An anti-interference texture enhancement positioning method, system and medium suitable for low-texture visual scenes

Abstract

This application provides an anti-interference texture enhancement localization method, system, and medium suitable for low-texture visual scenes. The method includes: acquiring an original image; preprocessing the original image; extracting the texture response of the preprocessed image in different directions and frequency dimensions based on multi-scale Gabor filtering to obtain multiple texture response maps; dividing the preprocessed image into multiple image partitions based on a partitioned LED array and calculating the local gradient contrast of each image partition; independently adjusting the light source brightness of different image partitions to obtain an illumination-optimized image; performing feature fusion processing on the multiple texture response maps and the illumination-optimized image to extract feature points for visual localization; improving the consistency of images acquired at different time periods by normalizing the frequency response of the original image, histogram matching, and adaptive brightness restoration; and improving the local contrast by using a partitioned LED array with spatial illumination modulation to improve visual localization accuracy.

Description

Technical Field

[0001] This application relates to the field of texture localization technology, and more specifically, to an anti-interference texture enhancement localization method, system, and medium suitable for low-texture visual scenes. Background Technology

[0002] In computer vision systems, feature point extraction is closely related to the quality of image texture structure. However, in typical low-texture scenes such as natural grass, fabric, and homogeneous materials, due to the gradual grayscale changes and sparse edge information, traditional feature extraction algorithms based on spatial gradients or corner responses (such as Harris, SIFT, and ORB) often fail to obtain enough robust feature points, which seriously affects the stability and accuracy of subsequent tasks, especially showing significant degradation in tasks such as precise localization and visual matching.

[0003] Taking agricultural robots as an example, during actual field navigation, the ground surface is mostly composed of weeds, soil, and dead leaves. The texture structure in the image presents an unstructured, low-contrast, random distribution. Under the influence of changes in natural lighting (such as dappled sunlight and overlapping shadows), the contrast of local images further decreases, resulting in insufficient density of image feature points and significant drift in the navigation trajectory. In the industrial field, defect detection and localization of textile fabrics also face similar problems: many textile patterns have a periodic structure, but in areas with monotonous colors or curved edges, the texture gradually weakens, making it impossible for the vision system to stably lock the defect boundary.

[0004] Existing solutions often rely on increasing camera resolution, extending exposure time, or introducing high-power light sources for illumination enhancement. However, these methods frequently result in increased power consumption, increased image noise, or motion blur, and do not improve detectability at the fundamental structural level of texture. Therefore, there is an urgent need to propose a texture enhancement localization method suitable for low-texture scenes, starting from the frequency domain perspective of image structural information and combining it with spatial enhancement strategies. This method aims to achieve a dual improvement in the quantity and quality of feature extraction, ultimately serving high-precision localization, detection, and recognition tasks. Summary of the Invention

[0005] The purpose of this application is to provide an anti-interference texture enhancement positioning method, system and medium suitable for low-texture visual scenes. By normalizing the frequency response processing of the original image, histogram matching and adaptive brightness restoration, the consistency of images acquired at different time periods is improved. By using a partitioned LED array to enhance local contrast through spatial illumination modulation, the visual positioning accuracy is improved.

[0006] This application also provides an anti-interference texture enhancement localization method suitable for low-texture visual scenes, including:

[0007] Obtain the original image, preprocess the original image to obtain the preprocessed image;

[0008] Based on multi-scale Gabor filtering, texture response of preprocessed images under different directions and frequency dimensions is extracted to obtain multiple texture response maps;

[0009] The preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated.

[0010] Based on local gradient contrast, the brightness of the light source is adjusted independently for different image regions to obtain an optimized illumination image;

[0011] Multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization.

[0012] Optionally, in the anti-interference texture enhancement and localization method for low-texture visual scenes described in the embodiments of this application, the original image is acquired, and the original image is preprocessed to obtain a preprocessed image, specifically including:

[0013] Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range.

[0014] Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering.

[0015] The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block.

[0016] The preprocessed image is obtained by improving local contrast through histogram equalization.

[0017] Optionally, in the anti-interference texture enhancement and localization method for low-texture visual scenes described in the embodiments of this application, multiple texture response maps are obtained by extracting the texture response of the preprocessed image in different directions and frequency dimensions based on multi-scale Gabor filtering, specifically including:

[0018] Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows: ; in: ; ; In the formula: The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle; It is the new ordinate after the same rotation; Wavelength, controlling frequency; Direction: Controls the direction of the filter's main axis; Phase shift; : Standard deviation of Gaussian envelope; Spatial aspect ratio controls the shape of the ellipse; After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained: ; : Indicates a specific scale (wavelength) and specific directions Below is the texture response map obtained after Gabor filtering; in , , where represents the number of samples for scale and direction, respectively.

[0019] Optionally, in the anti-interference texture enhancement and localization method for low-texture visual scenes described in the embodiments of this application, the image is divided into partitions. Each zone corresponds to an adjustable light source intensity. Its goal is to maximize the local gradient contrast function:

[0020]

[0021] It is the average gradient intensity within the partition, used to evaluate the visual information richness of the region;

[0022] It is the gradient of the image at point (x,y), reflecting the local changes at that point;

[0023] Local contrast is calculated using gradient operators such as Sobel. And define the illumination adjustment function:

[0024]

[0025] In the formula: This represents the brightness of the light source at the current moment, i.e., the [number]th [time]. Each image partition in the number of iterations LED light source brightness value at that time;

[0026] This indicates the brightness of the light source at the next moment, i.e., the brightness of the light source at the next moment. Each image partition in the number of iterations The updated LED light source brightness value;

[0027] Represents the local gradient contrast function, i.e., the th The average value of the image gradient within each partition reflects the texture richness of that region. This represents the partial derivative of contrast with respect to brightness, indicating the degree to which changes in the brightness of the current partition affect the contrast, and is used for gradient ascent optimization.

