An image recognition-based stacked object boundary extraction method, device, and medium

Abstract

The application discloses a kind of based on image recognition's stacked object boundary extraction method, equipment and medium, it is related to image processing technical field, including, obtain the stacked object image to be handled, and stacked object image is input to double branch feature encoder, output appearance feature map and material invariability feature map;Appearance feature map and material invariability feature map are fused, generate fusion feature map, and fusion feature map is input to boundary prediction decoder, obtain the initial boundary probability graph corresponding to each pixel;Based on initial boundary probability graph, multiple forward reasoning is executed to stacked object image, generate multiple boundary probability samples, and based on multiple boundary probability samples, obtain the integrated boundary probability value and cognitive uncertainty value of each pixel, form integrated boundary probability graph and cognitive uncertainty graph.The application reaches the effect that the feature discrimination of stacked occlusion scene is improved and texture noise false detection is inhibited.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method, device and medium for extracting the boundaries of stacked objects based on image recognition. Background Technology

[0002] With the rapid development of image recognition and deep learning technologies, object boundary extraction, as a fundamental prerequisite task for image segmentation, scene understanding, and robot grasping, has been widely applied in fields such as industrial automation inspection, logistics sorting, and intelligent monitoring. Traditional boundary extraction methods mainly rely on hand-designed differential operators such as Sobel and Canny, which are difficult to handle noise interference in complex backgrounds. In recent years, fully convolutional networks based on convolutional neural networks and encoder-decoder architectures have become mainstream, significantly improving feature representation capabilities and boundary localization accuracy through end-to-end learning.

[0003] Existing technologies have limitations when dealing with complex stacked objects: First, feature fusion strategies lack differentiated modeling of appearance texture and material invariance features, resulting in decreased discrimination power for illumination changes or texture repetition areas; Second, deterministic reasoning ignores the cognitive uncertainty of occluded areas, and using fixed thresholds easily produces false contours and lacks credibility assessment, making it difficult to meet the requirements of high-precision operations for boundary reliability. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for extracting the boundary of stacked objects based on image recognition to solve the problems of insufficient discrimination power of boundary feature fusion of stacked objects and false detection of boundaries and unreliable results caused by ignoring cognitive uncertainty.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for extracting the boundary of a stacked object based on image recognition, comprising: acquiring an image of the stacked object to be processed, and inputting the image of the stacked object into a dual-branch feature encoder to output an appearance feature map and a material invariance feature map; fusing the appearance feature map and the material invariance feature map to generate a fused feature map, and inputting the fused feature map into a boundary prediction decoder to obtain an initial boundary probability map corresponding to each pixel; performing multiple forward inferences on the image of the stacked object based on the initial boundary probability map to generate multiple boundary probability samples, and obtaining the integrated boundary probability value and cognitive uncertainty value of each pixel based on the multiple boundary probability samples to form an integrated boundary probability map and a cognitive uncertainty map; performing adaptive thresholding on the integrated boundary probability map using the cognitive uncertainty map to generate a binary boundary mask; using the binary boundary mask to extract the boundary contour of the stacked object, and calculating the confidence score of the boundary contour based on the integrated boundary probability value and cognitive uncertainty value of the region corresponding to the boundary contour to obtain boundary information with confidence evaluation.

[0008] As a preferred embodiment of the image recognition-based method for extracting the boundaries of stacked objects according to the present invention, the specific steps for outputting the appearance feature map and the material invariance feature map are as follows:

[0009] The stacked object image to be processed is obtained, and the stacked object image is normalized and subjected to steady-state wavelet transform to obtain a high-frequency texture image carrying texture information and a low-frequency structure image carrying structural information.

[0010] The high-frequency texture image is input into the appearance feature encoding branch, and the edge information is enhanced by a high-frequency attention mechanism to generate the initial appearance feature map.

[0011] The low-frequency structure image is decomposed to obtain the reflectance component map, and the reflectance component map is input into the material invariance coding branch to generate the initial material feature map;

[0012] The initial appearance feature map and the initial material feature map are orthogonalized to generate an appearance feature map and a material invariance feature map.

[0013] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for generating the fused feature map are as follows:

[0014] The channel-level local variance values ​​of the appearance feature map and the material invariance feature map are calculated separately to obtain their respective local variance information;

[0015] Based on local variance information, pixel-level fusion weight coefficients are calculated using a normalized exponential function.

[0016] The appearance feature map and the material invariance feature map are combined at the pixel level using pixel-level fusion weight coefficients and then convolutional smoothing is performed to generate a fused feature map.

[0017] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for obtaining the initial boundary probability map corresponding to each pixel are as follows:

[0018] The fused feature map is input into the boundary prediction decoder and the resolution is restored by transposed convolution to obtain a high-resolution feature map;

[0019] Gradient components are obtained by applying a fixed convolution kernel to the high-resolution feature map, and gradient feature information is obtained.

[0020] Gradient feature information is used to modulate the high-resolution feature map to generate an edge-enhanced feature map.

[0021] The edge enhancement feature map is processed using channel compression convolution and activation functions to obtain the initial boundary probability map corresponding to each pixel.

[0022] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for generating multiple boundary probability samples are as follows:

[0023] The initial boundary probability map is mapped to the initial memory state matrix, and the inference iteration counter and dynamic noise generator are initialized at the same time.

[0024] The spatial correlation noise field of the current iteration step is constructed based on the dynamic noise generator, and the boundary probability sample value of the current iteration step is calculated by combining the initial memory state matrix.

[0025] The boundary probability sample values ​​of the current iteration step are stored in the sample sequence set, and the inference iteration counter is incremented until the inference iteration counter reaches the preset number of inferences. The resulting sample sequence set is then used as multiple boundary probability samples.

[0026] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for forming the integrated boundary probability map and the cognitive uncertainty map are as follows:

[0027] Multiple boundary probability samples are aligned according to pixel position, and geometric transport aggregation is performed in the local neighborhood corresponding to each pixel to obtain the local integrated boundary distribution corresponding to each pixel.

[0028] Based on the local ensemble boundary distribution corresponding to each pixel, the transport bias of each boundary probability sample is obtained, and topological consistency analysis is performed in combination with the connectivity, discontinuity and branch differences in the local neighborhood to obtain the comprehensive conflict amount corresponding to each pixel.

