059] In equation (9), p(i,j|θ) represents the gray-level run-length matrix, where θ is 1, 2, 3, or 4, representing 0°, 45°, 90°, and 135° respectively. L represents the number of gray-level run-length matrices in different directions. The point (i,j|θ) represents the count of a pixel with gray level i appearing j times consecutively in the θ direction of the image, and N... g N represents the number of gray levels in an image. r This indicates the number of different runs on the graph.

[0060] Long-Term Low Gray-Level Emphasis (LRLGE) in GLRLM describes the characteristics of gray levels and run length in pollen grain images. A longer run length indicates a greater distribution of low gray-level values, resulting in a larger SRHGE value. The specific formula is as follows:

[0061]

[0062] In equation (10), p(i,j|θ) represents the gray-level run-length matrix, where θ is 1, 2, 3, or 4, representing 0°, 45°, 90°, and 135° respectively. L represents the number of gray-level run-length matrices in different directions. The point (i,j|θ) represents the count of a pixel with gray level i appearing j times consecutively in the θ direction of the image, and N... g N represents the number of gray levels in an image. r This indicates the number of different runs on the graph.

[0063] In the second-order statistics of the Gray-Level Co-occurrence Matrix (GLCM), contrast reflects the sharpness of an image and the depth of its texture. The greater the contrast, the deeper the texture, and the sharper the image.

[0064]

[0065] In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively. H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix.

[0066] Dissimilarity in the second-order statistics of the gray-level co-occurrence matrix (GLCM): Experiments show that the clearer the texture of the pollen grain image, the greater the dissimilarity value. The specific formula is as follows:

[0067]

[0068] In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively. H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix.

[0069] Homogeneity in the second-order statistics of the Gray-Level Co-occurrence Matrix (GLCM): Experiments show that the clearer the texture of pollen grain images, the smaller the homogeneity value. The specific formula is as follows:

[0070]

[0071] In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively. H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix.

[0072] Step 3.3 Normalize Texture Metrics: To eliminate the adverse effects of outlier data, the texture metrics extracted in the previous step are normalized using the sigmoid function, controlling the values ​​of each metric within the range of 0-1. The specific formula for the sigmoid function is shown below:

[0073]

[0074] In equation (14), x represents the value before normalization, and f(x) represents the value after normalization.

[0075] Step 4: Final Quality Evaluation Score

[0076] The texture feature indices of pollen grains obtained in the previous step are weighted and integrated to form the quality score of a single pollen grain. Then, the quality scores of all pollen grains in the pollen image are weighted and summed to form the quality score of the entire pollen image.

[0077] Step 4.1 Calculate the quality score of a single pollen grain image: To perceive the quality information of pollen grains through their texture information, texture indices selected from different texture methods are merged according to the principle of weighted average distribution. Positive correlation indices are assigned positive values, and negative correlation indices are assigned negative values. The merged texture quantization index is used as the quality score of a single pollen grain image.

[0078]

[0079] In equation (15), s_iqa l S_num represents the quality score of the l-th local pollen grain image in the entire image, and S_num represents the total number of texture indices. con HGRE, SRHGLE, CON, and DIS represent positively correlated indicators in texture metrics. crs LRLGLE and Homo represent negative correlation metrics in texture metrics. f(F con ), f(F crs f(HGRE), f(SRHGLE), f(LRLGLE), f(CON), f(DIS), and f(Homo) represent the normalized texture index quantization values.

[0080] Step 4.2 Calculate the total quality score: In order to obtain the quality score of the entire pollen image, the quality scores of each local pollen grain image in the pollen image are weighted and averaged, and finally used as the quality score of the pollen image.

[0081]

[0082] In equation (16), iqa represents the quality score of the entire image, and s_iqa l L_sum represents the quality score of the l-th local pollen grain image in the whole image, and L_sum represents the number of single pollen grain images in the whole image.