[0028] By independently adjusting the light intensity of each partition, the global contrast of the image is optimized, thus improving the response quality of the texture filter.

[0029] Optionally, in the anti-interference texture enhancement and localization method for low-texture visual scenes described in this application embodiment, feature fusion processing is performed on multiple texture response maps and the illumination-optimized avatar, specifically including:

[0030] Calculate the maximum response fusion for all texture response maps:

[0031]

[0032] Indicates the location At this location, the response value with the largest absolute value in all multi-scale, multi-directional Gabor texture response maps;

[0033] Alternatively, a weighted fusion approach may be adopted:

[0034]

[0035] Indicates the location At this point, the composite texture response value is obtained by weighted fusion of all Gabor response maps;

[0036] Among them, weight Dynamically adjust based on the mean response, directional prior, or feature feedback;

[0037] fused images The texture response-enhanced image is obtained by overlaying it with the original image:

[0038] ;

[0039] This represents the fused texture response map. or One of them;

[0040] This represents the final enhanced image, derived from the original image. and texture response map It is made by blending in a certain proportion;

[0041] in Control the fusion intensity.

[0042] Optionally, in the anti-interference texture enhancement and localization method for low-texture visual scenes described in this application embodiment, extracting feature points from the texture response enhancement image specifically includes: using the SIFT operator to detect interest points, with the response function being:

[0043]

[0044] in For Gaussian kernel, As a scale, Scale ratio;

[0045] Apply the operation to the response enhancement image Increase the number of interest point detections The lift rate is defined as:

[0046] ;

[0047] In the formula: This represents the feature point count improvement rate, which is the relative increase in the number of feature points in the enhanced image compared to the original image.

[0048] This represents the number of feature points detected in the enhanced image, using operators such as SIFT to enhance the texture response of the image. The result is obtained through feature extraction;

[0049] This represents the number of feature points detected in the original image, obtained by feature extraction from the unenhanced original image.

[0050] Secondly, embodiments of this application provide an anti-interference texture enhancement localization system suitable for low-texture visual scenes. The system includes a memory and a processor. The memory includes a program for an anti-interference texture enhancement localization method suitable for low-texture visual scenes. When the program for the anti-interference texture enhancement localization method suitable for low-texture visual scenes is executed by the processor, it implements the following steps:

[0051] Obtain the original image, preprocess the original image to obtain the preprocessed image;

[0052] Based on multi-scale Gabor filtering, texture response of preprocessed images under different directions and frequency dimensions is extracted to obtain multiple texture response maps;

[0053] The preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated.

[0054] Based on local gradient contrast, the brightness of the light source is adjusted independently for different image regions to obtain an optimized illumination image;

[0055] Multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization.

[0056] Optionally, in the anti-interference texture enhancement positioning system for low-texture visual scenes described in this application embodiment, the original image is acquired, and the original image is preprocessed to obtain a preprocessed image, specifically including:

[0057] Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range.

[0058] Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering.

[0059] The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block.

[0060] The preprocessed image is obtained by improving local contrast through histogram equalization.

[0061] Optionally, in the anti-interference texture enhancement and localization system for low-texture visual scenes described in this application embodiment, multiple texture response maps are obtained by extracting the texture response of the preprocessed image in different directions and frequency dimensions based on multi-scale Gabor filtering, specifically including:

[0062] Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows:

[0063] ;

[0064] in:

[0065]

[0066]

[0067] In the formula:

[0068] The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle;

[0069] It is the new ordinate after the same rotation;

[0070] Wavelength, controlling frequency;

[0071] Direction: Controls the direction of the filter's main axis;

[0072] Phase shift;

[0073] : Standard deviation of Gaussian envelope;

[0074] Spatial aspect ratio controls the shape of the ellipse;

[0075] After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained:

[0076]

[0077] : Indicates a specific scale (wavelength) and specific directions Below is the texture response map obtained after Gabor filtering;

[0078] in , , where represents the number of samples for scale and direction, respectively.

[0079] Thirdly, embodiments of this application also provide a computer-readable storage medium, which includes a program for an anti-interference texture enhancement and localization method suitable for low-texture visual scenes. When the program for the anti-interference texture enhancement and localization method suitable for low-texture visual scenes is executed by a processor, it implements the steps of the anti-interference texture enhancement and localization method suitable for low-texture visual scenes as described in any of the above claims.