[0029] Based on the local integration boundary distribution and integrated conflict amount corresponding to each pixel, the boundary positive evidence value and non-boundary negative evidence value corresponding to each pixel are generated. Based on the boundary positive evidence value and non-boundary negative evidence value, the integration boundary probability value and cognitive uncertainty value of each pixel are obtained.

[0030] The integration boundary probability value and cognitive uncertainty value of each pixel are written back according to the original pixel position to form the integration boundary probability map and cognitive uncertainty map.

[0031] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for generating the binary boundary mask are as follows:

[0032] Based on the integrated boundary probability map and the cognitive uncertainty map, an adaptive threshold neighborhood is constructed for each pixel.

[0033] Based on the adaptive threshold neighborhood corresponding to each pixel, the adaptive threshold corresponding to each pixel is obtained by combining the integrated boundary probability value and the cognitive uncertainty value.

[0034] Based on the adaptive threshold corresponding to each pixel, the integrated boundary probability map is compared to obtain strong boundary pixels, strong non-boundary pixels, and boundary pixels to be determined.

[0035] The strong boundary pixels are used to perform connectivity filtering on the boundary pixels to be judged to obtain the target boundary pixels;

[0036] The strong boundary pixels and the target boundary pixels are assigned to the boundary class, and the strong non-boundary pixels and the boundary pixels to be determined are assigned to the non-boundary class to generate a binary boundary mask.

[0037] As a preferred embodiment of the image recognition-based stacked object boundary extraction method of the present invention, the specific steps for obtaining boundary information with confidence evaluation are as follows:

[0038] Extract the skeleton of the connected region from the binary boundary mask to generate a set of boundary contours;

[0039] The boundary contour set is mapped to the integrated boundary probability map and cognitive uncertainty map, and the probability and uncertainty feature sequences of the corresponding pixels are obtained;

[0040] Based on the probability and uncertainty feature sequence, the comprehensive confidence score of the boundary contour is calculated;

[0041] The boundary contour set is bound to the comprehensive confidence score to generate boundary information with confidence assessment.

[0042] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein the computer program, when executed by the processor, implements any step of the image recognition-based method for extracting the boundaries of stacked objects as described in the first aspect of the present invention.

[0043] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the image recognition-based method for extracting the boundaries of stacked objects as described in the first aspect of the present invention.

[0044] The beneficial effects of this invention are as follows: by using steady-state wavelet transform and reflectivity component decomposition in image recognition, orthogonal separation of appearance texture features and material invariance features is achieved, thereby improving feature discrimination in stacked occlusion scenarios and suppressing false detections due to texture noise; by using multiple forward inferences and dynamic noise field construction in image recognition, pixel-level integrated boundary probability and cognitive uncertainty value are generated synchronously, thereby eliminating false contours in blurred boundary areas and providing reliable confidence scores. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of a method for extracting the boundaries of stacked objects based on image recognition.

[0047] Figure 2 The flowchart for generating appearance feature maps and material invariance feature maps.

[0048] Figure 3 A flowchart for generating boundary information with confidence assessment.

[0049] Figure 4 This is a graph showing the effect of changes in illumination on the boundary F1 value.

[0050] Figure 5 Heatmap to mask fuzzy coupling of cognitive uncertainty.

[0051] Figure 6 This is a comparison chart of the overall means for each method. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] Reference Figures 1-6 This is one embodiment of the present invention, which provides a method for extracting the boundaries of stacked objects based on image recognition, including the following steps:

[0056] S1. Obtain the image of the stacked object to be processed, and input the image of the stacked object into the dual-branch feature encoder to output the appearance feature map and the material invariance feature map.

[0057] The stacked object image to be processed is obtained, and the stacked object image is normalized and subjected to steady-state wavelet transform to obtain a high-frequency texture image carrying texture information and a low-frequency structure image carrying structural information.

[0058] The specific process includes: acquiring two-dimensional digital images by capturing images of multiple objects in a real scene that are mutually occluded or stacked; recording the appearance information of the stacked objects in the form of a pixel matrix; normalizing the images of the stacked objects by scaling the pixel values ​​of the images to a uniform range; performing steady-state wavelet transform; and using the multi-scale decomposition characteristics of steady-state wavelet transform to separate the images of the stacked objects into high-frequency texture images containing details and edge information and low-frequency structural images containing overall shape and structural information.

[0059] It should be noted that steady-state wavelet transform refers to the translation-invariant wavelet transform method. By performing redundant sampling on the signal at each scale decomposition and keeping the sub-band resolution unchanged, it effectively avoids the shift sensitivity problem caused by downsampling in traditional discrete wavelet transform.

[0060] The high-frequency texture image is input into the appearance feature encoding branch, and the edge information is enhanced by a high-frequency attention mechanism to generate the initial appearance feature map.

[0061] The specific process includes inputting the high-frequency texture image into the appearance feature encoding branch, which extracts local features from the high-frequency texture image layer by layer through convolution operations. At the same time, a high-frequency attention mechanism is embedded in the feature extraction process. The high-frequency attention mechanism obtains attention weights based on the response intensity of pixels in the high-frequency domain and enhances the feature channels of the edge region, so that the texture details related to the object boundary have higher activation values ​​in the feature map, thereby outputting the initial appearance feature map.

[0062] It should be noted that the high-frequency attention mechanism is a method that adaptively assigns attention weights based on the local response intensity of the image in the high-frequency domain to enhance edge and texture detail features.

[0063] The low-frequency structure image is decomposed to obtain the reflectivity component map, and the reflectivity component map is input into the material invariance coding branch to generate the initial material feature map.

[0064] The specific process includes decomposing the low-frequency structural image based on the relationship between illumination and reflection, separating the reflectance component map related to the material properties of the object surface. This decomposition process is based on the principle that image brightness is jointly determined by illumination intensity and surface reflectance. After estimating or fixing the illumination components, the reflectance component map is extracted and input into the material invariance coding branch. The material invariance coding branch extracts feature representations related to the material category and unaffected by illumination changes through convolution operations, generating an initial material feature map.

[0065] The initial appearance feature map and the initial material feature map are orthogonalized to generate an appearance feature map and a material invariance feature map.

[0066] The specific process includes orthogonalizing the initial appearance feature map and the initial material feature map. The orthogonalization process is based on the principle of orthogonal projection in the feature space. The components in the initial appearance feature map that are in the same direction as the initial material feature map are removed, and the components in the initial material feature map that are in the same direction as the initial appearance feature map are also removed, so that the two feature maps after processing are orthogonal to each other in the vector space. This process retains the unique information related to texture and edge in the initial appearance feature map to form the appearance feature map, and at the same time retains the unique information related to the material properties of the object surface in the initial material feature map that is not affected by changes in lighting or viewing angle to form the material invariance feature map.

