Comprehensive evaluation method for underwater image quality assessment performance based on dataset classification

Abstract

The invention relates to the field of image processing technology, and discloses a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification, S1: the data set is classified according to the standards of the distortion level image sequence and the applied subjective evaluation mode; S2: the four types of data sets obtained after classification are analyzed, and the main indicators and secondary indicators for evaluating the underwater image quality evaluation method in each data set are extracted; S3: the comprehensive performance of each type is calculated according to the value range of the main indicators and secondary indicators and the analysis of the effect on the data set; S4: a quadrilateral radar map is constructed for the comprehensive performance of each type; S5: the area and balance of the closed quadrilateral in the quadrilateral radar map are calculated to obtain the comprehensive performance of the underwater image quality evaluation method. Experiments show that the underwater image quality evaluation method can be explained intuitively and comprehensively using QDE, which provides a reference for improving the performance of the underwater image quality evaluation method in the future.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification. Background Art

[0002] As a scientific exploration tool, underwater vision is becoming increasingly important in diverse fields such as marine research, archaeology, and biology. Image Quality Assessment (IQA) evaluates the quality of images and helps determine whether the images meet the expected visual standards. In recent years, natural IQA methods have developed rapidly, and many algorithms have been proposed, such as machine learning-based methods BRISQUE, NIQE, DILT, ILNIQE, SNPNIQE, and deep learning-based methods DLIQA, RankIQA, DB-CNN, HyperIQA, UNIQUE, MB-CNN, etc. However, the quality measurement of underwater images is challenged by light absorption and scattering, uneven illumination, and noise. Directly applying existing natural image quality assessment methods to underwater images cannot accurately reflect the quality of perceived images. Underwater Image Quality Evaluation (UIQE) methods can be divided into subjective and objective methods. Subjective UIQE methods provide a quality benchmark, but are usually labor-intensive and time-consuming. In contrast, objective UIQE offers advantages such as repeatability and automation, making it very attractive to researchers. Early studies on objective UIQE mainly focused on the quality assessment of grayscale images. For underwater color images, evaluation indicators are usually formulated according to the particularity of the underwater environment to achieve quality assessment of underwater images.

[0003] Humans are the ultimate recipients of images, and the performance of IQA methods is usually verified by related metrics, such as Pearson Linear Correlation Coefficient (PLCC), Spearman Rank Order Correlation Coefficient (SROCC), Kendall Rank Order Correlation Coefficient (KROCC), and Root Mean Square Error (RMSE). PLCC measures the linear correlation between the algorithm's prediction value and the Mean Opinion Score (MOS), while SROCC evaluates the monotonic relationship between the two and KROCC quantifies their ordinal relationship. RMSE measures the difference between the IQA value and MOS, and values ​​closer to 0 indicate better performance of the IQA algorithm. In addition, pristine / distorted image discriminability test (D-test), listwise ranking consistency test (L-test), and pairwise preference consistency test (P-test) are also used. Recently, researchers have proposed the maximum difference (MAximumDifferentiation, MAD) and group maximum difference (group MAximumDifferentiation, gMAD) methods to compare the perceptual distinguishability performance between two or more IQA methods. MAD works by using another model to refute the comparison model, and the model with more challenging refutation is considered better; gMAD quantifies the attack performance by searching for the maximum difference predicted by other models in the same quality range by the attacker, and quantifies the defense performance by the degree to which the defender is immune to the attack.

[0004] Since IQA usually relies on image quality assessment datasets, the comprehensive performance of IQA algorithms also needs to consider factors such as the richness of dataset categories and the diversity of evaluation perspectives. At present, the following problems exist in the field of underwater image quality assessment: UIQE methods are usually developed based on certain specific datasets, focusing on a single aspect of performance, such as the consistency of MOS values ​​with the predicted values ​​of image sequences with different distortion levels, the consistency of ranking with the predicted values ​​of enhanced images, or the correlation with the predicted values ​​of cross-content images. However, reliance on specific datasets may cause the model to perform poorly in other datasets or scenarios; this limitation may limit the applicability and effectiveness of UIQE methods in practical applications. For this reason, a comprehensive evaluation method for underwater image quality assessment performance based on dataset classification is provided. Summary of the invention

