An image edge detection method based on biomimetic retina multi-scale receptive field iteration

Abstract

The application discloses a kind of image edge detection methods of biological retina multiscale receptive field iteration in the field of image processing and computer vision, including obtaining image to be handled, and pre-processing image to obtain input image.The application is by with center point belonging to an edge edge point in local slice and center point is classified as first edge, the rest is classified as second edge;Subsequently, with the tangent of center point, local slice is divided, and in the partition, from center point outwardly emit ray, if the direction of ray exists second edge point, then the average value of image value in the region within second edge is assigned to the region outside second edge, form compensation slice, the compensation operation is corrected to the pixel distribution in the detection process in adjacent second edge direction, so that the response of center point when subsequent second derivative is reconvolved is weakened, the positioning stability under adjacent interference is improved, positioning deviation caused by adjacent edge interference, multiscale result inconsistency and other problems are avoided.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, specifically a biomimetic retinal multi-scale receptive field iterative image edge detection method. Background Technology

[0002] Edge detection is a fundamental problem in image processing, widely used in computer vision, pattern recognition, and other fields. Current technologies primarily employ classical second-derivative edge detection theory, multi-scale LoG operators, Laplacian operator edge detection methods, and lateral suppression mechanisms. Classical second-derivative edge detection theory, inspired by image detection and physiology, uses center-periphery type second-derivative operators (such as the LoG operator) to filter the image at different scales, representing edges by identifying zero-crossing line segments in the filter response. Multi-scale LoG and Laplacian operator edge detection methods propose running Laplacian and LoG-type edge detectors at different scales to achieve multi-scale detection. The specific detection process of this method is as follows: first, smoothing the image and calculating the Laplacian or LoG operator edge responses at multiple scales; then, obtaining edges at each scale and discussing the localization accuracy and efficiency. Lateral suppression mechanism edge detection methods utilize the suppression interaction between neighborhoods to enhance the contrast of signal change areas, thereby achieving edge detection or edge enhancement.

[0003] However, existing edge detection methods suffer from problems such as positioning offset and inconsistent results at multiple scales due to interference from nearby edges in complex scenarios, resulting in poor edge detection performance. Summary of the Invention

[0004] The purpose of this invention is to provide a biomimetic retinal multi-scale receptive field iterative image edge detection method to solve the problems mentioned above, such as positioning offset and inconsistent multi-scale results caused by interference from adjacent edges in complex scenes, resulting in poor edge detection performance.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A biomimetic retinal multi-scale receptive field iterative image edge detection method includes: S1: Obtain the image to be processed and preprocess the image to obtain the input image; S2: Set a set of receptive field scales for the input image, which includes multiple receptive field scales, and construct a two-dimensional Gaussian distribution second-order derivative convolution kernel for each receptive field scale in the set of receptive field scales; S3: Convolve the input image using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain an initial convolution response map and extract initial edge points; S4: Extract local slices for each initial edge point at each receptive field scale, calculate the local slope of the local slices, and divide the first edge and the second edge according to the local slope; S5: Divide the local slice into a first sub-region and a second sub-region according to the local slope, and emit rays to determine the interference edge. Obtain the compensation slice based on the interference edge. S6: The compensation slice is reconvolved using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain the updated response value at the center point. The updated response value is then used to replace the response value at the center point of the initial convolution response map to obtain the updated response map. S7: Repeat steps S4 to S6 for iteration until the updated response map reaches the convergence condition, and obtain stable response maps at each scale; S8: Extract edge results from the stable response map and combine them to obtain the final edge detection result.

[0006] As a further aspect of the present invention: in step S1, the image to be processed includes a grayscale image and a color image; The RGB three-channel components of the color image are converted into single-channel grayscale values, and then fused into a converted grayscale image. The grayscale image and the converted grayscale image are normalized and have their boundaries filled to obtain the input image.