[0067] It should be noted that the dual-branch feature encoder refers to a feature extraction structure consisting of an appearance feature encoding branch and a material invariance encoding branch. The appearance feature encoding branch is used to process high-frequency texture images and combine a high-frequency attention mechanism to extract features related to object edges and surface details. The material invariance encoding branch is used to process reflectivity component maps to extract features related to the object's material properties and unaffected by changes in illumination. The appearance feature map reflects the visual performance of stacked objects under specific imaging conditions caused by texture, edges, and illumination, highlighting high-frequency details to support accurate boundary localization. The material invariance feature map describes the inherent material properties of the object's surface, remains stable under different illuminations, and is used to enhance the robustness of boundary discrimination to material changes.

[0068] like Figure 4 The graph shows the effect of illumination variation on the boundary F1 value. As the illumination variation continuously increases, experimental group E consistently maintains a high boundary F1 value, while control groups A through D all show varying degrees of performance degradation, with control group A showing the most significant decrease. This result demonstrates that the present invention, by performing steady-state wavelet transform on the stacked object image to generate high-frequency texture and low-frequency structure images, and further obtaining a material invariance feature map based on the reflectivity component map, combined with orthogonalization of the appearance feature map and the material invariance feature map, can effectively reduce the interference of illumination variation on boundary recognition results and improve feature discrimination and boundary extraction stability under complex imaging conditions. To verify the specific contribution of each technical step of the present invention to the final boundary extraction effect, the experimental method was divided into control groups A through D and experimental group E. Control group A uses a single-branch fixed-threshold method, extracting boundaries based on only a single path and outputting boundary results with a fixed threshold. Control group B uses a dual-branch processing structure but does not orthogonalize the appearance feature map and material invariance feature map, used to verify the impact of the orthogonalization separation step on texture noise suppression and feature discrimination. Control group C, after completing dual-branch feature extraction and orthogonalization processing, does not perform multiple forward inferences, used to verify the impact of the multiple boundary probability sample construction process on boundary stability. Control group D, after performing multiple forward inferences, does not use the cognitive uncertainty map to guide the subsequent boundary screening and optimization process, used to verify the role of the cognitive uncertainty map in false contour suppression and result reliability improvement. Experimental group E is the complete method of this invention. These grouping methods correspond to the item-by-item ablation of the dual-branch feature encoding, orthogonalization separation, multiple forward inferences, cognitive uncertainty map guidance, and confidence assessment link of this invention.

[0069] S2. The appearance feature map and the material invariance feature map are fused to generate a fused feature map, and the fused feature map is input into the boundary prediction decoder to obtain the initial boundary probability map corresponding to each pixel.

[0070] The channel-level local variance values ​​of the appearance feature map and the material invariance feature map are calculated separately to obtain their respective local variance information.

[0071] The specific process includes calculating the channel-level local variance values ​​of the appearance feature map and the material invariance feature map respectively. At each pixel location, a spatial neighborhood of a fixed size is selected with the pixel as the center. The local mean is first calculated for the feature values ​​at all spatial locations and in all channels within the spatial neighborhood. Then, the square of the difference between each feature value and the local mean is calculated. These squared differences are averaged over the spatial neighborhood and channel dimensions to obtain the local variance information corresponding to the appearance feature map and the local variance information corresponding to the material invariance feature map.

[0072] The expression for calculating the channel-level local variance of the appearance feature map and the material invariance feature map is as follows:

[0073] ;

[0074] ;

[0075] in, Represents the appearance feature map in spatial coordinates Local variance at the channel level; Represents spatial location coordinates; Indicates the side length of the local neighborhood window; Indicates the total number of channels in the feature map; Indicates the feature channel index variable; Indicates the convolution radius; Indicates the x-coordinate offset of the neighborhood; Indicates the offset of the ordinate of the neighborhood; Represents the coordinates of the appearance feature map in the local neighborhood space. First The eigenvalues ​​of the channel; Represents the appearance feature map in spatial coordinates Local mean at the channel level; Represents the material invariance characteristic map in spatial coordinates Local variance at the channel level; Represents the coordinates of the material invariance feature map in the local neighborhood space. First The eigenvalues ​​of the channel; Material invariance characteristic map in spatial coordinates Local mean at the channel level.

[0076] It should be noted that, Represents the appearance feature map in spatial coordinates The channel-level local mean at a given location is obtained by taking the arithmetic mean of the eigenvalues ​​of all locations and all channels within a fixed-size neighborhood centered on that spatial coordinate. Material invariance characteristic map in spatial coordinates The local mean at the channel level is a value obtained by calculating the arithmetic mean of the characteristic values ​​of all positions and all channels within a fixed-size neighborhood centered on the spatial coordinates.

[0077] Based on local variance information, pixel-level fusion weight coefficients are calculated using a normalized exponential function, expressed as follows:

[0078] ;

[0079] in, Represents spatial coordinates Pixel-level fusion weight coefficients at the location; Represents the natural exponential function; This represents a temperature coefficient used to adjust the sharpness of the pixel-level fusion weight coefficient distribution. This represents a smoothing constant to prevent the denominator from being zero.

[0080] It should be noted that before the expression is calculated, the local variance information corresponding to the appearance feature map and the local variance information corresponding to the material invariance feature map have been normalized to make them fall within the same numerical scale range, thereby achieving dimensional uniformity.

[0081] The specific process includes: based on the local variance information corresponding to the appearance feature map and the local variance information corresponding to the material invariance feature map, performing exponential operations on the two local variance values ​​at each pixel position, and using the exponential result corresponding to the appearance feature map, the sum of the two exponential results, and a very small constant to form a normalized denominator, and obtaining the pixel-level fusion weight coefficient for fusing the appearance feature map and the material invariance feature map at that pixel position in the form of a ratio.

[0082] The appearance feature map and the material invariance feature map are combined at the pixel level using pixel-level fusion weight coefficients and then convolutional smoothing is performed to generate a fused feature map.

[0083] The specific process includes: using pixel-level fusion weight coefficients to combine the appearance feature map and the material invariance feature map at the pixel level; at each spatial coordinate position, using the pixel-level fusion weight coefficient as the weight of the feature value at the corresponding position of the appearance feature map, and using the complement of the pixel-level fusion weight coefficient as the weight of the feature value at the corresponding position of the material invariance feature map; adding the two sets of weighted feature values ​​channel by channel to form a preliminary fusion result; and then applying a convolution smoothing operation to the preliminary fusion result to suppress local discontinuities and enhance spatial consistency to generate a fused feature map.