[0080] As can be seen from the above, the anti-interference texture enhancement localization method, system, and medium provided in this application embodiment, suitable for low-texture visual scenes, involves acquiring an original image, preprocessing the original image to obtain a preprocessed image, extracting the texture response of the preprocessed image in different directions and frequency dimensions based on multi-scale Gabor filtering to obtain multiple texture response maps, dividing the preprocessed image into multiple image partitions based on a partitioned LED array, and calculating the local gradient contrast of each image partition, independently adjusting the light source brightness of different image partitions based on the local gradient contrast to obtain an illumination-optimized image, performing feature fusion processing on the multiple texture response maps and the illumination-optimized image to obtain a texture response enhancement image, and extracting feature points from the texture response enhancement image for visual localization, improving the consistency of images acquired at different time periods through normalized frequency response processing, histogram matching, and adaptive brightness restoration of the original image, and enhancing local contrast through spatial illumination modulation using a partitioned LED array to improve visual localization accuracy. Attached Figure Description

[0081] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0082] Figure 1 A flowchart illustrating an anti-interference texture enhancement localization method for low-texture visual scenes provided in this application embodiment;

[0083] Figure 2 A block diagram of an anti-interference texture enhancement positioning system for low-texture visual scenes provided in this application embodiment;

[0084] Figure 3 Schematic diagram of LCTF band selection and voltage control principle for an anti-interference texture enhancement positioning system suitable for low-texture visual scenes provided in the embodiments of this application;

[0085] Figure 4 A flowchart illustrating the band subset acquisition and neural network reconstruction process for an anti-interference texture enhancement positioning system suitable for low-texture visual scenes, provided in an embodiment of this application. Detailed Implementation

[0086] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0087] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0088] Please refer to Figure 1 , Figure 1 This is a flowchart of an anti-interference texture enhancement and localization method for low-texture visual scenes, as described in some embodiments of this application. This anti-interference texture enhancement and localization method for low-texture visual scenes is used in a terminal device and includes the following steps:

[0089] S101, Obtain the original image, preprocess the original image to obtain the preprocessed image;

[0090] S102, based on multi-scale Gabor filtering, extract the texture response of the preprocessed image under different directions and frequency dimensions to obtain multiple texture response maps;

[0091] S103, the preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated;

[0092] S104, Based on local gradient contrast, the brightness of the light source is independently adjusted for different image partitions to obtain an optimized illumination image;

[0093] S105, multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points of the texture response enhancement image are extracted for visual localization.

[0094] It should be noted that this application proposes a frequency domain-spatial domain joint filtering algorithm, which uses Gabor wavelet transform to enhance the texture response of images at specific directions and scales;

[0095] A dynamic illumination compensation device based on a controllable LED zone array is integrated to enhance local contrast through spatial illumination modulation.

[0096] A feature extraction feedback mechanism is constructed to dynamically adjust the filter response parameters using local image texture information as the driving force.

[0097] By employing a multi-channel texture enhancement result fusion strategy, seamless feature transfer between the enhanced image and the original image is achieved.

[0098] According to an embodiment of the present invention, an original image is acquired, and the original image is preprocessed to obtain a preprocessed image, specifically including:

[0099] Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range.

[0100] Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering.

[0101] The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block.

[0102] The preprocessed image is obtained by improving local contrast through histogram equalization.

[0103] It should be noted that the original images of low-texture scenes may have inconsistent pixel grayscale value distribution ranges due to differences in acquisition devices (such as industrial cameras, drone cameras) and exposure parameters during shooting (for example, some images have pixel values ​​ranging from 0-100, while others range from 0-255). Normalization, through linear or non-linear transformations, maps all pixel grayscale values ​​of the original images to a fixed range (usually 0-1 or 0-255).

[0104] Choose the appropriate algorithm based on the type of noise: For salt-and-pepper noise, median filtering is commonly used (to eliminate isolated noise points by replacing the original pixel with the median value in the neighborhood of the pixel); for Gaussian noise, Gaussian filtering is commonly used (to construct a Gaussian kernel centered on the pixel and smooth the image by weighted averaging to reduce continuous noise); if the noise is complex, bilateral filtering can be used (to consider both spatial distance and gray-level similarity, and to preserve key texture structures such as image edges while reducing noise).

[0105] The image is divided into multiple non-overlapping or partially overlapping local blocks. A gray-level histogram is calculated for each block. Local contrast is improved by histogram equalization (stretching the concentrated gray-level range to a wider range). For example, dark pixels in shadow areas are stretched to higher gray-level values, and bright pixels in reflective areas are adjusted to more appropriate gray-level values. At the same time, interpolation is used to eliminate abrupt gray-level changes at the boundaries between blocks.

[0106] According to an embodiment of the present invention, texture responses of a preprocessed image in different directions and frequency dimensions are extracted based on multi-scale Gabor filtering to obtain multiple texture response maps, specifically including:

[0107] Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows:

[0108] ;

[0109] in:

[0110]

[0111]

[0112] In the formula:

[0113] The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle;

[0114] It is the new ordinate after the same rotation;

[0115] Wavelength, controlling frequency;

[0116] Direction: Controls the direction of the filter's main axis;

[0117] Phase shift;

[0118] : Standard deviation of Gaussian envelope;

[0119] Spatial aspect ratio controls the shape of the ellipse;

[0120] After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained:

[0121]

[0122] : Indicates a specific scale (wavelength) and specific directions Below is the texture response map obtained after Gabor filtering;

[0123] in , , where represents the number of samples for scale and direction, respectively.

[0124] It should be noted that the Gabor filter can be set to multiple directions (such as 0°, 45°, 90°, 135°, or even more subdivided directions), which can specifically extract textures with different orientations (such as the horizontal fibers of fabrics and the oblique veins of grass), avoiding the omission of features due to the single texture orientation.

[0125] Scale selectivity: The scale corresponds to the coarseness of the texture—small-scale (high-frequency) filters can capture fine textures (such as tiny scratches on a metal surface), while large-scale (low-frequency) filters can capture macroscopic textures (such as the overall undulation of a grassland). By setting multiple scales (usually 3-5 scales, such as σ=1, 2, 3, where σ is the standard deviation of the Gaussian kernel, and the larger σ is, the larger the scale), comprehensive coverage of textures of different coarseness can be achieved.