[0007] As a further aspect of the present invention: in step S2, the input image is set to include multiple receptive field scales. receptive field scale set ; For receptive field scale set Each receptive field scale Construct a two-dimensional Gaussian distributed second-order derivative convolution kernel The two-dimensional Gaussian distributed second-order derivative convolution kernel A two-dimensional Gaussian probability density function radial The second derivative; ; ; in, To experience the scale of the wild A two-dimensional Gaussian distribution second-order derivative convolution kernel, y is the horizontal variable of the two-dimensional coordinates in the input image, and y is the vertical variable of the two-dimensional coordinates in the input image.

[0008] As a further aspect of the present invention: In step S3, a two-dimensional Gaussian distributed second-order derivative convolution kernel is used to process each receptive field scale in the input image. Perform convolution to obtain the initial convolutional response map; ; in, This is the initial convolutional response map. Input image; Initial edge points are extracted from the initial convolutional response map based on the positive and negative zero-crossing positions, and the remaining pixels of the input image are classified as non-edge points, forming an initial edge point set. .

[0009] As a further aspect of the present invention: in step S4, for each initial edge point at each receptive field scale, a local slice containing several edge points is extracted with the initial edge point as the center point and a preset radius. Within a local slice of the neighborhood of the center point, select several adjacent edge points, calculate the local slope of the edge where the center point is located, and calculate the edge slope between the edge point and its adjacent edge points. Based on the proximity threshold and connectivity constraints, if the absolute value of the difference between the edge slope and the local slope is less than or equal to the proximity threshold and a cluster of edge points connected to the center point is classified as the first edge, otherwise it is classified as the second edge.

[0010] As a further aspect of the present invention: the preset radius is the same as the current receptive field scale value.

[0011] As a further aspect of the present invention: based on the local slope, a tangent line is drawn through the center point, dividing the local slice into a first sub-region and a second sub-region of equal area. Multiple rays are emitted from the center point outwards from the first sub-region and the second sub-region at preset angle intervals. If there is an edge point of the second edge in the ray direction, i.e., the intersection of the second edge, the second edge intersection point closest to the center point on the ray is determined. The average pixel value of the region between the center point and the second edge intersection point is calculated, and the average pixel value is assigned to the local slice pixels corresponding to the region outside the second edge intersection point of the ray. If there is no edge point of the second edge in the ray direction, the pixels in the ray direction of the local slice remain unchanged. After processing each ray direction, a compensated slice is obtained.

[0012] As a further aspect of the present invention: In step S6, the compensation slice is processed using a scale similar to that of the receptive field. The corresponding two-dimensional Gaussian distribution second-order derivative convolution kernel is used to recalculate the convolution response, obtain the updated response value at the center point, and replace the response value at the center point of the initial convolution response map with the updated response value to obtain the updated response map.

[0013] As a further aspect of the present invention: in step S7, the preset convergence includes the number of iterations reaching a preset number of iterations or a preset convergence condition. The preset convergence condition is that the change in the update response value of the edge point is less than a threshold, or the proportion of edge points that meet the change threshold reaches a preset value.

[0014] As a further aspect of the present invention: in step S8, positive and negative zero-crossing positions are extracted from the stable response map at each scale to obtain the edge results at each scale, and the edge results obtained at different receptive field scales are combined to obtain the final edge detection result.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention extracts local slices at initial edge points at various scales. First, it calculates the local slope based on the edge points in the neighborhood of the center point. Using a "similar direction threshold + connectivity constraint," edge points within the local slice that belong to the same edge as the center point are classified as first edges, and the rest as second edges. Then, the local slice is divided into partitions by the tangent of the center point, and rays are emitted outward from the center point within each partition. If there are second edge points in the ray direction, the average image value of the region within the second edge is assigned to the region outside the second edge, forming a compensation slice. This compensation operation corrects the pixel distribution in the direction of the adjacent second edge during the detection process, so that the response of the center point during the subsequent second-order derivative reconvolution is less affected by the superposition of the second edge. Through the process suppression mechanism of "first edge / second edge differentiation + tangent partition ray compensation," the positioning stability under neighboring interference is improved, and problems such as positioning offset and inconsistent results at multiple scales due to neighboring edge interference are avoided.