Claims

1. An image quality evaluation method based on multi-texture feature fusion of interest region, characterized in that The method comprises the following steps: Step 1: image preprocessing The large-scale image is cropped and screened to obtain multi-layer optimal size pollen images; then image registration is performed to finally obtain aligned multi-layer optimal size pollen images of pollen particles in spatial position; Step 2: local pollen particle extraction All local pollen particle regions of interest in the optimal size pollen image are extracted, the input of this step is the optimal size pollen image, and the output result is all pollen particle images in the optimal size pollen image; Step 2.1 image denoising: first, the pollen particle image is smoothed to reduce noise points in the image; then the background noise in the pollen image is removed; Step 2.2 edge detection: first, the image gradient amplitude and direction are calculated to obtain a possible edge set; second, non-maximum suppression is performed on the gradient amplitude, and false positives are removed by using a double-threshold screening operation; Step 2.3 detection of circular contour: on the basis of the edge detection in the previous step, the Hough transform is used to adjust the threshold to detect the circular contour in the pollen image, and finally the circular pollen particles, bubbles and impurities are obtained; Step 2.4 exclude bubbles and impurities: convert the RGB color space to the HSV color space, set the threshold range of (0, 0, 100) to (180, 30, 255), and screen out the pollen particles; Step 2.5 collect pollen particles: first, calculate the minimum rectangle containing the pollen contour, then collect the minimum rectangle in the pollen image, and finally obtain all pollen particle images; the total number of all local pollen particle images in a pollen image is set as L_sum; Step 3: multi-texture feature quantitative index extraction A plurality of different texture methods are used to extract texture feature indexes; the input of this module is the pollen particle image obtained in the previous stage, and the output is the normalized texture feature indexes extracted from the pollen particle image; Step 3.1 image grayscale conversion: first, the pollen particle image is grayed to prepare for the next step of texture feature extraction; Step 3.2 multi-texture feature extraction: different texture extraction methods extract different texture features, and the effects are also different; the indexes that can highlight the texture effect are selected as the texture feature extraction indexes; these indexes include the roughness and contrast in the Tamura texture feature, the high gray level run emphasis, short-term high gray level emphasis and long-term low gray level emphasis in GLRLM, and the contrast, dissimilarity and homogeneity in the second-order statistics based on the gray level co-occurrence matrix GLCM; The positive correlation indexes include: the contrast in the Tamura texture feature, the high gray level run emphasis and short-term high gray level emphasis in GLRLM, and the contrast and dissimilarity in the second-order statistics based on the gray level co-occurrence matrix GLCM; The negative correlation indexes include: the roughness in the Tamura texture feature, the long-term low gray level emphasis in GLRLM, and the homogeneity in the second-order statistics based on the gray level co-occurrence matrix (GLCM); In addition, the mean value of the selected indicators in GLRLM and GLCM is taken in different directions; because the texture information content extracted in different directions is not the same, the reflection is also different, so four different directions are selected to extract texture information, and the texture characteristic information of pollen grains is comprehensively reflected; The roughness in Tamura texture feature describes the roughness of pollen grain image; the fewer the primitive repetition times, the greater the roughness; the specific calculation method is as follows: In formula (1), A k (x,y) represents the average brightness of each pixel point in a rectangular window of size 2 k x2 k , (x,y) represents the position of the selected window in the entire image, g(i,j) represents the pixel brightness value of the (i,j) point in the selected window, and k determines the pixel range; E k,h = |A k (x+2 k-1 ,y)-A k (x-2 k-1 ,y)| (2) E k,v = |A k (x,y+2 k-1 )-A k (x,y-2 k-1 )| (3) S best (x,y) = 2 k (4) AND k =E max = max(E1,E2,…,E h ) (5) In equation (2), E k,h E represents the difference in the horizontal direction corresponding to the pixel; in equation (3), E k,v S represents the difference in the vertical direction corresponding to the pixel; in equation (4), S best This represents the optimal size that maximizes the E value; E k k represents the value of E in any direction, as shown in equation (5); in equation (6), the roughness F of the pollen image crs By calculating S in the entire image best The average value is obtained; where m and n are the length and width of the image; The contrast in Tamura texture feature describes the brightness difference of light and dark areas in pollen grain image; the greater the difference range of pixels, the greater the contrast; the specific calculation method is as follows: In formula (7), σ represents a standard deviation of the image gray scale value, and a4 indicates a kurtosis of the image gray scale value, wherein μ4 is a fourth moment mean, σ 2 is a variance of