Low-quality gating image enhancement processing method suitable for underwater imaging

Abstract

The invention relates to a low-quality gating image enhancement processing method suitable for underwater imaging. The method comprises the following steps: 1) gating image acquisition: acquiring an original image through a gating camera; 2) environment parameter and image quality judgment: making environment parameter judgment according to a laser diffusion angle, a detection distance, a water quality condition and the like set by current equipment; determining feature extraction, feature matching parameters and frame superposition quantity in subsequent steps by combining the PSNR of the original image; 3) image preprocessing: carrying out filtering processing on the image; 4) spatial feature extraction: performing feature extraction on adjacent frames of images by using an SIFT algorithm to obtain key points and descriptors of each image; 5) feature matching and image transformation: after the SIFT features, namely key points and descriptors, of the image are extracted, feature matching and image transformation are carried out; 6) sliding type weighted frame superposition: in the video frame sequence, processing the image frame by frame in a sliding window mode to generate a frame of new composite image; and 7) outputting the processed image.

Description

Technical Field

[0001] The present invention relates to a method for enhancing low-quality grayscale images of underwater gated camera imaging, mainly using feature matching and image enhancement technologies to process low-quality, low-light, and high-noise images in underwater gated imaging.

[0002] Through this method, even in the case of severe attenuation of the laser signal, images with high signal-to-noise ratio and high contrast can be obtained, thereby increasing the detection angle or detection distance of the gated device. This method can achieve higher-quality image output at the same distance and extend the detection distance and angle at the same quality. Background Art

[0003] In the existing underwater imaging process, a combination of a gated camera and a pulsed laser is used to accurately capture underwater targets to meet the high requirements for image clarity, contrast, and detail capture. However, due to the complex underwater environment, during the light propagation process, the water medium has very serious attenuation and scattering effects on light. Even when using gated imaging, at long distances, the signal will be weaker than the noise in the gated area. In order to ensure image quality and imaging effect, it is necessary to conduct research on intelligent analysis and image processing technologies for key equipment in the underwater imaging process, provide effective guidance for improving image processing parameters, and improve the quality and efficiency of underwater imaging.

[0004] Although significant progress has been made in underwater imaging technology, there are still some challenges in practical applications. Traditional image processing methods are easily affected by various factors during operation, such as water turbidity, light refraction, and different object materials and reflectivities. In order to improve the detection efficiency of the gated camera, expand the detection area, and increase the detection distance, aiming at the problem that in the gated imaging process, due to excessive signal attenuation, the imaging target signal is weaker than the scattered noise in the gated area, resulting in blurred or distorted images, it is particularly important to propose an image processing method that can improve the clarity of underwater imaging under the conditions of a larger detection area and a farther detection distance at long distances. Summary of the Invention

[0005] Traditional image processing methods are created for single problems and single environments, but there are various situations in underwater gated imaging. Due to the complex underwater environment, poor light conditions, large variations in target material and reflectivity, the control requirements for the dynamic characteristics of the equipment during the imaging process are extremely high; and under the influence of light attenuation and scattering by the medium, the imaging signal-to-noise ratio and contrast are low; and there will be significant motion blur when using long exposure in a moving platform.

[0006] Based on the above problems, the present invention proposes a method that combines frame superposition technology with an image transformation algorithm based on feature matching, giving full play to the advantages of both, creating an image processing method that can adjust in real time according to the current environment and image quality, taking into account both the frame rate and the motion imaging quality, and solving the problem of low underwater imaging quality on moving platforms.

[0007] The present invention uses a mathematical model of water body scattering attenuation, and comprehensively discriminates the imaging quality in real time by comparing the limit detection performance of the device and the acceptable minimum PSNR according to the detection distance, detection angle set by the current device, and the image PSNR (Peak signal-to-noise ratio). According to the discrimination result, the parameters of feature extraction and matching are determined, and improvements are made according to the characteristics of weak underwater long-distance gated image signals, strong scattering noise, and unclear features, and an algorithm for enhancing underwater gated images in a moving state is proposed.