[0005] The purpose of the present invention is to provide a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification in order to address the defects of the prior art and solve the problems raised by the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solution: a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification, the specific steps are as follows:

[0007] S1: Datasets The underwater image quality assessment datasets are classified according to the criteria of the distortion level image sequences and the applied subjective evaluation mode;

[0008] S2: Analysis of the four types of datasets obtained after classification, extracting the main and secondary indicators for evaluating underwater image quality assessment methods in each dataset;

[0009] S3: Calculate the comprehensive performance of each type based on the value range of the primary and secondary indicators and the analysis of the effect on the data set;

[0010] S4: Construct a quadrilateral radar chart based on the comprehensive performance of each type;

[0011] S5: Calculate the area and balance of the closed quadrilaterals in the quadrilateral radar image to obtain the comprehensive performance of the underwater image quality evaluation method.

[0012] As a preferred technical solution of the present invention, in step S1, the underwater image quality evaluation dataset is classified into a distortion level sequence dataset and a non-distortion level sequence dataset according to whether the image contains a distortion level sequence. The distortion level sequence dataset is divided into a distortion level sequence dataset for changing the imaging distance and a distortion level sequence dataset for changing the water turbidity according to the transformation conditions. At the same time, according to whether the images are grouped with the same content, the non-distortion level sequence dataset is subdivided into distortion-free level sequence datasets in grouping mode and non-grouping mode.

[0013] As a preferred technical solution of the present invention, the main index b of the distortion level sequence data set of changing the imaging distance and changing the water quality turbidity and the distortion level sequence data set of the grouping mode i1 For L-test, secondary indicator b i2 is D-test, the main indicator of the undistorted level sequence data set in the non-grouping mode b i1 For PLCC, secondary indicator b i2 For P-test.

[0014] As a preferred technical solution of the present invention, the main indicator b in step S3 i1 For the first aspect, the main indicator b i1 Absolute values ​​and secondary indicatorsb i2 Multiply as the second aspect, calculate the comprehensive index a for each type of data set i :a i (i=1,2,3,4,5)=w1×b i1 +w2×|b i1 |×b i2 ; Where w1 and w2 are weighting coefficients.

[0015] As a preferred technical solution of the present invention, the area S of the closed quadrilateral in the quadrilateral radar map in step S5 is n for The balance is in a5=a1.

[0016] Compared with the existing technologies, the beneficial effects of the present invention are as follows: a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification is proposed, in which the data set is divided according to whether it is a distortion level sequence and whether it is a group evaluation mode, and the four types of data sets are selected according to their characteristics. The main and secondary indicators are used to form a quadrilateral radar chart to calculate the area and balance, and the weight distribution method is applied to variance analysis to evaluate its rationality. In addition, QDE is applied to the advanced UIQE method. From the observation data, the normalized area of ​​the UEIQM algorithm has an advantage in overall performance compared with other algorithms. In terms of balance, the significant value of UEIQM indicates that its performance in various aspects is relatively balanced; according to the radar chart analysis, the UEIQM algorithm is always at the forefront in the performance evaluation of the four types of data sets, which further confirms the ability to use QDE to comprehensively evaluate underwater image quality evaluation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is the overall flow chart of the present invention;

[0018] Figure 2 It is a schematic diagram of a QDE radar diagram of the present invention;

[0019] Figure 3 It is the QDE radar chart of the UIQE method of the present invention. DETAILED DESCRIPTION

[0020] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby making a clearer and more definite definition of the protection scope of the present invention.