the image gray scale value; The high gray run length emphasis HGRE in GLRLM emphasizes the distribution of gray value of pollen grain image; the more the high gray value distribution, the greater the value; the specific formula is as follows: In formula (8), p(i, j|θ) represents a gray run matrix, and θ is 1, 2, 3 or 4, representing 0°, 45°, 90° and 135° respectively; L represents the number of gray run matrices in different directions; (i, j|θ) represents the count of pixels with gray i appearing continuously j times in the image in the direction θ; N g represents the number of different runs in the image; and r represents the number of different runs in the image. The short-term high gray emphasis SRHGE in GLRLM describes the features of gray level and run length in pollen grain image; the shorter the run length, the more the high gray value distribution, the greater the SRHGE value; the specific formula is as follows: In formula (9), p(i, j|θ) represents a gray run matrix, and θ is 1, 2, 3, or 4, representing 0°, 45°, 90°, and 135°, respectively; L represents the number of gray run matrices in different directions; (i, j|θ) represents the count of pixels with a gray value of i that appear continuously j times in the image in the θ direction; and N g represents the number of gray levels on the image r represents the number of different runs on the image; The long-term low gray emphasis LRLGE in GLRLM describes the features of gray level and run length in pollen grain image; the longer the run length, the more the low gray value distribution, the greater the LRLGE value; the specific formula is as follows: In formula (10), p(i,j|θ) represents a gray run matrix, θ here takes 1, 2, 3, 4, which respectively represent 0°, 45°, 90° and 135°; L represents the number of gray run matrices in different directions; (i,j|θ) represents the count of pixels with gray i appearing continuously j times in the image θ direction, N g represents the number of different runs on the image, N r represents the number of different runs on the image; The contrast in the second-order statistics of gray level co-occurrence matrix reflects the definition and texture depth of the image; the greater the contrast, the deeper the texture, and the clearer the image presented; In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively; H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix. The dissimilarity in the second-order statistics of gray level co-occurrence matrix, the clearer the texture of pollen grain image, the greater the dissimilarity value, and the specific formula is as follows: In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively; H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix. The homogeneity in the second-order statistics of gray level co-occurrence matrix, the clearer the texture of pollen grain image, the smaller the homogeneity value; the specific formula is as follows: In the formula, P(g1,g2) h This represents the grayscale value of each pair of points, where h represents different directions, which are 1, 2, 3, and 4, representing 0°, 45°, 90°, and 135° respectively; H represents the number of co-occurrence matrices extracted for different directions, and (g1, g2) represents the position coordinates of the grayscale co-occurrence matrix. Step 3.3 Normalization of texture indicators: in order to eliminate the adverse effects caused by singular sample data, the texture indicators extracted in the last step are normalized by sigmoid function, and the values of each texture indicator are controlled between 0 and 1; the specific formula of sigmoid function is as follows: In formula (14), x represents the value before normalization, and f(x) represents the value after normalization; Step 4: final score of quality evaluation The texture characteristic indicators of pollen grains obtained in the last step are integrated by equal weight, which are taken as the quality score of single pollen grain; then, the quality scores of all pollen grains in the pollen image are added by equal weight, which are taken as the quality score of the whole pollen image; Step 4.1 calculation of quality score of single pollen grain image: in order to perceive the quality information of pollen grain through its texture information, the selected texture indicators from different texture methods are combined according to the principle of equal weight; among them, the positive correlation indicators take positive values, and the negative correlation indicators take negative values; the combined texture quantitative indicators are taken as the quality score of single pollen grain image; In formula (15), s_iqa l S_num represents the total number of texture indexes; F con , HGRE, SRHGLE, CON, DIS represent positive correlation indexes in the texture indexes; F crs , LRLGLE, Homo represent negative correlation indexes in the texture indexes; f(F con ), f(F crs ), f(HGRE), f(SRHGLE), f(LRLGLE), f(CON), f(DIS), f(Homo) represent quantized values of the normalized texture indexes; Step 4.2 calculation of total score: in order to obtain the quality score of the whole pollen image, the quality scores of each local pollen grain image in the pollen image are weighted and averaged, and finally the quality score of the pollen image is obtained. In Equation (16), iqa represents the quality score of the whole image, s_iqa l represents the quality score of the l-th local pollen grain image in the whole image, and L_sum represents the number of single pollen grain images in the whole image.

Images