[0008] First, the original image is obtained by a gated camera cooperating with a pulsed laser. According to the currently set shutter delay and laser divergence angle, the detection distance, detection area, and image quality are calculated to determine the parameters for image feature extraction and transformation alignment and the superposition quantity.

[0009] Then, the determined parameters are passed in, and adaptive gamma transformation and dual-platform filtering processing are performed according to the parameters to enhance features and reduce the influence of noise on feature extraction, and increase the accuracy of feature localization;

[0010] Next, the Scale-invariant feature transform (SIFT) algorithm is used to extract features from adjacent frames, and the keypoints and descriptors of each image are obtained and recorded and saved;

[0011] After that, a feature point matching algorithm is used for preliminary screening to remove obviously invalid or unmatched noise feature points, so as to reduce the subsequent iteration data volume and accelerate image processing;

[0012] Then, the RANSAC algorithm is used to further iterate the feature pairs to obtain highly matched feature pairs, and the matching accuracy of the feature pairs is sorted;

[0013] Then, according to the feature matching accuracy, a multi-anchor point weighting method is adopted, and weights are assigned according to the matching accuracy to reduce the influence of perspective distortion and high noise, and improve the image transformation accuracy;

[0014] Next, the image is transformed according to the transformation matrix, and a sliding window weighted superposition method is adopted to perform weighted superposition according to the time series of frames, reducing noise and enhancing the signal strength while maintaining the frame rate.

[0015] Finally, an image with high signal-to-noise ratio and high contrast is obtained, and the frame rate is maintained without decrease. Brief Description of the Drawings

[0016] Figure 1 It shows a flowchart of the method of the present invention for improving the quality of underwater gated imaging on a mobile platform;

[0017] Figure 2(a) and Figure 2(b) respectively show the Gaussian pyramid and the Gaussian difference structure;

[0018] Figure 3 It shows the key point direction and descriptor;

[0019] Figure 4 It shows the schematic of feature matching;

[0020] Figure 5 It shows the sliding window algorithm. Detailed Embodiment

[0021] The present invention will be further described below with reference to the drawings.

[0022] The method for enhancing the processing of low-quality gated images applicable to underwater imaging in this example is as follows:

[0023] Step 1) Gated image acquisition: At present, underwater image acquisition cannot be accurately inferred through a mathematical model. It involves multiple factors such as water quality, laser diffusion angle, detection distance, laser power, and signal collection efficiency. Therefore, there is only a rough prediction for the imaging result. Therefore, it is difficult to predict the phenomenon that the target signal is lost or a large amount of scattered noise appears during the detection process, covering the true signal and causing the image to be blurred or distorted. Therefore, according to this method, first, the original image obtained by the gated camera is further processed according to the original image and the environment.

[0024] Step 2) Environmental parameter and image quality determination: Make an environmental parameter judgment based on the laser diffusion angle, detection distance, water quality situation, and the type of laser, laser power, and optical collection efficiency set by the current device, and match the PSNR of the original image to determine the subsequent feature extraction, feature matching parameters, and the number of frame superpositions, so as to achieve the adaptation of multiple environments and multiple imaging qualities.

[0025] The laser beam expansion angle is denoted as θ; the detection distance is obtained by setting the gated imaging system and denoted as z; the laser power is set by the laser and denoted as P 0 ; the water quality situation is determined by the attenuation coefficient, and the attenuation coefficient measurement is determined by the attenuation instrument. Its measurement formula is where P d is the laser power after attenuation over a distance d;

[0026] The optical collection efficiency is Where D is the effective diameter of the lens and z is the detection distance.

[0027] Step 3) Image preprocessing: Aiming at the characteristics of high noise in the underwater original image, reducing noise feature points and increasing the accuracy of feature points, an adaptive gamma curve and dual-platform threshold method combined filtering preprocessing is adopted. Before feature extraction and matching, the original image is preprocessed to smooth the high-frequency noise in the underwater gated image, reduce the number of invalid feature points and invalid feature pairs caused by noise, and greatly improve the accuracy of feature point positioning. While reducing the subsequent feature matching calculation amount, the positioning accuracy is improved, thus improving the speed and accuracy of the transformation matrix.