[0126] According to an embodiment of the present invention, the image is partitioned as follows: Each zone corresponds to an adjustable light source intensity. Its goal is to maximize the local gradient contrast function:

[0127]

[0128] It is the average gradient intensity within the partition, used to evaluate the visual information richness of the region;

[0129] It is the gradient of the image at point (x,y), reflecting the local changes at that point;

[0130] Local contrast is calculated using gradient operators such as Sobel. And define the illumination adjustment function:

[0131] ;

[0132] In the formula: This represents the brightness of the light source at the current moment, i.e., the [number]th [time]. Each image partition in the number of iterations LED light source brightness value at that time;

[0133] This indicates the brightness of the light source at the next moment, i.e., the brightness of the light source at the next moment. Each image partition in the number of iterations The updated LED light source brightness value;

[0134] Represents the local gradient contrast function, i.e., the th The average value of the image gradient within each partition reflects the texture richness of that region. This represents the partial derivative of contrast with respect to brightness, indicating the degree to which changes in the brightness of the current partition affect the contrast, and is used for gradient ascent optimization.

[0135] By independently adjusting the light intensity of each partition, the global contrast of the image is optimized, thus improving the response quality of the texture filter.

[0136] It should be noted that for areas with low local gradient contrast (i.e., low-contrast areas, such as shadow areas), the brightness of the corresponding LED beads is increased to enhance the light intensity in that area and expand the pixel grayscale difference within the area; for areas with high local gradient contrast (i.e., high-contrast areas, such as bright areas), the brightness of the LED beads is maintained or reduced to avoid overexposure.

[0137] According to an embodiment of the present invention, feature fusion processing is performed on multiple texture response maps and a lighting-optimized avatar, specifically including:

[0138] Calculate the maximum response fusion for all texture response maps:

[0139]

[0140] Indicates the location At this location, the response value with the largest absolute value in all multi-scale, multi-directional Gabor texture response maps;

[0141] Alternatively, a weighted fusion approach may be adopted:

[0142] ;

[0143] Indicates the location At this point, the composite texture response value is obtained by weighted fusion of all Gabor response maps;

[0144] Among them, weight Dynamically adjust based on the mean response, directional prior, or feature feedback;

[0145] fused images The texture response-enhanced image is obtained by overlaying it with the original image:

[0146] ;

[0147] This represents the fused texture response map. or One of them;

[0148] This represents the final enhanced image, derived from the original image. and texture response map It is made by blending in a certain proportion;

[0149] in Control the fusion intensity.

[0150] It should be noted that the multi-scale-directional texture response maps each emphasize different dimensions of texture (e.g., the small-scale response map emphasizes fine textures, while the large-scale response map emphasizes macroscopic textures), while the illumination-optimized image emphasizes the basic image structure after illumination equalization. The purpose of fusion is to integrate and complement this information, avoiding the limitations of a single information source.

[0151] According to an embodiment of the present invention, extracting feature points from a texture response-enhanced image specifically includes: using the SIFT operator to detect interest points, with the response function being:

[0152]

[0153] in For Gaussian kernel, As a scale, Scale ratio;

[0154] Apply the operation to the response enhancement image Increase the number of interest point detections The lift rate is defined as:

[0155] ;

[0156] In the formula: This represents the feature point count improvement rate, which is the relative increase in the number of feature points in the enhanced image compared to the original image.

[0157] This represents the number of feature points detected in the enhanced image, using operators such as SIFT to enhance the texture response of the image. The result is obtained through feature extraction;

[0158] This represents the number of feature points detected in the original image, obtained by feature extraction from the unenhanced original image.

[0159] Experiments show that this ratio can reach over 5.3 on typical grassland images, and is improved by an average of 4.8 times in textile fabric detection tasks, which is significantly better than existing enhancement technologies such as CLAHE and Retinex.

[0160] According to embodiments of the present invention, the application of texture enhancement and positioning systems in typical scenarios is expanded.

[0161] I. Application Details in Agricultural Navigation Scenarios

[0162] In intelligent agricultural equipment (such as driverless tractors and harvesting robots), texture information is an important visual cue for path recognition, fruit localization, and crop health status analysis. However, texture in natural scenes is often affected by the following factors:

[0163] Changes in light (sunlight, shadow, dawn and dusk);

[0164] Low-contrast texture between soil and vegetation;

[0165] Crops shading and overlapping cause blurred boundaries.

[0166] This system adaptively enhances fine-grained textures in low-contrast environments through a multi-scale texture enhancement network module and a spectral channel fusion mechanism. The system's core capabilities are as follows:

[0167] Using frequency domain enhancement residual blocks on the input image Enhanced response after multi-scale convolution:

[0168]

[0169] In the formula, Indicates the location of the enhanced image. The grayscale value at that location; Indicates the original image at the 1st... Response maps after Laplacian convolution at various scales;

[0170] in For the first Laplace operator of scale, The enhanced weights are obtained through training.

[0171] In multi-band scenarios, channel attention mechanism is used to select the spectral segments most relevant to texture changes, thereby improving the separability of micro-features such as soil cracks and crop leaf veins.

[0172] In agricultural navigation experiments, the system was applied to sugarcane field and strawberry picking route planning, reducing the average boundary deviation by 22% and improving the navigation line extraction stability rate by more than 30% compared with traditional CNN navigation systems.