[0084] It should be noted that the fused feature map combines the texture edge information of the appearance feature map with the material semantic information of the material invariance feature map, thereby achieving complementary representation of variable appearance and stable material properties in the boundary discrimination of stacked objects, and improving the accuracy and robustness of boundary extraction.

[0085] The fused feature map is input into the boundary prediction decoder and the resolution is restored by transposed convolution to obtain a high-resolution feature map.

[0086] The specific process includes inputting the fused feature map into the boundary prediction decoder, which uses multi-layer transposed convolution operations to upsample the fused feature map step by step. At each level, the low-resolution feature map is expanded to a higher resolution space through a learnable convolution kernel, while gradually restoring spatial detail information. After several levels of transposed convolution, a high-resolution feature map with the same spatial size as the original stacked object image is output.

[0087] It should be noted that the boundary prediction decoder is a network structure used to recover boundary details from the fused feature map and output a high-resolution boundary representation. It consists of multiple cascaded transposed convolutional layers. The high-resolution feature map preserves fine spatial details aligned with the original stacked object image, providing pixel-level semantic and structural support for accurately extracting the boundaries of the stacked objects.

[0088] Gradient components are obtained by applying a fixed convolution kernel to the high-resolution feature map, and gradient feature information is obtained.

[0089] The specific process includes applying a fixed convolution kernel to the high-resolution feature map. The fixed convolution kernel uses a predefined gradient detection operator to perform convolution operations in the horizontal and vertical directions to obtain the rate of change of feature values ​​in the neighborhood of each pixel position, thereby obtaining the horizontal gradient component and the vertical gradient component. These two gradient components together constitute the gradient feature information.

[0090] It should be noted that the predefined gradient detection operator is determined based on commonly used edge detection operators in image processing, such as the Sobel operator or the Prewitt operator, and its convolution kernel weights are fixed before use.

[0091] Gradient feature information is used to modulate the high-resolution feature map to generate an edge-enhanced feature map.

[0092] The specific process includes using gradient feature information to modulate the high-resolution feature map. The gradient feature information includes horizontal gradient components and vertical gradient components. By obtaining the magnitude of the gradient feature information, the edge intensity response at each pixel location is obtained. The edge intensity response is used as a modulation factor and applied to the corresponding position of the high-resolution feature map. The feature values ​​of each channel in the high-resolution feature map are scaled element-wise according to the modulation factor, so that the feature activation in the strong gradient region, i.e., the potential boundary, is amplified, and the feature activation in the weak gradient region is weakened, thus generating an edge-enhanced feature map that highlights the boundary of the stacked object.

[0093] It should be noted that the edge enhancement feature map modulates the high-resolution feature map by fusing gradient feature information, which enhances the feature response of the boundary region of the stacked objects and effectively improves the localization accuracy and edge clarity of subsequent boundary prediction.

[0094] The edge enhancement feature map is processed using channel compression convolution and activation functions to obtain the initial boundary probability map corresponding to each pixel.

[0095] The specific process includes: using channel compression convolution for edge enhancement feature maps, which uses a single output channel kernel to weight and fuse all input channels, compressing the multi-channel edge enhancement feature map into a single-channel feature map, and inputting the single-channel feature map into an activation function, which converts the feature value of each pixel location into a value between zero and one through non-linear mapping. This value represents the confidence that the pixel belongs to the object boundary, thus obtaining the initial boundary probability map corresponding to each pixel.

[0096] It should be noted that the initial boundary probability map represents the confidence level of each pixel belonging to the boundary of the stacked object, providing a basic probability distribution for subsequent boundary refinement and effectively reflecting the rough location and continuity of potential boundary regions.

[0097] S3. Based on the initial boundary probability map, perform multiple forward inferences on the stacked object image to generate multiple boundary probability samples, and obtain the integrated boundary probability value and cognitive uncertainty value of each pixel based on the multiple boundary probability samples to form an integrated boundary probability map and a cognitive uncertainty map.

[0098] The initial boundary probability map is mapped to the initial memory state matrix, and the inference iteration counter and dynamic noise generator are initialized at the same time.

[0099] The specific process includes taking each pixel value in the initial boundary probability map as the initial state of the corresponding position in the initial memory state matrix. This mapping process maintains a one-to-one correspondence between spatial positions and keeps the numerical range unchanged, thereby transforming the boundary confidence distribution into a memory representation that can be used for iterative inference. At the same time, the inference iteration counter is set to the initial value of zero to record the number of subsequent iterations. The dynamic noise generator is initialized so that it is in a ready state to generate spatially correlated noise that conforms to preset statistical characteristics.

[0100] It should be noted that the inference iteration counter is an integer variable used to record the number of iterations performed in the current inference process, and is composed of incrementally updatable numerical storage units; the dynamic noise generator is a component used to generate spatially correlated random perturbation signals in each iteration, and is composed of a random number generation mechanism that follows a specific probability distribution and spatial filtering operations; the preset statistical characteristics are set according to task requirements and experience, and adopt a standard normal distribution or uniform distribution and generate noise patterns with local correlation through spatial low-pass filtering.

[0101] The spatial correlation noise field for the current iteration step is constructed based on a dynamic noise generator, and the boundary probability sample value for the current iteration step is calculated by combining the initial memory state matrix. The expression is as follows:

[0102] ;

[0103] in, Indicates the current iteration step In spatial coordinates Boundary probability sample values ​​at; Indicates the index of the current iteration step; Indicates the initial memory retention coefficient; The initial memory state matrix is ​​represented in coordinates. The initial boundary probability value of the mapping; Indicates the spatial coordinates of the previous iteration step. Boundary probability sample values ​​at; Indicates noise adjustment gain; Indicates the current iteration step In spatial coordinates The spatially correlated noise field value at that location.

[0104] It should be noted that the initial memory retention coefficient is a scalar parameter used to control the retention ratio of the initial memory state matrix during the iteration process, and it is a fixed value set through validation set tuning. The initial memory state matrix is ​​represented in coordinates. The initial boundary probability value of the mapping refers to the pixel value of the initial boundary probability map at that coordinate position. It is obtained by directly assigning values ​​to the initial boundary probability map and the initial memory state matrix according to their spatial positions. The noise adjustment gain is a scalar parameter used to control the intensity of the influence of the spatially correlated noise field on the boundary probability sample value. It is a fixed value set by evaluating the balance effect of different values ​​on the stability and accuracy of boundary prediction on the validation set.