[0028] Step 4) Spatial feature extraction: In image processing, in order to extract significant and stable spatial features from an image, the SIFT (Scale-Invariant Feature Transform) algorithm adopts a series of steps to ensure that the features have scale invariance and rotation invariance.

[0029] Adjust the required number of feature points N according to the known environmental parameters in step 2 f , and the specific calculation method is as follows:

[0030]

[0031] Where k is the proportionality coefficient; PSNR 0 is the PSNR of the first frame image within the window.

[0032] The following are the spatial feature extraction steps of the SIFT algorithm:

[0033] a) Scale space extreme value detection: At different scales, the image is processed by the Difference of Gaussians (DoG) method to generate a series of scale spaces. As shown in Figures 2(a) and 2(b), the Gaussian pyramid images at different scales and their Gaussian difference images.

[0034] b) Key point localization: The potential feature points detected in the scale space (i.e., the extreme points in the DoG space) are further accurately located. Specifically, low-contrast points and edge response points are removed to ensure the stability of the key points, refer to Figure 3 .

[0035] c) Direction assignment: One or more directions are assigned to each key point, which is calculated based on the gradient direction histogram within the neighborhood of the key point. This can ensure that the features extracted in the subsequent steps have rotation invariance. Figure 3 The direction assignment of the key points can be seen on the left, and the gradient direction is represented by an arrow.

[0036] d) Key point description: Around each key point, extract an image region and calculate the gradient information within this region. Represent this information as a 128-dimensional feature vector (SIFT descriptor) for subsequent feature matching. Figure 3 The right side shows the SIFT descriptor extracted based on gradient information.

[0037] Step 5) Feature matching and image transformation: Refer to Figure 4 , after extracting the SIFT features of the image, the next task is to perform feature matching and image transformation to achieve image registration and stitching.

[0038] The following are the steps of the SIFT algorithm in feature matching and image transformation:

[0039] a) Feature matching: By calculating the Euclidean distance between SIFT descriptors, find the matching point pairs between images. As Figure 4 shown, the feature points in two images are matched through nearest neighbor search. To improve the accuracy of matching, the nearest neighbor distance ratio (NNDR) is often used for screening, and select the matching pairs with the ratio of the nearest neighbor to the second nearest neighbor distance less than a certain threshold.

[0040] b) Transformation model estimation: Use the Random Sample Consensus (RANSAC) algorithm to estimate the transformation model between images to exclude the influence of mis-matched point pairs. The transformation models adopted include the homography matrix and the fundamental matrix. The homography matrix is applicable to the case where there is a perspective transformation between images, while the fundamental matrix is applicable to depth estimation and 3D reconstruction in stereo vision. The homography matrix estimated iteratively by the RANSAC algorithm is used to transform the image through the homography transformation matrix, and calculate the distance of the feature points after transformation to evaluate the quality of the feature pairs.

[0041] c) Multi-anchor weighted transformation matrix: Select anchor points from high-quality matching feature pairs and perform weighted averaging on their transformation matrices, aiming to achieve more accurate image transformation. Effectively reduce the perspective distortion problem brought by a single transformation matrix, making the transformation of the overlapping area smoother and more natural. It can improve the quality of image matching in high-noise and complex scenes, and enhance the robustness and accuracy of the transformation.

[0042] d) Image registration and transformation: Apply the calculated transformation matrix to transform one image into the coordinate system of another image to achieve image registration.

[0043] Step 6) Sliding Frame Integration: Refer to Figure 5 , in the video frame sequence, process the images frame by frame in a sliding window manner. By performing weighted superposition on multiple frames within the sliding window, a new composite image is generated.

[0044] Specifically, each frame within the window is assigned a weight according to its distance from the target frame in the time series. The weight gradually decreases as the time series distance increases, thereby effectively reducing the deformation impact caused by the time span. At the same time, through weighted superposition, the target signal intensity can be significantly enhanced, and the influence of random scattering noise can be effectively reduced, improving the imaging quality.