[0021] Embodiment: The present invention provides a technical solution: a comprehensive evaluation method for underwater image quality evaluation performance based on data set classification, which classifies underwater image quality evaluation data sets according to whether they contain distortion levels and the applied subjective evaluation mode, and obtains distortion level sequence data sets with changing imaging distance and changing water turbidity, as well as distortion-free level sequence data sets in grouping mode and non-grouping mode;

[0022] Analyze each type of data set and extract the main and secondary indicators of the underwater image quality evaluation method of the data set; calculate the comprehensive performance of each type based on the value range of the main and secondary indicators and the analysis of their effects on the data set;

[0023] The comprehensive performance of each type is used to construct a quadrilateral radar map; the area and balance of the closed quadrilaterals in the quadrilateral radar map are calculated to obtain the comprehensive performance of the underwater image quality evaluation method;

[0024] The overall flow chart of the evaluation method is as follows: Figure 1 As shown;

[0025] Primary and secondary indicator evaluation process:

[0026] For each underwater image quality assessment dataset, it is randomly divided into 80% as a training set and 20% as a test set. The training / testing split process is repeated 1000 times, and the results after each split are calculated, and the average performance of each dataset is finally reported.

[0027] Performance on different distortion level sequence datasets: In order to compare the consistency between different UIQE algorithms in predicting the degree of underwater absorption and scattering and the subjective MOS value, the performance comparison was performed on the UID-LEIA dataset using L-test and D-test. L-test is used to evaluate the sensitivity of the algorithm to the underwater distortion level, and D-test evaluates the ability of the algorithm to distinguish between the original image and the distorted image. The UID-LEIA dataset is a distortion-free level sequence dataset based on changes in imaging distance and water turbidity, and the performance when changing the imaging distance and water turbidity is calculated respectively.

[0028] Performance on group evaluation mode datasets: In order to compare the consistency between different UIQE algorithms in predicting different transformed images of the same content and subjective MOS values, the performance of UIQE algorithms on SAUD and UID2021 datasets was evaluated using L-test and D-test. For such datasets, there is no correlation between the MOS values ​​of image groups with different contents. Therefore, L-test is performed on each group to evaluate the consistency of the algorithm in ranking the enhancement results.

[0029] Performance on datasets without group evaluation mode: In order to compare the correlation between different UIQE algorithms and subjective MOS values ​​in predicting cross-content images, PLCC and P-test were used for evaluation on UWIQA and UIED datasets.

[0030] Comprehensive indicator calculation:

[0031] Use quad radar charts to visualize the performance of various UIQE metrics, such as Figure 2 As shown. In the radar chart, a1 and a2 represent the comprehensive performance of the algorithm on the distorted level sequence data set at different imaging distances and water turbidity, respectively. A3 and a4 reflect the overall performance of the algorithm on the undistorted level sequence data set in the grouped evaluation mode and the ungrouped evaluation mode, respectively. These four overall performances form a closed quadrilateral. The larger the area of ​​the closed quadrilateral and the more uniform the distribution, the better the overall performance of the algorithm. When all four vertices are on the circumference of the circle, it means that the performance has reached the optimal level. Therefore, the proposed QDE involves two factors:

[0032]

[0033]

[0034]

[0035]

[0036] The normalized area S n is the normalized area of ​​the closed area surrounded by adjacent vertices in the quadrilateral radar chart. The balance parameter σ is introduced to describe the balance of various indicators, and μ is the average value of the four indicators. n This indicates that the algorithm has good accuracy in image quality assessment for images with different distortion levels, different underwater enhanced images, and across dataset content.

[0037] Considering that PLCC and L-test range from -1 to 1, the origin of the QDE graph is set to -1. i(i=1,2,3,4) represent the performance of four types of data sets respectively, and a5=a1. i Includes two sub-performance b i1 and b i2 , expressed as:

[0038] a i =w1×b i1 +w2×|b i1 |×b i2 (5)

[0039] Among them, w1 and w2 are weighting coefficients.

[0040] Experiment and analysis:

[0041] Experiments were conducted on the underwater image quality test set, and seven advanced UIQE methods, including UIQM, UCIQE, CCF, FDUM, UEIQM, NUIQ and TwiceMixing, were selected for comparison on four types of data sets. The main and secondary indicators of this scheme were calculated, and the four overall indicators and QDE comprehensive indicators obtained are shown in Tables 1 and 2. The radar chart is shown in Figure 3 shown.