[0016] 2. This invention performs a two-dimensional Gaussian second-order derivative convolution on the compensated slice after local compensation, and replaces the original response value with the updated center point convolution response value. This process is repeated iteratively until a preset number of iterations is reached or the convergence condition is met. This "compensation-reconvolution-replacement update" iterative process changes the edge response from a one-time judgment to a gradual correction, which helps to reduce the impact of the initial response error on the final zero-crossing output. Iterative reconvolution replacement update reduces the amplification effect of a one-time judgment on the final result, further improving the accuracy and robustness of edge localization. This allows the response value of the edge point to gradually approach its true edge features, effectively alleviating the overshoot or undershoot problems that may be caused by a single convolution. Even when facing complex backgrounds, noise interference, and edge blurring, it can still maintain stable and accurate edge detection performance, resulting in good edge detection effect. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A natural image is provided for Embodiment 1 of the present invention; Figure 3 The present invention provides a heatmap of a two-dimensional Gaussian second-order guided convolution kernel according to Embodiment 1; Figure 4 This invention provides cross-sectional views of Gaussian second-order guided convolution kernels under different parameters in Embodiment 1. Figure 5This invention provides response maps of natural images after performing Gaussian second-order convolutions of different sizes, as shown in Embodiment 1. Figure 6 The present invention provides response maps of natural images after performing Gaussian second-order convolutions of different sizes in Embodiment 1. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-4 This embodiment provides a biomimetic retinal multi-scale receptive field iteration image edge detection method, including: S1: Obtain the image to be processed and preprocess the image to obtain the input image; S2: Set a set of receptive field scales for the input image, which includes multiple receptive field scales, and construct a two-dimensional Gaussian distribution second-order derivative convolution kernel for each receptive field scale in the set of receptive field scales; S3: Convolve the input image using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain an initial convolution response map and extract initial edge points; S4: Extract local slices for each initial edge point at each receptive field scale, calculate the local slope of the local slices, and divide the first edge and the second edge according to the local slope; S5: Divide the local slice into a first sub-region and a second sub-region according to the local slope, and emit rays to determine the interference edge. Obtain the compensation slice based on the interference edge. S6: The compensation slice is reconvolved using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain the updated response value at the center point. The updated response value is then used to replace the response value at the center point of the initial convolution response map to obtain the updated response map. S7: Repeat steps S4 to S6 for iteration until the updated response map reaches the convergence condition, and obtain stable response maps at each scale; S8: Extract edge results from the stable response map and combine them to obtain the final edge detection result.

[0020] Specifically, this invention completes edge detection based on the second-order derivative convolution response of a two-dimensional Gaussian distribution and a deterministic iterative compensation update rule. The detection process does not rely on neural network parameter learning, does not require the construction of a large-scale labeled dataset, and does not require model training, repeated parameter tuning, or training computing power investment. It also eliminates the need for training data and training process, thus reducing implementation and deployment costs.

[0021] This invention extracts slices, or local slices, at the initial edge points at each scale. First, the local slope is calculated based on the edge points in the neighborhood of the center point. Using a "similar direction threshold + connectivity constraint," edge points within the local slice that belong to the same edge as the center point are classified as first edges, and the rest as second edges. Then, the local slice is divided into partitions by the tangent of the center point, and rays are emitted outward from the center point within each partition. If a second edge point exists in the ray direction, the average image value of the region within the second edge is assigned to the region outside the second edge, forming a compensation slice. This compensation operation corrects the pixel distribution in the direction of the adjacent second edge during the detection process, so that the response of the center point during the subsequent second-order derivative reconvolution is less affected by the superposition of the second edge. Through the process suppression mechanism of "first edge / second edge differentiation + tangent partition ray compensation," the positioning stability under neighboring interference is improved, avoiding problems such as positioning offset and inconsistent results at multiple scales caused by neighboring edge interference.