[0105] The specific process includes generating a spatially correlated noise field for the current iteration step based on a dynamic noise generator. The spatially correlated noise field is obtained by sampling from a probability distribution defined by preset statistical characteristics and applying spatial filtering operations, and has a random perturbation structure with local inter-pixel correlation. The initial memory state matrix and the spatially correlated noise field are added element-wise at the same spatial location to make the superposition result contain random perturbations. The superposition result is then input into a nonlinear activation function for compression mapping so that the output value falls within the interval of zero to one, thereby calculating the boundary probability sample value of the current iteration step.

[0106] The boundary probability sample values ​​of the current iteration step are stored in the sample sequence set, and the inference iteration counter is incremented until the inference iteration counter reaches the preset number of inferences. The resulting sample sequence set is then used as multiple boundary probability samples.

[0107] The specific process includes saving the boundary probability sample values ​​of the current iteration step into the sample sequence set according to their spatial correspondence, incrementing the inference iteration counter by one, and continuing to the next iteration. This process repeats the operations of constructing a spatially correlated noise field by the dynamic noise generator, calculating new boundary probability sample values ​​by combining the initial memory state matrix, and storing the boundary probability sample values ​​into the sample sequence set. Meanwhile, the inference iteration counter is continuously incremented. When the value of the inference iteration counter reaches the preset number of inferences, the iteration process is terminated. At this point, the sample sequence set has accumulated and stored all the boundary probability sample values ​​generated from the first iteration to the preset number of inference iterations, serving as multiple boundary probability samples.

[0108] It should be noted that the preset number of inferences is a fixed value selected based on the complexity and computational efficiency requirements of the boundary prediction task. This value is chosen by evaluating the performance of different number of iterations on the validation set to ensure stable results and acceptable computational overhead.

[0109] Multiple boundary probability samples are aligned according to their pixel positions, and geometric transport aggregation is performed in the local neighborhood corresponding to each pixel to obtain the local integrated boundary distribution corresponding to each pixel.

[0110] The specific process includes aligning multiple boundary probability samples according to the same spatial coordinate position to ensure that the probability value of each sample at each pixel position corresponds to the same image region. At each pixel position, a local neighborhood of a fixed size is selected with that position as the center. The geometric transport aggregation method is applied to the probability values ​​from multiple boundary probability samples at all positions within the local neighborhood. By minimizing the geometric distance between distributions, the information of multiple samples is fused to obtain the local integrated boundary distribution corresponding to each pixel.

[0111] It should be noted that the geometric transport aggregation method is a multi-distribution fusion strategy based on optimal transport theory. It performs weighted aggregation by minimizing the Wasserstein distance between each boundary probability sample and the local integrated boundary distribution. The probability metric framework and spatial correlation constraints adopted are naturally derived.

[0112] Based on the local integrated boundary distribution corresponding to each pixel, the transport bias of each boundary probability sample is obtained, and topological consistency analysis is performed in combination with the connectivity, discontinuity and branch differences in the local neighborhood to obtain the comprehensive conflict amount corresponding to each pixel.

[0113] The specific process includes: based on the local ensemble boundary distribution corresponding to each pixel, measuring the difference between the probability value of each boundary probability sample at the pixel location and the probability value of the local ensemble boundary distribution at the same location, and using geometric transport distance to obtain this difference, forming the transport bias of the boundary probability sample at the pixel location; within the local neighborhood centered on the pixel, analyzing the connectivity of the boundary structure (whether adjacent boundary pixels form a continuous path), the discontinuity (whether there is a phenomenon of uninterrupted boundary being cut off), and the branching difference (whether there is an unreasonable forking or convergence structure), quantifying local topological conflict by evaluating the consistency of these topological attributes among multiple boundary probability samples; and weightedly fusing the transport bias with the conflict index obtained from the topological consistency analysis to generate a comprehensive conflict quantity reflecting the overall inconsistency of multi-source boundary predictions at the pixel location.

[0114] It should be noted that the comprehensive conflict quantity refers to the multi-hypothesis inconsistency measure quantified in the boundary prediction process, based on the transport deviation between the local integrated boundary distribution corresponding to each pixel and multiple boundary probability samples, and combined with the connectivity, discontinuity and branching differences within the local neighborhood.

[0115] Based on the local ensemble boundary distribution and the overall conflict amount corresponding to each pixel, the boundary positive evidence value and non-boundary negative evidence value corresponding to each pixel are generated. Based on the boundary positive evidence value and non-boundary negative evidence value, the ensemble boundary probability value and cognitive uncertainty value of each pixel are obtained.

[0116] The specific process includes: based on the local ensemble boundary distribution and the comprehensive conflict amount corresponding to each pixel, the value of the local ensemble boundary distribution is regarded as the positive support strength that the pixel belongs to the boundary, and the support strength is suppressed according to the comprehensive conflict amount to generate the boundary positive evidence value corresponding to each pixel; the complement value of the local ensemble boundary distribution to unit 1 is regarded as the negative support strength that the pixel does not belong to the boundary, and the negative support strength is also suppressed according to the comprehensive conflict amount to generate the non-boundary negative evidence value corresponding to each pixel; by substituting the boundary positive evidence value and the non-boundary negative evidence value into the evidence synthesis rule, the ensemble boundary probability value of each pixel is obtained, and the cognitive uncertainty value of each pixel is quantified based on the relative difference between the boundary positive evidence value and the non-boundary negative evidence value.

[0117] It should be noted that the evidence synthesis rule refers to combining boundary positive evidence values ​​and non-boundary negative evidence values ​​through normalization, using the boundary positive evidence value as the numerator and the sum of the boundary positive evidence value and the non-boundary negative evidence value as the denominator to construct a ratio, thereby obtaining the integrated boundary probability value.

[0118] The integration boundary probability value and cognitive uncertainty value of each pixel are written back according to the original pixel position to form the integration boundary probability map and cognitive uncertainty map.

[0119] The specific process includes backfilling the integrated boundary probability value and cognitive uncertainty value of each pixel according to their original spatial coordinates in the stacked object image, ensuring that the integrated boundary probability value at each spatial location accurately reflects the overall confidence that the location belongs to the boundary of the stacked object, and at the same time, ensuring that the cognitive uncertainty value at each spatial location accurately represents the degree to which the boundary prediction result at that location is affected by hypothesis conflict, and finally constructing an integrated boundary probability map and cognitive uncertainty map that are consistent with the spatial size of the stacked object image.