[0045] The formula for weighted superposition is as follows:

[0046]

[0047] I final is the final image generated by weighting; I t is the t-th frame image within the window; λ is the empirical coefficient controlling the weight decay; t is the frame sequence number; n is the total number of frames required for superposition.

[0048] The calculation method for the total number of frames n required for superposition is as follows:

[0049]

[0050] where C is a proportionality constant.

[0051] Step 7) Image output.

Claims

1. A low-quality gated image enhancement processing method suitable for underwater imaging, the steps comprising: Step 1) Sorted image acquisition: the original image is acquired by the strobed camera; The method is characterized in that the steps further include: Step 2) Environmental parameters and image quality determination: Make environmental parameter judgments based on the laser diffusion angle, detection distance, water quality, laser type, laser power, and optical collection efficiency set by the current device; and determine feature extraction, feature matching parameters, and the number of frame overlays in subsequent steps by combining the peak signal-to-noise ratio (PSNR) of the original image; Step 3) Image preprocessing: filter the image to smooth the high-frequency noise in the underwater gated image; Step 4) Spatial feature extraction: Use the scale-invariant feature transformation (SIFT) algorithm to extract features from adjacent frame images and obtain key points and descriptors for each image; Step 5) Feature matching and image transformation: After extracting the SIFT features of the image, i.e., key points and descriptors, feature matching and image transformation are performed to achieve image registration and splicing; Step 6) Sliding weighted frame superposition: In the video frame sequence, the image is processed frame by frame in a sliding window manner, and all frames in the window are weightedly superimposed according to the temporal distance from the target image to generate a new synthetic image; Step 7) Output the processed image.

2. The low-quality gated image enhancement processing method suitable for underwater imaging according to claim 1 is characterized in that In step 3), the image is filtered and preprocessed using an adaptive gamma curve and a dual-platform threshold method.

3. The low-quality gated image enhancement processing method suitable for underwater imaging according to claim 1, characterized in that In step 4), the spatial feature extraction step of the SIFT algorithm includes: 4.1) Scale space extrema detection: At different scales, the image is processed by the Gaussian difference DoG method to generate a series of scale spaces; 4.1) Key point positioning: The potential feature points detected in the scale space are the extreme points in the DoG space, and are further accurately positioned by removing low-contrast points and edge response points to ensure the stability of the key points; 4.1) Direction assignment: Assign one or more directions to each key point, calculated based on the gradient direction histogram in the key point neighborhood; 4.1) Key point description: Extract an image area around each key point and calculate the gradient information in the area; represent the gradient information as a 128-dimensional feature vector, namely SIFT descriptor, for subsequent feature matching.

4. The low-quality gated image enhancement processing method suitable for underwater imaging according to claim 1 is characterized in that In step 5), the steps of SIFT algorithm in feature matching and image transformation include: 5.1) Feature matching: Find matching point pairs between images by calculating the Euclidean distance between SIFT descriptors; The feature points in the two images are matched by nearest neighbor search; the nearest neighbor distance ratio is used for screening, and the matching pairs whose nearest neighbor to next nearest neighbor distance ratio is less than a certain threshold are selected; 5.2) Transformation model estimation: Use the random sampling consistent RANSAC algorithm to estimate the transformation model between images; transform the image through the transformation model and calculate the distance of the feature points after the transformation to evaluate the quality of the feature pair; 5.3) Multi-anchor weighted transformation matrix: Select anchor points in the matching feature pairs and perform weighted averaging on their transformation matrices; 5.4) Image registration and transformation: Apply the calculated transformation matrix to transform one image into the coordinate system of another image to achieve image registration.

5. The low-quality gated image enhancement processing method suitable for underwater imaging according to claim 4 is characterized in that In step 5.2), the transformation model includes a homography matrix or a basic matrix; the homography matrix is ​​applicable to the case where there is a perspective transformation between images, and the basic matrix is ​​applicable to depth estimation and three-dimensional reconstruction in stereo vision.

Images