[0173] II. Application Details in Textile Defect Detection

[0174] In the textile manufacturing process, fabrics have complex and repetitive texture backgrounds (such as checks, twill, printed patterns, etc.). Traditional algorithms based on grayscale / edge detection are prone to misidentifying texture repetition as defects or missing out on real defects such as minor breaks, skipped stitches, and stains.

[0175] The advantages of this application are:

[0176] We can build a texture-aware encoder-decoder architecture that can generate a background prediction map after modeling repetitive textures. Then, the difference between the original image and the result is used to obtain the abnormal response image:

[0177]

[0178] in, Indicates the location of the abnormal response graph. The value at that location indicates the degree of difference between that location and the background; the larger the value, the more likely it is to be a defective area. This indicates the location of the background prediction image generated by the texture-aware encoder-decoder. The grayscale value or feature value at the location represents a flawless textured background.

[0179] Simultaneously, the wavelet transform + spatial frequency reconstruction module is used to distinguish between periodic texture disturbances and non-periodic defect responses, thereby improving the system's fault tolerance.

[0180] In the Textile-500 image library test, this system achieved F1-scores of 0.93 and 0.96 for patterned and solid-color fabrics, respectively, which are significantly better than classic methods such as SIFT / SVM and U-Net, and its performance is particularly stable on complex printed fabrics.

[0181] like Figures 2-4 As shown, Figure 2 The system includes an optical acquisition module, comprising an industrial camera and a liquid crystal tunable filter (LCTF) for dynamic band acquisition; a data reconstruction module, a band completion system composed of neural networks, which reconstructs complete images of 10+ bands based on a few acquired bands of data; and a system scheduling and synchronization control module, which coordinates filter voltage control, camera exposure timing, and neural network triggering to ensure high-speed and stable operation. The overall logic follows an input-central processing-output flow, reflecting the integrated architecture of physical hardware and algorithm software.

[0182] Specifically, Figure 2 In this context, Optical Acquisition Module represents the optical acquisition module; Liquid Crystal Tunable Filter represents the liquid crystal tunable filter; High-Speed ​​Industrial Camera represents the high-speed industrial camera; Input represents the input; Band Reconstruction Module represents the band reconstruction module; System Scheduling and Synchronization Control Module represents the system scheduling and synchronization control module; and Overall System Architecture represents the overall system architecture.

[0183] Figure 3 This section explains the LCTF response mechanism and fast switching method. The horizontal axis represents the driving voltage V, and the vertical axis represents the center transmission wavelength λ. The figure shows the nonlinear variation curve of the filter center wavelength with voltage. The right side also indicates the "asymmetric sinusoidal pre-bias excitation" in the fast response strategy, whose control function is: ;

[0184] In the formula: This represents the driving voltage as it changes over time, and the unit is volts (V). It represents the bias voltage, providing a base voltage level for presetting the initial state of the liquid crystal molecules; It represents the amplitude of the sinusoidal excitation, controlling the magnitude of voltage fluctuations; It represents the angular frequency, which controls the oscillation frequency of the sine wave. ,in For frequency; Represents a time variable; This indicates a phase offset, used to adjust the starting phase of a sine wave and optimize the switching timing.

[0185] Specifically, Figure 3 In this context, "liquid crystal" refers to liquid crystal; "transparent electrodes" refers to transparent electrodes; "incident light" refers to incident light; "selected light" refers to selected light; and "LCTF wavelength selection and voltage control principle" refers to the LCTF wavelength selection and voltage control principle.

[0186] This control strategy can reduce liquid crystal response hysteresis, accelerate band switching, and effectively shorten the filter response time.

[0187] Figure 4 This describes the neural network processing procedure for few-band acquisition to full-band reconstruction. The left side shows the actual acquired image subset (e.g., 3–5 bands), the middle shows the convolutional neural network structure (including encoder, skip connections, and decoder), and the right side outputs the fully reconstructed multi-band image sequence. Specifically, "band subset acquisition" represents the acquisition of a subset of bands.

[0188] Neural network reconstruction refers to the reconstruction of a neural network.

[0189] This process embodies a core innovation: utilizing a learning-based mapping function to reduce reliance on the number of physical filters.

[0190] Secondly, embodiments of this application provide an anti-interference texture enhancement localization system suitable for low-texture visual scenes. The system includes a memory and a processor. The memory includes a program for an anti-interference texture enhancement localization method suitable for low-texture visual scenes. When the program for the anti-interference texture enhancement localization method suitable for low-texture visual scenes is executed by the processor, it implements the following steps:

[0191] Obtain the original image, preprocess the original image to obtain the preprocessed image;

[0192] Based on multi-scale Gabor filtering, texture response of preprocessed images under different directions and frequency dimensions is extracted to obtain multiple texture response maps;

[0193] The preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated.

[0194] Based on local gradient contrast, the brightness of the light source is adjusted independently for different image regions to obtain an optimized illumination image;

[0195] Multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization.

[0196] It should be noted that this system consists of the following modules: an image preprocessing module, which performs normalization, noise reduction, and local histogram equalization;

[0197] A multi-scale Gabor filtering module extracts texture responses at different directions and frequencies;

[0198] The illumination compensation control module adjusts the LED array intensity matrix in real time to improve the response amplitude in low-contrast areas;

[0199] The feature fusion module constructs a texture-aware enhanced image based on response intensity;

[0200] The feature point extraction module uses enhanced images to extract features using algorithms such as SIFT / ORB, resulting in a high-density, highly robust feature set.