[0120] It should be noted that the integrated boundary probability map represents the comprehensive confidence that each pixel in the stacked object image belongs to the object boundary. It integrates the statistical consistency and topological rationality of multiple rounds of iterative samples, effectively improving the accuracy and robustness of boundary localization. The cognitive uncertainty map represents the degree of cognitive uncertainty of each pixel in the stacked object image during the boundary prediction process due to multiple hypothesis conflicts or insufficient evidence, providing a reliable assessment basis for subsequent boundary optimization. The hypothesis conflict impact refers to the boundary attribution uncertainty caused by inconsistent predictions of multiple boundary probability samples at the same pixel location. It is obtained by acquiring the transport deviation between each boundary probability sample and the local integrated boundary distribution and combining it with the differences in local neighborhood topology.

[0121] like Figure 5A heatmap of cognitive uncertainty under occlusion and fuzzy coupling was used to verify the ability of this invention to identify unreliable boundary regions under occlusion and fuzzy coupling conditions. Statistical analysis was performed on the average cognitive uncertainty values ​​under different combinations of stacked occlusion rates and boundary fuzziness intensities. The results are as follows: Figure 5 As shown. By Figure 5 It can be seen that as the stacking occlusion rate and the intensity of boundary blurring increase together, the cognitive uncertainty value generally shows an upward trend, indicating that the present invention can effectively characterize boundary regions where multiple hypothesis conflicts are enhanced and insufficient evidence is aggravated. Furthermore, by forming an integrated boundary probability map and a cognitive uncertainty map based on multiple boundary probability samples, the present invention can provide a reliable basis for subsequent adaptive threshold neighborhood construction and dynamic threshold adjustment, thereby more effectively suppressing false activations, eliminating false contours in boundary blurring regions, and improving the credibility of the final boundary results.

[0122] S4. Adaptive thresholding is applied to the integrated boundary probability map using the cognitive uncertainty map to generate a binary boundary mask.

[0123] Based on the integrated boundary probability map and the cognitive uncertainty map, an adaptive threshold neighborhood is constructed for each pixel.

[0124] The specific process includes obtaining the corresponding integrated boundary probability value and cognitive uncertainty value for each pixel location based on the integrated boundary probability map and the cognitive uncertainty map. An initial threshold is set based on the integrated boundary probability value, and the initial threshold is dynamically adjusted according to the cognitive uncertainty value. When the cognitive uncertainty value is high, the threshold neighborhood range is expanded to enhance robustness, and when the cognitive uncertainty value is low, the threshold neighborhood range is narrowed to improve positioning accuracy, thereby constructing an adaptive threshold neighborhood for each pixel that is adapted to its local reliability.

[0125] Based on the adaptive threshold neighborhood corresponding to each pixel, and combined with the integrated boundary probability value and the cognitive uncertainty value, the adaptive threshold corresponding to each pixel is obtained.

[0126] The specific process includes: based on the adaptive threshold neighborhood corresponding to each pixel, extracting the ensemble boundary probability value of all pixels within the adaptive threshold neighborhood from the ensemble boundary probability map, and obtaining the statistical central trend as the basic threshold reference. At the same time, the basic threshold reference is adjusted in combination with the cognitive uncertainty value of the corresponding neighborhood in the cognitive uncertainty map. When the overall cognitive uncertainty value in the neighborhood is high, the threshold is increased to suppress false activation in uncertain regions. When the overall cognitive uncertainty value in the neighborhood is low, the threshold is decreased to retain weak but reliable boundary responses. Finally, an adaptive threshold matching its local context and prediction reliability is determined for each pixel.

[0127] Based on the adaptive threshold corresponding to each pixel, the integrated boundary probability map is compared to obtain strong boundary pixels, strong non-boundary pixels, and boundary pixels to be determined.

[0128] The specific process includes comparing the integration boundary probability value of each pixel position in the integration boundary probability map with the adaptive threshold at that position based on the adaptive threshold corresponding to each pixel. When the integration boundary probability value is greater than the adaptive threshold and exceeds the preset high confidence margin, it is determined to be a strong boundary pixel. When the integration boundary probability value is less than the adaptive threshold and lower than the preset low confidence margin, it is determined to be a strong non-boundary pixel. When the difference between the integration boundary probability value and the adaptive threshold is in the middle range, it is determined to be a boundary pixel to be determined.

[0129] It should be noted that the preset high confidence margin is a fixed offset set by evaluating the impact of different margin values ​​on the accuracy of strong boundary pixel determination on the confidence requirements of the boundary discrimination task for strong boundary responses. The low confidence margin refers to the threshold offset used to determine strong non-boundary pixels. It is a fixed offset set by evaluating the impact of different margin values ​​on the accuracy of strong non-boundary pixel determination on the validation set by the confidence requirements of the boundary discrimination task for non-boundary regions.

[0130] The strong boundary pixels are used to perform connectivity filtering on the boundary pixels to be judged, and the target boundary pixels are obtained.

[0131] The specific process includes using strong boundary pixels to perform connectivity filtering on the boundary pixels to be determined. By checking whether the boundary pixels to be determined form an eight-neighbor or four-neighbor connection with at least one strong boundary pixel in the spatial neighborhood, the boundary pixels to be determined that meet the connectivity conditions are retained, and isolated boundary pixels to be determined that are not connected to any strong boundary pixels are removed, thereby obtaining the target boundary pixels.

[0132] The strong boundary pixels and the target boundary pixels are assigned to the boundary class, and the strong non-boundary pixels and the boundary pixels to be determined are assigned to the non-boundary class to generate a binary boundary mask.

[0133] The specific process includes: uniformly setting the values ​​of strong boundary pixels and target boundary pixels at corresponding positions in the output image with the same spatial size as the stacked object image to a fixed category label representing the boundary, which is usually one; uniformly setting the values ​​of strong non-boundary pixels and the boundary pixels to be determined at corresponding positions in the output image to a fixed category label representing the non-boundary, which is usually zero; through this pixel-by-pixel classification and assignment operation, a binary boundary mask with the same spatial size as the stacked object image and containing only two discrete values, boundary class and non-boundary class, is generated.