[0022] This invention performs a second-order Gaussian convolution on the compensated slice after local compensation, and replaces the original response value with the updated center point convolution response value. This process is repeated iteratively until a preset number of iterations is reached or the convergence condition is met. This iterative process of "compensation-reconvolution-replacement update" transforms the edge response from a one-time judgment to a gradual correction, which helps to reduce the impact of the initial response error on the final zero-crossing output. The iterative reconvolution replacement update reduces the amplification effect of the one-time judgment on the final result, further improving the accuracy and robustness of edge localization. This allows the response value of the edge point to gradually approach its true edge features, effectively alleviating the overshoot or undershoot problems that may be caused by a single convolution. Even when facing complex backgrounds, noise interference, and edge blurring, it can still maintain stable and accurate edge detection performance, resulting in good edge detection effect.

[0023] This invention includes radial quadratic convolution of a multi-scale two-dimensional Gaussian probability density function, local slope / tangent direction estimation and connectivity criteria, tangent partitioning and ray direction compensation assignment rules, compensation-based reconvolution and response substitution update, and iteration stopping conditions. The model is interpretable, has a biomimetic theoretical basis, and is easy to analyze and control.

[0024] This invention performs initial detection and iterative compensation at different receptive field scales to obtain edge results at each scale, and can perform cross-scale combination in the output stage: the large receptive field result is used as the main object contour edge, and the small receptive field result is used as the detail edge, which makes it convenient for subsequent tasks to select the main contour and detail information as needed. The main contour and detail edges are output in layers by multiple receptive fields, which is suitable for the multi-scale expression needs of complex images.

[0025] In this embodiment, in step S1, the image to be processed includes a grayscale image and a color image; The RGB three-channel components of a color image are converted into single-channel grayscale values, and then fused into a converted grayscale image. The specific conversion process is as follows: For each pixel in the color image, the values ​​of its R (red), G (green), and B (blue) channels are extracted. Then, the grayscale value Y corresponding to that pixel is calculated using the internationally accepted luminance formula Y = 0.299*R + 0.587*G + 0.114*B. By performing this operation on all pixels in the image, the original three-dimensional color image data is converted into a two-dimensional single-channel grayscale image. This provides a unified input format for subsequent multi-scale receptive field edge detection processing, ensuring that edge detection can be performed using a consistent process for different types of images (grayscale or color). Furthermore, the color image can be converted into a luminance component and then fused into a converted grayscale image to avoid interference from different color channels on edge detection, while retaining key structural information related to luminance in the image. When processing natural scene images containing rich colors, directly performing edge detection on the RGB channels separately may produce redundant or conflicting edge responses due to the different sensitivities of each channel to illumination and color. However, after converting to a grayscale image based on the luminance component, it can focus more on the light and dark boundaries formed on the object surface due to changes in illumination, thereby extracting more representative structural edges and laying a stable foundation for subsequent multi-scale receptive field iterative processing. The grayscale image and the converted grayscale image are normalized and have their boundaries filled to obtain the input image.

[0026] In this embodiment, in step S2, the input image is set to include multiple receptive field scales. receptive field scale set ,in, .

[0027] For receptive field scale set Each receptive field scale Construct a two-dimensional Gaussian distributed second-order derivative convolution kernel Two-dimensional Gaussian distribution second-order derivative convolution kernel A two-dimensional Gaussian probability density function radial The second derivative; ; ; in, To experience the scale of the wild A two-dimensional Gaussian distribution second-order derivative convolution kernel, Let y be the horizontal variable of the two-dimensional coordinates in the input image, and let y be the vertical variable of the two-dimensional coordinates in the input image. Where different... Corresponding to different sizes of receptive fields.