[0201] Furthermore, this system also includes a multi-scale frequency-aware texture enhancement module:

[0202] Enhance weak texture features using trainable high-pass filters and residual coupling structures;

[0203] A robust image representation is implemented to facilitate use by downstream localization and detection modules.

[0204] Spectral channel selective fusion mechanism:

[0205] By fusing multispectral information and dynamically selecting the channel with the best texture discrimination capability, the ability to distinguish blurred areas is improved;

[0206] Channel attention mapping function:

[0207]

[0208] in For the first Channel feature map, This represents the channel weight.

[0209] Self-supervised texture modeling and differential localization mechanism:

[0210] Generate predicted background images using unsupervised texture modeling;

[0211] Differential strategies are used to locate abnormal regions and construct a self-labeled sample space for use by subsequent meta-learning modules.

[0212] Spectral entropy-driven reconstruction verification mechanism:

[0213] Image frequency domain spectral entropy is introduced as an image quality index for image self-detection and automatic back-down mechanisms;

[0214] When the image texture information entropy is lower than the set threshold, the system automatically reduces the location confidence or requests a second sampling.

[0215] According to an embodiment of the present invention, an original image is acquired, and the original image is preprocessed to obtain a preprocessed image, specifically including:

[0216] Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range.

[0217] Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering.

[0218] The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block.

[0219] The preprocessed image is obtained by improving local contrast through histogram equalization.

[0220] According to an embodiment of the present invention, texture responses of a preprocessed image in different directions and frequency dimensions are extracted based on multi-scale Gabor filtering to obtain multiple texture response maps, specifically including:

[0221] Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows:

[0222] ;

[0223] in:

[0224]

[0225]

[0226] In the formula:

[0227] The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle;

[0228] It is the new ordinate after the same rotation;

[0229] Wavelength, controlling frequency;

[0230] Direction: Controls the direction of the filter's main axis;

[0231] Phase shift;

[0232] : Standard deviation of Gaussian envelope;

[0233] Spatial aspect ratio controls the shape of the ellipse;

[0234] After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained:

[0235]

[0236] in , , where represents the number of samples for scale and direction, respectively.

[0237] It should be noted that the Gabor filter can be set to multiple directions (such as 0°, 45°, 90°, 135°, or even more subdivided directions), which can specifically extract textures with different orientations (such as the horizontal fibers of fabrics and the oblique veins of grass), avoiding the omission of features due to the single texture orientation.

[0238] Scale selectivity: "Scale" corresponds to the "coarseness" of the texture—small-scale (high-frequency) filters can capture fine textures (such as tiny scratches on a metal surface), while large-scale (low-frequency) filters can capture macroscopic textures (such as the overall undulation of a grassland). By setting multiple scales (usually 3-5 scales, such as σ=1, 2, 3, where σ is the standard deviation of the Gaussian kernel, and the larger σ is, the larger the scale), comprehensive coverage of textures of different coarseness can be achieved.

[0239] According to an embodiment of the present invention, the image is partitioned as follows: Each zone corresponds to an adjustable light source intensity. Its goal is to maximize the local gradient contrast function:

[0240]

[0241] It is the average gradient intensity within the partition, used to evaluate the visual information richness of the region;

[0242] It is the gradient of the image at point (x,y), reflecting the local changes at that point;

[0243] Local contrast is calculated using gradient operators such as Sobel. And define the illumination adjustment function:

[0244]

[0245] By independently adjusting the light intensity of each partition, the global contrast of the image is optimized, thus improving the response quality of the texture filter.

[0246] It should be noted that for areas with low local gradient contrast (i.e., low-contrast areas, such as shadow areas), the brightness of the corresponding LED beads is increased to enhance the light intensity in that area and expand the pixel grayscale difference within the area; for areas with high local gradient contrast (i.e., high-contrast areas, such as bright areas), the brightness of the LED beads is maintained or reduced to avoid overexposure.

[0247] According to an embodiment of the present invention, feature fusion processing is performed on multiple texture response maps and a lighting-optimized avatar, specifically including:

[0248] Calculate the maximum response fusion for all texture response maps:

[0249]

[0250] Indicates the location At this location, the response value with the largest absolute value in all multi-scale, multi-directional Gabor texture response maps;

[0251] Alternatively, a weighted fusion approach may be adopted:

[0252] ;

[0253] Indicates the location At this point, the composite texture response value is obtained by weighted fusion of all Gabor response maps;

[0254] Among them, weight Dynamically adjust based on the mean response, directional prior, or feature feedback;

[0255] fused images The texture response-enhanced image is obtained by overlaying it with the original image:

[0256] ;

[0257] This represents the fused texture response map. or One of them;

[0258] This represents the final enhanced image, derived from the original image. and texture response map It is made by blending in a certain proportion;

[0259] in Control the fusion intensity.

[0260] It should be noted that the multi-scale-directional texture response maps each emphasize different dimensions of texture (e.g., the small-scale response map emphasizes fine textures, while the large-scale response map emphasizes macroscopic textures), while the illumination-optimized image emphasizes the basic image structure after illumination equalization. The purpose of fusion is to integrate and complement this information, avoiding the limitations of a single information source.

[0261] According to an embodiment of the present invention, extracting feature points from a texture response-enhanced image specifically includes: using the SIFT operator to detect interest points, with the response function being:

[0262]

[0263] in For Gaussian kernel, As a scale, Scale ratio;

[0264] Apply the operation to the response enhancement image Increase the number of interest point detections The lift rate is defined as:

[0265] ;

[0266] In the formula: This represents the feature point count improvement rate, which is the relative increase in the number of feature points in the enhanced image compared to the original image.