[0134] S5. Use a binary boundary mask to extract the boundary contours of the stacked objects, and calculate the confidence score of the boundary contours based on the integrated boundary probability value and cognitive uncertainty value of the region corresponding to the boundary contours, thereby obtaining boundary information with confidence assessment.

[0135] Extract the skeleton of the connected region from the binary boundary mask to generate a set of boundary contours.

[0136] The specific process includes extracting all four-neighbor or eight-neighbor connected regions composed of boundary-type pixels from the binary boundary mask, applying morphological skeletonization operation to each connected region, shrinking the connected region into a centerline structure with a single pixel width, preserving the topological connectivity and geometric orientation of each connected region, and organizing the skeleton line set corresponding to all connected regions into a boundary contour set.

[0137] The boundary contour set is mapped to the integrated boundary probability map and cognitive uncertainty map, and the probability and uncertainty feature sequences of the corresponding pixels are obtained.

[0138] The specific process includes treating each boundary contour in the boundary contour set as a continuous curve composed of a series of ordered spatial coordinate points. Based on the position of these coordinate points in the stacked object image, the pixel values ​​with the same coordinates in the integrated boundary probability map are queried point by point to obtain a one-dimensional integrated boundary probability value sequence corresponding to the boundary contour. At the same time, based on the same coordinate points, the pixel values ​​with the same coordinates in the cognitive uncertainty map are queried point by point to obtain a one-dimensional cognitive uncertainty value sequence corresponding to the boundary contour. These two sequences are paired in a point-to-point order to form the probability and uncertainty feature sequence corresponding to the boundary contour. All boundary contours are processed in this way to obtain a set of probability and uncertainty feature sequences that correspond one-to-one with the boundary contour set.

[0139] Based on the probability and uncertainty feature sequence, the comprehensive confidence score of the boundary contour is calculated, and the expression is:

[0140] ;

[0141] in, A comprehensive confidence score representing the boundary contour; The total arc length representing the boundary profile; The integral path representing the boundary contour; Represents spatial coordinates The integration boundary probability value at that point. Represents spatial coordinates The value of cognitive uncertainty at the location; Represents the zero-smoothing constant; Indicates the curvature weighting coefficient; Represents spatial coordinates The local curvature value at that location; This represents the arc length parameter.

[0142] It should be noted that this expression has been calculated using a zero-prevention smoothing constant before computation. A physical quantity explicitly defined as having a unit of length, such that local curvature and Multiplication yields a dimensionless result, thus ensuring... The two quantities have the same dimensions; at the same time, the integrated boundary probability value With cognitive uncertainty value Both are dimensionless quantities, and their product is also dimensionless. Therefore, both terms in the integrand are dimensionless quantities. (This is followed by a series of infinitives related to arc length.) After integration, the whole has a unit of length, which is then combined with the total arc length. Divide and the final overall confidence score As a dimensionless quantity, it achieves dimensional unification across all formulas.

[0143] Furthermore, Represents spatial coordinates The integrated boundary probability value refers to the value stored in the spatial coordinate position of the integrated boundary probability map. It is derived from the integrated boundary confidence of each pixel after geometric transport aggregation and evidence synthesis of multiple boundary probability samples. Represents spatial coordinates The cognitive uncertainty value at a given location refers to the value stored in the cognitive uncertainty map at the spatial coordinate position. It is derived from the cognitive uncertainty measure of each pixel obtained based on the local ensemble boundary distribution and the comprehensive conflict quantity. The curvature weight coefficient is a dimensionless scalar parameter used to adjust the degree of influence of local curvature on the comprehensive confidence score. It is a fixed value set by optimization through the validation set based on the importance of the geometric characteristics of the boundary contour. Represents spatial coordinates The local curvature value refers to the degree of geometric curvature of the boundary profile at that coordinate position. It is obtained by parameterizing the boundary profile, obtaining the first and second derivatives, and substituting them into the curvature formula.

[0144] The specific process includes: based on the probability and uncertainty feature sequence, normalizing the integrated boundary probability value at the spatial coordinates along the integral path of the boundary contour; forming a ratio between the integrated boundary probability value and the sum of the cognitive uncertainty value and the regularization term; simultaneously combining the local curvature value and weighting it with a zero-prevention smoothing constant to give higher weight to high curvature regions; then performing arc length integration on the weighted response along the entire path; and forming a ratio between the integration result and the total arc length of the boundary contour to obtain the comprehensive confidence score of the boundary contour.

[0145] The boundary contour set is bound to the comprehensive confidence score to generate boundary information with confidence assessment.

[0146] The specific process includes pairing each boundary contour in the boundary contour set with a comprehensive confidence score obtained through probability and uncertainty feature sequences, according to the corresponding order or identifier during the generation process, to ensure that each boundary contour is associated with a unique comprehensive confidence score. The comprehensive confidence score reflects the overall credibility determined by the response strength of the boundary contour in the integrated boundary probability map, the reliability in the cognitive uncertainty map, and the curvature characteristics along the contour geometry. By organizing the geometric coordinate sequence of the boundary contour and its corresponding comprehensive confidence score into a unified data structure, an output containing complete contour morphology information and quantitative confidence assessment results is generated, thus generating boundary information with confidence assessment.

[0147] like Figure 6 The overall mean comparison chart of each method is presented to comprehensively verify the overall effectiveness of this invention in terms of boundary extraction accuracy, texture noise suppression capability, and result reliability. The boundary F1 score, high-confidence contour accuracy, and texture region false detection suppression capability of each method are compared. The results are as follows: Figure 6 As shown. By Figure 6 It can be seen that experimental group E is superior to control groups A to D in all of the above indicators, indicating that the present invention does not only achieve improvement in a certain local aspect, but also achieves simultaneous improvement in boundary extraction accuracy, false contour suppression ability and result credibility through the complementary fusion of appearance feature map and material invariance feature map, integrated reasoning of multiple boundary probability samples, synchronous construction of cognitive uncertainty map, and the synergistic effect of adaptive threshold processing and boundary contour confidence scoring. This makes it more in line with the actual application requirements of high precision and high reliability for stacked object boundary extraction tasks.

[0148] This embodiment also provides a computer device applicable to the image recognition-based method for extracting the boundaries of stacked objects, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the image recognition-based method for extracting the boundaries of stacked objects as proposed in the above embodiment.