[0028] In this embodiment, in step S3, a two-dimensional Gaussian distributed second-order derivative convolution kernel is used to process each receptive field scale in the input image. Perform convolution to obtain the initial convolutional response map; ; in, This is the initial convolutional response map. Input image; Initial edge points are extracted from the initial convolutional response map based on the positive and negative zero-crossing positions, and the remaining pixels of the input image are classified as non-edge points, forming an initial edge point set. .

[0029] In this embodiment, in step S4, for each initial edge point at each receptive field scale, a local slice containing several edge points is extracted with the initial edge point as the center point and a preset radius. Within a local slice of the neighborhood of the center point, select several adjacent edge points, calculate the local slope of the edge where the center point is located, and calculate the edge slope between the edge point and its adjacent edge points. Based on the proximity threshold and connectivity constraints, if the absolute value of the difference between the edge slope and the local slope is less than or equal to the proximity threshold and a cluster of edge points connected to the center point is classified as the first edge, otherwise it is classified as the second edge.

[0030] Among them, local slice extraction is distinguished from the first edge and the second edge. For each and each of its edge points (No. In the next iteration, the following processing is performed: Extracting a local slice from the center Window size is ,radius The numerical values ​​are the same as the receptive field values ​​currently used. An edge point set is collected within the slice. .

[0031] To distinguish between "the first edge that shares the same edge as the center point" and "the second edge that does not share the same edge as the center point," the following determination method is adopted: First, select several adjacent edge points in the neighborhood of the center point and estimate the local slope (or local tangent direction) of the edge where the center point is located. Then, using thresholds with similar directions. Using connectivity constraints (such as 8-neighbor connectivity or distance thresholds) as criteria, it will be compared with... A cluster of edge points that are similar in direction and connected to the center point is grouped into the first edge set. Finally, the remaining edge points within the slice that do not satisfy the conditions of similar orientation and connectivity are grouped into the second edge set. .

[0032] Specifically, the direction vector of the line connecting the center point and adjacent edge points is calculated, and the angle value of the edge slope is obtained through the arctangent function. The similarity threshold can be set to 5 to 15 degrees according to the actual image characteristics. When the difference between the slope angle of adjacent edge points and the local slope angle is within this threshold range, they are determined to be in the same direction. The connectivity constraint is achieved by judging whether adjacent edge points are within a preset neighborhood (such as a 3×3 pixel area). If both the similarity of direction and connectivity conditions are met, these edge points are classified as first edges, forming edge segments with continuous direction; edge points that do not meet the conditions are marked as second edges, which are usually isolated points or noise points with abrupt changes in direction. In this way, edge segments with structural significance can be effectively screened from the initial set of edge points, laying the foundation for subsequent iterative optimization.

[0033] In this embodiment, the preset radius is the same as the current receptive field scale value.

[0034] Specifically, for example, when the receptive field scale is 3, the preset radius is set to 3 pixels. In this case, the captured local slice is a square area centered on the initial edge point with a side length of 7 pixels (i.e., 2*3+1). When the receptive field scale is adjusted to 5, the preset radius becomes 5 pixels accordingly, and the side length of the local slice expands to 11 pixels (2*5+1). This design, which binds the preset radius to the receptive field scale value, ensures that the size of the local slice matches the perceptual range of the edge features at the current scale. This avoids the loss of information on adjacent edge points due to slices that are too small, and also prevents the introduction of too many irrelevant pixels that interfere with the accuracy of slope calculation due to slices that are too large.

[0035] In this embodiment, based on the local slope, a tangent line passing through the center point is drawn, dividing the local slice into a first sub-region of equal area. ) and the second sub-region ( Multiple rays are emitted from the center point outwards at preset angle intervals within the first and second sub-regions. If there is an edge point of the second edge in the ray direction, i.e., the intersection of the second edge, the second edge intersection point closest to the center point on the ray is determined. The average pixel value of the region between the center point and the second edge intersection point is calculated, and the average pixel value is assigned to the local slice pixel corresponding to the region outside the second edge intersection point of the ray. If there is no edge point of the second edge in the ray direction, the pixel value in the ray direction of the local slice remains unchanged. After processing each ray direction, a compensated slice is obtained.