[0267] This represents the number of feature points detected in the enhanced image, using operators such as SIFT to enhance the texture response of the image. The result is obtained through feature extraction;

[0268] This represents the number of feature points detected in the original image, obtained by feature extraction from the unenhanced original image.

[0269] Experiments show that this ratio can reach over 5.3 on typical grassland images, and is improved by an average of 4.8 times in textile fabric detection tasks, which is significantly better than existing enhancement technologies such as CLAHE and Retinex.

[0270] According to embodiments of the present invention, the application of texture enhancement and positioning systems in typical scenarios is expanded.

[0271] I. Application Details in Agricultural Navigation Scenarios

[0272] In intelligent agricultural equipment (such as driverless tractors and harvesting robots), texture information is an important visual cue for path recognition, fruit localization, and crop health status analysis. However, texture in natural scenes is often affected by the following factors:

[0273] Changes in light (sunlight, shadow, dawn and dusk);

[0274] Low-contrast texture between soil and vegetation;

[0275] Crops shading and overlapping cause blurred boundaries.

[0276] This system adaptively enhances fine-grained textures in low-contrast environments through a multi-scale texture enhancement network module and a spectral channel fusion mechanism. The system's core capabilities are as follows:

[0277] Using frequency domain enhancement residual blocks on the input image Enhanced response after multi-scale convolution:

[0278]

[0279] in For the first Laplace operator of scale, The enhanced weights are obtained through training.

[0280] In multi-band scenarios, channel attention mechanism is used to select the spectral segments most relevant to texture changes, thereby improving the separability of micro-features such as soil cracks and crop leaf veins.

[0281] In agricultural navigation experiments, the system was applied to sugarcane field and strawberry picking route planning, reducing the average boundary deviation by 22% and improving the navigation line extraction stability rate by more than 30% compared with traditional CNN navigation systems.

[0282] II. Application Details in Textile Defect Detection

[0283] In the textile manufacturing process, fabrics have complex and repetitive texture backgrounds (such as checks, twill, printed patterns, etc.). Traditional algorithms based on grayscale / edge detection are prone to misidentifying texture repetition as defects or missing out on real defects such as minor breaks, skipped stitches, and stains.

[0284] The advantages of this application are:

[0285] We can build a texture-aware encoder-decoder architecture that can generate a background prediction map after modeling repetitive textures. Then, the difference between the original image and the result is used to obtain the abnormal response image:

[0286]

[0287] Simultaneously, the wavelet transform + spatial frequency reconstruction module is used to distinguish between periodic texture disturbances and non-periodic defect responses, thereby improving the system's fault tolerance.

[0288] In the Textile-500 image library test, this system achieved F1-scores of 0.93 and 0.96 for patterned and solid-color fabrics, respectively, which are significantly better than classic methods such as SIFT / SVM and U-Net, and its performance is particularly stable on complex printed fabrics.

[0289] A third aspect of the present invention provides a computer-readable storage medium including a program for an anti-interference texture enhancement localization method suitable for low-texture visual scenes. When the program for the anti-interference texture enhancement localization method suitable for low-texture visual scenes is executed by a processor, it implements the steps of the anti-interference texture enhancement localization method suitable for low-texture visual scenes as described in any of the above claims.

[0290] This invention discloses an anti-interference texture enhancement and localization method, system, and medium suitable for low-texture visual scenes. The method involves acquiring an original image, preprocessing it to obtain a preprocessed image, extracting the texture response of the preprocessed image in different directions and frequency dimensions using multi-scale Gabor filtering to obtain multiple texture response maps, dividing the preprocessed image into multiple image partitions using a partitioned LED array, and calculating the local gradient contrast of each partition. Based on the local gradient contrast, the brightness of the light source is independently adjusted for different image partitions to obtain an illumination-optimized image. The multiple texture response maps are then fused with the illumination-optimized image to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization. By normalizing the frequency response of the original image, histogram matching, and adaptive brightness restoration, the consistency of images acquired at different time periods is improved. The local contrast is enhanced by spatial illumination modulation using a partitioned LED array, thereby improving visual localization accuracy.

[0291] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0292] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0293] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0294] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0295] Alternatively, if the integrated units of the present invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a readable storage medium. Based on this understanding, the technical solutions of the embodiments of the present invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A method for anti-interference texture enhancement and localization suitable for low-texture visual scenes, characterized in that, include: Obtain the original image, preprocess the original image to obtain the preprocessed image; Based on multi-scale Gabor filtering, texture response of preprocessed images under different directions and frequency dimensions is extracted to obtain multiple texture response maps; The preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated. Based on local gradient contrast, the brightness of the light source is adjusted independently for different image regions to obtain an optimized illumination image; Multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization.

2. The anti-interference texture enhancement and localization method for low-texture visual scenes according to claim 1, characterized in that, Obtain the original image, perform preprocessing on the original image to obtain the preprocessed image, specifically including: Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range. Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering. The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block. The preprocessed image is obtained by improving local contrast through histogram equalization.

3. The anti-interference texture enhancement and localization method for low-texture visual scenes according to claim 2, characterized in that, Based on multi-scale Gabor filtering, texture responses of preprocessed images are extracted under different directions and frequency dimensions to obtain multiple texture response maps, specifically including: Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows: ； in: ； ； In the formula: The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle; It is the new ordinate after the same rotation; Wavelength, controlling frequency; Direction: Controls the direction of the filter's main axis; Phase shift; : Standard deviation of Gaussian envelope; Spatial aspect ratio controls the shape of the ellipse; After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained: ； : Indicates at a specific scale and specific directions Below is the texture response map obtained after Gabor filtering; in , , where represents the number of samples for scale and direction, respectively.