[0149] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0150] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the image recognition-based method for extracting the boundaries of stacked objects as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0151] In summary, this invention achieves orthogonal separation of appearance texture features and material invariance features through steady-state wavelet transform and reflectivity component decomposition in image recognition, thereby improving feature discrimination in stacked occlusion scenarios and suppressing false detections due to texture noise. Through multiple forward inferences and dynamic noise field construction in image recognition, it achieves the synchronous generation of pixel-level integrated boundary probabilities and cognitive uncertainty values, thereby eliminating false contours in blurred boundary areas and providing reliable confidence scores.

[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for extracting the boundaries of stacked objects based on image recognition, characterized in that, include: Acquire images of stacked objects to be processed, and input these images into a dual-branch feature encoder to output appearance feature maps and material invariance feature maps; The appearance feature map and the material invariance feature map are fused to generate a fused feature map, which is then input into the boundary prediction decoder to obtain the initial boundary probability map corresponding to each pixel. Based on the initial boundary probability map, multiple forward inferences are performed on the stacked object image to generate multiple boundary probability samples. Based on the multiple boundary probability samples, the integrated boundary probability value and cognitive uncertainty value of each pixel are obtained to form an integrated boundary probability map and a cognitive uncertainty map. An adaptive thresholding process is applied to the ensemble boundary probability map using a cognitive uncertainty graph to generate a binary boundary mask. Binary boundary masks are used to extract the boundary contours of stacked objects, and the confidence score of the boundary contours is calculated based on the integrated boundary probability value and cognitive uncertainty value of the corresponding regions to obtain boundary information with confidence assessment.

2. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1, characterized in that, The specific steps for outputting the appearance feature map and the material invariance feature map are as follows: The stacked object image to be processed is obtained, and the stacked object image is normalized and subjected to steady-state wavelet transform to obtain a high-frequency texture image carrying texture information and a low-frequency structure image carrying structural information. The high-frequency texture image is input into the appearance feature encoding branch, and the edge information is enhanced by a high-frequency attention mechanism to generate the initial appearance feature map. The low-frequency structure image is decomposed to obtain the reflectance component map, and the reflectance component map is input into the material invariance coding branch to generate the initial material feature map; The initial appearance feature map and the initial material feature map are orthogonalized to generate an appearance feature map and a material invariance feature map.

3. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1, characterized in that, The specific steps for generating the fused feature map are as follows: The channel-level local variance values ​​of the appearance feature map and the material invariance feature map are calculated separately to obtain their respective local variance information; Based on local variance information, pixel-level fusion weight coefficients are calculated using a normalized exponential function. The appearance feature map and the material invariance feature map are combined at the pixel level using pixel-level fusion weight coefficients and then convolutional smoothing is performed to generate a fused feature map.

4. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1 or 3, characterized in that, The specific steps for obtaining the initial boundary probability map corresponding to each pixel are as follows: The fused feature map is input into the boundary prediction decoder and the resolution is restored by transposed convolution to obtain a high-resolution feature map; Gradient components are obtained by applying a fixed convolution kernel to the high-resolution feature map, and gradient feature information is obtained. Gradient feature information is used to modulate the high-resolution feature map to generate an edge-enhanced feature map. The edge enhancement feature map is processed using channel compression convolution and activation functions to obtain the initial boundary probability map corresponding to each pixel.

5. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1, characterized in that, The specific steps for generating multiple boundary probability samples are as follows: The initial boundary probability map is mapped to the initial memory state matrix, and the inference iteration counter and dynamic noise generator are initialized at the same time. The spatial correlation noise field of the current iteration step is constructed based on the dynamic noise generator, and the boundary probability sample value of the current iteration step is calculated by combining the initial memory state matrix. The boundary probability sample values ​​of the current iteration step are stored in the sample sequence set, and the inference iteration counter is incremented until the inference iteration counter reaches the preset number of inferences. The resulting sample sequence set is then used as multiple boundary probability samples.

6. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1 or 5, characterized in that, The specific steps for forming the integrated boundary probability map and the cognitive uncertainty map are as follows: Multiple boundary probability samples are aligned according to pixel position, and geometric transport aggregation is performed in the local neighborhood corresponding to each pixel to obtain the local integrated boundary distribution corresponding to each pixel. Based on the local ensemble boundary distribution corresponding to each pixel, the transport bias of each boundary probability sample is obtained, and topological consistency analysis is performed in combination with the connectivity, discontinuity and branch differences in the local neighborhood to obtain the comprehensive conflict amount corresponding to each pixel. Based on the local integration boundary distribution and integrated conflict amount corresponding to each pixel, the boundary positive evidence value and non-boundary negative evidence value corresponding to each pixel are generated. Based on the boundary positive evidence value and non-boundary negative evidence value, the integration boundary probability value and cognitive uncertainty value of each pixel are obtained. The integration boundary probability value and cognitive uncertainty value of each pixel are written back according to the original pixel position to form the integration boundary probability map and cognitive uncertainty map.

7. The method for extracting the boundary of stacked objects based on image recognition as described in claim 1, characterized in that, The specific steps for generating the binary boundary mask are as follows: Based on the integrated boundary probability map and the cognitive uncertainty map, an adaptive threshold neighborhood is constructed for each pixel. Based on the adaptive threshold neighborhood corresponding to each pixel, the adaptive threshold corresponding to each pixel is obtained by combining the integrated boundary probability value and the cognitive uncertainty value. Based on the adaptive threshold corresponding to each pixel, the integrated boundary probability map is compared to obtain strong boundary pixels, strong non-boundary pixels, and boundary pixels to be determined. The strong boundary pixels are used to perform connectivity filtering on the boundary pixels to be judged to obtain the target boundary pixels; The strong boundary pixels and the target boundary pixels are assigned to the boundary class, and the strong non-boundary pixels and the boundary pixels to be determined are assigned to the non-boundary class to generate a binary boundary mask.

8. The method for extracting the boundary of stacked objects based on image recognition as described in claim 7, characterized in that, The specific steps for obtaining boundary information with confidence assessment are as follows: Extract the skeleton of the connected region from the binary boundary mask to generate a set of boundary contours; The boundary contour set is mapped to the integrated boundary probability map and cognitive uncertainty map, and the probability and uncertainty feature sequences of the corresponding pixels are obtained; Based on the probability and uncertainty feature sequence, the comprehensive confidence score of the boundary contour is calculated; The boundary contour set is bound to the comprehensive confidence score to generate boundary information with confidence assessment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the image recognition-based stacked object boundary extraction method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the image recognition-based stacked object boundary extraction method according to any one of claims 1 to 8.

Images