[0036] Among them, in slices Inside, based on the local tangent direction of the center point obtained in step 4 Draw a tangent line through the center point The slice is divided into two sub-regions of equal area. and Then, ray search compensation is performed in each sub-region: ① From the center point At preset angle intervals on the outer perimeter Emit multiple rays (The rays fell on) and (inner); ② for each ray Check along the ray direction to see if there is a second edge point, i.e., determine Is it empty? If not empty, determine the second edge intersection point on the ray that is closest to the center point. ③ Calculate the distance from the center point to the ray. The average image value of the region between (within the second edge) ④ The average value Assigning a value to the ray The slice pixels corresponding to the area outside the second edge are used to form the compensated slice. When there is no second edge point on a ray, the pixel values ​​in that ray direction remain unchanged.

[0037] The purpose of the above compensation is to replace the pixel values ​​of the region outside the second edge with the statistical values ​​of the region inside the second edge in the direction where the second edge exists, thereby weakening the influence of the second edge on the local convolution response of the center edge point.

[0038] In this embodiment, in step S6, the compensation slice is processed using the receptive field scale. The corresponding two-dimensional Gaussian distribution second-order derivative convolution kernel is used to recalculate the convolution response, obtain the updated response value at the center point, and replace the response value at the center point of the initial convolution response map with the updated response value to obtain the updated response map. The preset convergence includes the number of iterations reaching the preset number of iterations or the preset convergence condition. The preset convergence condition is that the change in the updated response value of the edge point is less than a threshold, or the proportion of edge points that meet the change threshold reaches a preset value.

[0039] Among them, the compensated slices Adoption and Scale The corresponding two-dimensional Gaussian second-order derivative convolution kernel Recalculate the convolution response to obtain the updated response value at the center point. .use Replace the original response diagram At point The response value at the location is used to obtain the updated response graph. The alternative update is preferably performed on edge points on a point-by-point basis; non-edge points may remain unchanged, or in alternative implementations, the update may be extended according to the same rules.

[0040] Iterative execution and conditions for convergence or termination. For each scale Repeat steps 4 through 6 for a total of Second-rate, The number of iterations should be 10-30, or a convergence criterion should be used to stop the process early, such as the change in response at edge points. Alternatively, the proportion of all edge points satisfying the change threshold reaches a preset value. After iteration, a stable response map at this scale is obtained. .

[0041] In this embodiment, in step S8, positive and negative zero-crossing positions are extracted from the stable response map at each scale to obtain the edge results at each scale, and the edge results obtained at different receptive field scales are combined to obtain the final edge detection result.

[0042] Among them, the stable response diagrams at various scales Extract the positive and negative zero-crossing positions to obtain the edge result at this scale. The edge results at different receptive field scales are then combined: the edge results corresponding to the large receptive field scale are used as the main object contour edges in the image, and the edge results corresponding to the small receptive field scale are used as detail edges. The combination method can be layered overlay, priority coverage, or scale-weighted fusion, and deduplication, thinning, and connectivity filtering can be performed during the fusion process.

[0043] Example 1: Taking a natural scene image as the input image, this paper presents a biomimetic retinal multi-scale receptive field iterative image edge detection method, as follows: Input as follows Figure 2 An image of a natural scene shown Setting up multi-receptive field scale sets ,in ,different Representing different receptive field sizes; the two-dimensional Gaussian probability density function is: The "Gaussian second-order derivative convolution kernel" used in this invention is a two-dimensional Gaussian probability density function along the radial direction. Second derivative: When discretizing the convolution kernel, the kernel size can be determined according to... The heatmap of the two-dimensional Gaussian second-order derivative convolution kernel after setting a fixed kernel size is as follows: Figure 3 As shown, the cross-sectional diagrams of the Gaussian second-order derivative convolution kernel under different parameters are as follows. Figure 4 As shown, the input image is padded (mirrored or copied) to ensure complete convolution calculation. The radius of the local slice window is further set. Consistent with the current receptive field (i.e., kernel size), the number of iterations... (Preferred) ); and optional response amplitude threshold. Used to suppress zero-crossing of weak responses.