4. The anti-interference texture enhancement and localization method for low-texture visual scenes according to claim 3, characterized in that, ... Image partitioning Each zone corresponds to an adjustable light source intensity. Its goal is to maximize the local gradient contrast function: ； It is the average gradient intensity within the partition, used to evaluate the visual information richness of the region; It is the gradient of the image at point (x,y), reflecting the local changes at that point; Local contrast is calculated using gradient operators such as Sobel. And define the illumination adjustment function: ； In the formula: This represents the brightness of the light source at the current moment, i.e., the [number]th [time]. Each image partition in the number of iterations LED light source brightness value at that time; This indicates the brightness of the light source at the next moment, i.e., the brightness of the light source at the next moment. Each image partition in the number of iterations The updated LED light source brightness value; Represents the local gradient contrast function, i.e., the th The average value of the image gradient within each partition reflects the texture richness of that region. This represents the partial derivative of contrast with respect to brightness, indicating the degree to which changes in the brightness of the current partition affect the contrast, and is used for gradient ascent optimization.

5. The anti-interference texture enhancement and localization method for low-texture visual scenes according to claim 4, characterized in that, The feature fusion process involves combining multiple texture response maps with the lighting-optimized avatar, specifically including: Calculate the maximum response fusion for all texture response maps: ； Indicates the location At this location, the response value with the largest absolute value in all multi-scale, multi-directional Gabor texture response maps; Alternatively, a weighted fusion approach may be adopted: ； Indicates the location At this point, the composite texture response value is obtained by weighted fusion of all Gabor response maps; Among them, weight Dynamically adjust based on the mean response, directional prior, or feature feedback; fused images The texture response-enhanced image is obtained by overlaying it with the original image: ； This represents the fused texture response map. or One of them; This represents the final enhanced image, derived from the original image. and texture response map It is made by blending in a certain proportion; in Control the fusion intensity.

6. The anti-interference texture enhancement and localization method for low-texture visual scenes according to claim 5, characterized in that, Extracting feature points from the texture response-enhanced image specifically includes: using the SIFT operator for interest point detection, with the response function being: ； in For Gaussian kernel, As a scale, Scale ratio; Apply the operation to the response enhancement image Increase the number of interest point detections The lift rate is defined as: ； In the formula: This represents the feature point count improvement rate, which is the relative increase in the number of feature points in the enhanced image compared to the original image. This represents the number of feature points detected in the enhanced image, using operators such as SIFT to enhance the texture response of the image. The result is obtained through feature extraction; This represents the number of feature points detected in the original image, obtained by feature extraction from the unenhanced original image.

7. An anti-interference texture enhancement positioning system suitable for low-texture visual scenes, characterized in that, The system includes a memory and a processor. The memory includes a program for an anti-interference texture enhancement and localization method suitable for low-texture visual scenes. When the program for the anti-interference texture enhancement and localization method suitable for low-texture visual scenes is executed by the processor, it performs the following steps: Obtain the original image, preprocess the original image to obtain the preprocessed image; Based on multi-scale Gabor filtering, texture response of preprocessed images under different directions and frequency dimensions is extracted to obtain multiple texture response maps; The preprocessed image is divided into multiple image partitions based on the partitioned LED array, and the local gradient contrast of each image partition is calculated. Based on local gradient contrast, the brightness of the light source is adjusted independently for different image regions to obtain an optimized illumination image; Multiple texture response maps are fused with the lighting-optimized avatar to obtain a texture response enhancement image, and feature points are extracted from the texture response enhancement image for visual localization.

8. The anti-interference texture enhancement positioning system for low-texture visual scenes according to claim 7, characterized in that, Obtain the original image, perform preprocessing on the original image to obtain the preprocessed image, specifically including: Obtain the original image, normalize the original image, and map the pixel grayscale values ​​of all original images to a fixed range. Isolated noise points are eliminated using median filtering, and Gaussian noise is eliminated using Gaussian filtering. The image is divided into multiple non-overlapping or partially overlapping local blocks, and a gray-level histogram is calculated for each local block. The preprocessed image is obtained by improving local contrast through histogram equalization.

9. The anti-interference texture enhancement positioning system for low-texture visual scenes according to claim 8, characterized in that, Based on multi-scale Gabor filtering, texture responses of preprocessed images are extracted under different directions and frequency dimensions to obtain multiple texture response maps, specifically including: Let the original image be A two-dimensional Gabor wavelet filter is applied to it. The Gabor kernel function is defined as follows: ； in: ； ； In the formula: The original image coordinate system Rotate counterclockwise around the origin The new x-coordinate after the angle; It is the new ordinate after the same rotation; Wavelength, controlling frequency; Direction: Controls the direction of the filter's main axis; Phase shift; : Standard deviation of Gaussian envelope; Spatial aspect ratio controls the shape of the ellipse; After performing a multi-scale, multi-directional Gabor transform on the image, a set of texture response maps is obtained: ； in , , where represents the number of samples for scale and direction, respectively.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a method program for anti-interference texture enhancement and localization suitable for low-texture visual scenes. When the method program is executed by a processor, it implements the steps of the method program for anti-interference texture enhancement and localization suitable for low-texture visual scenes as described in any one of claims 1 to 6.

Images