[0042] Table 1: Four indicators of QDE radar chart

[0043] method <![CDATA[a1]]> <![CDATA[a2]]> <![CDATA[a3]]> <![CDATA[a4]]> UIQM <![CDATA[ 0.9747 ]]> 0.4763 0.6008 0.3079 UCIQE 0.9559 0.5207 0.5976 <![CDATA[ 0.3472 ]]> CCF 0.9661 0.4592 <![CDATA[ 0.6216 ]]> 0.2885 FDUM 0.9745 0.6198 0.6179 0.2535 UEIQM 0.9796 <![CDATA[ 0.5284 ]]> 0.8024 0.3570 NUIQ 0.1498 0.4068 0.6136 0.2567 TwiceMixing 0.7191 -0.0106 0.5344 0.2524

[0044] Table 2: Results of the QDE method

[0045] Comprehensive indicators UIQM UCIQE CCF FDUM UEIQM NUIQ TwiceMixing <![CDATA[S n ]]> 0.6222 0.6369 0.6161 0.6451 0.6820 0.4600 0.4559 σ 0.2453 0.2218 0.2502 0.2549 0.2405 0.1742 0.2773

[0046] From the observation data, the normalized area S of the UEIQM algorithm n The value of UEIQM is 0.6820, indicating that it has an advantage in overall performance compared with other algorithms. In terms of balance, the significant value of UEIQM shows that its performance in all aspects is relatively balanced, ranking only lower than NUIQ and UCIQE. Figure 3 According to the analysis in Tables 1 and 2, the UEIQM algorithm is always at the forefront in the performance evaluation of the four types of datasets. This further confirms its ability to evaluate image quality using QDE in various scenarios. In particular, the QDE chart intuitively shows the strengths and weaknesses of each UIQE method.

[0047] The above embodiments only express the implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

Claims

1. A comprehensive evaluation method for underwater image quality assessment performance based on dataset classification, characterized by: The specific steps are as follows: S1: Datasets The underwater image quality assessment datasets are classified according to the criteria of the distortion level image sequences and the applied subjective evaluation mode; The underwater image quality evaluation dataset is classified into a distortion level sequence dataset and a non-distortion level sequence dataset according to whether the image contains a distortion level sequence. The distortion level sequence dataset is divided into a distortion level sequence dataset with a changed imaging distance and a distortion level sequence dataset with a changed water turbidity according to the transformation conditions. At the same time, according to whether the images are grouped with the same content, the non-distortion level sequence dataset is subdivided into a distortion level sequence dataset in a grouping mode and a distortion level sequence dataset in a non-grouping mode. S2: Analysis of the four types of datasets obtained after classification, extracting the main and secondary indicators for evaluating underwater image quality assessment methods in each dataset; Main indicators of the distorted grade series dataset with changing imaging distance and changing water turbidity and the undistorted grade series dataset with grouping mode b i1 For L-test, secondary indicator b i2 is D-test, the main indicator of the undistorted graded sequence data set without grouping mode b i1 For PLCC, secondary indicator b i2 is P-test; S3: Calculate the comprehensive performance of each type based on the value range of the primary and secondary indicators and the analysis of the effect on the data set; Main indicatorsb i1 For the first aspect, the main indicator b i1 Absolute values ​​and secondary indicatorsb i2 Multiply as the second aspect, calculate the comprehensive index a for each type of data set i :a i (i=1,2,3,4)=w1×b i1 +w2×|b i1 |×b i2 ; Where w1 and w2 are weighting coefficients; S4: Construct a quadrilateral radar chart based on the comprehensive performance of each type; S5: Calculate the area and balance of the closed quadrilaterals in the quadrilateral radar image to obtain the comprehensive performance of the underwater image quality evaluation method.

2. The method for comprehensive evaluation of underwater image quality assessment performance based on data set classification according to claim 1 is characterized in that: The area S of the closed quadrilateral in the quadrilateral radar diagram in step S5 n for: The balance is in a5=a1.

Images