[0044] For each receptive field scale The input image is convolved with the corresponding convolution kernel to obtain the initial second-order derivative response map, such as... Figure 5 As shown: exist The zero-crossing position of the sign change is used as the initial edge position; to reduce zero-crossing points caused by noise and weak textures, the zero-crossing point can be adjusted accordingly. Set threshold The selection process is performed. This yields the initial set of edge points at this scale. These points are then marked as edge points, while the remaining pixels are designated as non-edge points. Subsequent steps are performed independently at each scale to obtain the edge results for each scale.

[0045] For any scale any initial edge point (No. The iteration, initially... ),by Extracting a local slice from the center This yields the set of inner edge points of the slice. To distinguish between the "first edge" (which shares the same edge as the center point) and the "second edge" (which does not), several adjacent edge points are selected near the center point to estimate the local slope of the edge containing the center point. Then, with the statement "the directional difference is less than..." Furthermore, based on the criterion of "connected to the center point," a set of edge points that are close to and connected to the center point in the same direction is designated as the first edge set. The remaining edge points are divided into the second edge set. .

[0046] Based on the local direction of the center point Draw a tangent line through the center point slice Divided into two sub-regions of equal area and Within each sub-region, increment by angular steps outward from the center point. Emit multiple rays For each ray, check along the ray direction for the existence of a second edge point (i.e., ...). If it exists, then take the second edge intersection point that is closest to the center point. Calculate the center point along this ray. Average pixel value within the range (inside the second edge) and will Assign to The pixel values ​​corresponding to the ray region outside (outside the second edge) are used to obtain the compensated slice. If a ray fails to detect a second edge point, the pixel values ​​along that ray remain unchanged. This compensation corrects the pixel distribution along the direction of the second edge interference, thereby reducing the impact of the second edge on the subsequent convolution response of the center point.

[0047] For the compensated slices Adoption and Scale Same convolution kernel Reconvolve and calculate the updated response value at the center point. . Place the response graph at the point The original response value at that point is replaced with the updated response value, that is: The above "compensation-reconvolution-alternation update" process is performed sequentially on all edge points at this scale.

[0048] Repeat steps (2) to (4) for a total of Next, the stable response diagram was obtained. In an alternative approach, a convergence criterion can be used to terminate the iteration early, for example, when a majority of edge points satisfy a convergence criterion. Stop when the time comes.

[0049] The stable response zero-crossing output is combined with cross-scale results.

[0050] Stability response diagrams at various scales Extract the positive and negative zero-crossing positions to obtain the edge result at this scale. ,like Figure 6 As shown. Finally, the edge results at different receptive field scales are combined: the results at the large receptive field scale (larger) are combined. The corresponding edge is used as the main object outline edge, and the small receptive field scale (smaller) is used as the edge. The corresponding edges are displayed as detailed edges. Layered overlay or priority overlay methods can be used to obtain a comprehensive edge output that includes both the main outline and retained details.

[0051] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A biomimetic retinal multi-scale receptive field iterative image edge detection method, characterized in that, include: S1: Obtain the image to be processed and preprocess the image to obtain the input image; S2: Set a set of receptive field scales for the input image, which includes multiple receptive field scales, and construct a two-dimensional Gaussian distribution second-order derivative convolution kernel for each receptive field scale in the set of receptive field scales; S3: Convolve the input image using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain an initial convolution response map and extract initial edge points; S4: Extract local slices for each initial edge point at each receptive field scale, calculate the local slope of the local slices, and divide the first edge and the second edge according to the local slope; S5: Divide the local slice into a first sub-region and a second sub-region according to the local slope, and emit rays to determine the interference edge. Obtain the compensation slice based on the interference edge. S6: The compensation slice is reconvolved using a two-dimensional Gaussian distribution second-order derivative convolution kernel to obtain the updated response value at the center point. The updated response value is then used to replace the response value at the center point of the initial convolution response map to obtain the updated response map. S7: Repeat steps S4 to S6 for iteration until the updated response map reaches the convergence condition, and obtain stable response maps at each scale; S8: Extract edge results from the stable response map and combine them to obtain the final edge detection result.

2. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 1, characterized in that, In step S1, the image to be processed includes a grayscale image and a color image; The RGB three-channel components of the color image are converted into single-channel grayscale values, and then fused into a converted grayscale image. The grayscale image and the converted grayscale image are normalized and have their boundaries filled to obtain the input image.

3. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 2, characterized in that, In step S2, the input image is set to include multiple receptive field scales. receptive field scale set ; For receptive field scale set Each receptive field scale Construct a two-dimensional Gaussian distributed second-order derivative convolution kernel The two-dimensional Gaussian distributed second-order derivative convolution kernel A two-dimensional Gaussian probability density function radial The second derivative; ； ； in, To experience the scale of the wild A two-dimensional Gaussian distribution second-order derivative convolution kernel, y is the horizontal variable of the two-dimensional coordinates in the input image, and y is the vertical variable of the two-dimensional coordinates in the input image.

4. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 3, characterized in that, In step S3, a two-dimensional Gaussian distributed second-order derivative convolution kernel is used to process each receptive field scale in the input image. Perform convolution to obtain the initial convolutional response map; ； in, This is the initial convolutional response map. Input image; Initial edge points are extracted from the initial convolutional response map based on the positive and negative zero-crossing positions, and the remaining pixels of the input image are divided into non-edge points, forming an initial edge point set.

5. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 4, characterized in that, In step S4, for each initial edge point at each receptive field scale, a local slice containing several edge points is extracted with the initial edge point as the center point and a preset radius. Within a local slice of the neighborhood of the center point, select several adjacent edge points, calculate the local slope of the edge where the center point is located, and calculate the edge slope between the edge point and its adjacent edge points. Based on the proximity threshold and connectivity constraints, if the absolute value of the difference between the edge slope and the local slope is less than or equal to the proximity threshold and a cluster of edge points connected to the center point is classified as the first edge, otherwise it is classified as the second edge.

6. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 5, characterized in that, The preset radius is the same as the current receptive field scale value.

7. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 6, characterized in that, In step S5, based on the local slope, a tangent line is drawn through the center point to divide the local slice into a first sub-region and a second sub-region with equal areas. Multiple rays are emitted from the center point outwards from the first sub-region and the second sub-region at preset angle intervals. If there is an edge point of the second edge in the ray direction, i.e., the intersection of the second edge, the second edge intersection point closest to the center point on the ray is determined. The average pixel value of the region between the center point and the second edge intersection point is calculated, and the average pixel value is assigned to the local slice pixels in the region outside the second edge intersection point of the ray. If there is no edge point of the second edge in the ray direction, the pixel values ​​in the ray direction of the local slice remain unchanged. After processing each ray direction, a compensated slice is obtained.

8. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 7, characterized in that, In step S6, the compensation slices are processed using the same scale as the receptive field. The corresponding two-dimensional Gaussian distribution second-order derivative convolution kernel is used to recalculate the convolution response, obtain the updated response value at the center point, and replace the response value at the center point of the initial convolution response map with the updated response value to obtain the updated response map.

9. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 8, characterized in that, In step S7, the preset convergence includes reaching a preset number of iterations or a preset convergence condition. The preset convergence condition is that the change in the update response value of the edge point is less than a threshold, or the proportion of edge points that meet the change threshold reaches a preset value.

10. The image edge detection method based on the biomimetic retinal multi-scale receptive field iteration according to claim 9, characterized in that, In step S8, positive and negative zero-crossing positions are extracted from the stable response map at each scale to obtain the edge results at each scale. The edge results obtained at different receptive field scales are then combined to obtain the final edge detection result.

Images