Projection-near-end agricultural instance segmentation method and system based on capacity limitation

Abstract

The invention provides a projection-near-end agricultural instance segmentation method and system based on capacity limitation. The method comprises the following steps: establishing a capacity-calculation contract, and constraining model parameter drift and frame rate; an orthogonal residual aggregation module is introduced into the feature space, and the edge consistency gradient is enhanced through orthogonal disturbance in a Cayley form; introducing an anisotropic offset correction module into a geometric space, and utilizing conical cutting and group-level low-rank sharing to stabilize gallery type scene sampling; an overlapping prototype clamping module is introduced into an output space, and fuzzy prediction is inhibited through near-end updating and exclusive loss. According to the method, the real-time reasoning requirement of embedded equipment is met, and meanwhile, the segmentation precision and connectivity of crop rows, travelable areas and obstacles in an unstructured environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, artificial intelligence, and agricultural robot perception technology, and in particular to a capacity-constrained projection-near-end agricultural instance segmentation method and system. Specifically, it relates to a capacity-constrained projection-near-end agricultural instance segmentation method and system for resource-constrained edge computing devices, capable of achieving high-precision perception in adverse weather and unstructured environments. Background Technology

[0002] With the rapid development of smart agriculture, autonomous agricultural machinery (such as intelligent tractors, autonomous harvesting robots, and plant protection drones) is playing an increasingly important role in modern agricultural operations. The visual perception system, as the "eyes" of agricultural robots, has the core task of accurately segmenting key elements in the farmland environment (such as crop rows, drivable areas, obstacles, weeds, and soil) to provide reliable semantic and geometric information for downstream path planning and operational control.

[0003] However, compared with urban road autonomous driving scenarios, agricultural unstructured scenarios have significant unique characteristics and challenges, and existing general instance segmentation technologies face many bottlenecks in practical applications: First, the unique geometric structure of agricultural scenes. Farmland environments typically exhibit a long-corridor structure with strong anisotropy, meaning they have great vertical depth and repetitive horizontal textures. Existing feature extraction networks (such as CNNs based on square convolutional kernels) often struggle to adapt to these elongated, deep geometric features, leading to unstable sampling at the far ends of crop rows or narrow ridges, resulting in segmentation breaks or structural discontinuities.

[0004] Second, the robustness of perception under adverse weather conditions is poor. Agricultural machinery operations often require 24 / 7 operation and frequently face adverse weather conditions such as fog, snow, strong glare, low light, and dust. These factors can lead to image contrast compression and texture degradation, making the boundaries between crops and the background (such as weeds and soil) blurred. Traditional lightweight segmentation models lack enhancement mechanisms for these degraded features, making them prone to missed detections or boundary drift.

[0005] Third, there is class overlap and semantic ambiguity. In complex farmland environments, objects of different categories often overlap or are adjacent at the pixel level. For example, weeds may grow between crop rows, and obstacles may be partially buried in the soil. This spatial aliasing can cause the model to produce ambiguous predictions in the Logit space at the output, that is, outputting multiple high-confidence categories for the same pixel, which affects the safety of the operation.

[0006] Fourth, there is a contradiction between the limited resources of edge computing devices and the high precision requirements. To meet the low-cost and low-power consumption requirements of agricultural operations, agricultural robots are usually equipped with embedded edge computing devices (such as the NVIDIA Jetson series). Currently, the mainstream high-precision segmentation models in academia (such as Mask2Former based on Transformer) have a huge number of parameters and high computational complexity, making it impossible to achieve real-time inference on edge devices; while lightweight models that sacrifice accuracy for speed (such as the original YOLO series), although meeting the speed requirements, often cannot meet the segmentation accuracy (especially the boundary IoU) requirements for safe navigation under the aforementioned harsh conditions and complex geometric structures.

[0007] Therefore, there is an urgent need to develop an agricultural instance segmentation method and system that can strictly adhere to the computing and storage capacity limitations of edge devices (capacity-constrained) and can also address problems such as severe weather, long corridor geometry, and category overlap through an effective structured feature correction mechanism. Summary of the Invention

[0008] The purpose of this invention is to address the challenge of existing agricultural autonomous driving perception systems struggling to maintain robustness and high-precision boundary localization under complex and adverse weather conditions (such as fog, snow, and strong light) when edge computing resources are limited. This invention proposes a capacity-constrained projection-near-end agricultural instance segmentation method and system. The method establishes a capacity-computation contract and, through three lightweight correction modules in the feature space, geometric space, and output space, improves the segmentation accuracy and connectivity of key navigation regions in unstructured environments without significantly increasing the number of model parameters.

[0009] The present invention is achieved through the following technical solution: the present invention proposes a projection-proximal agricultural instance segmentation method based on capacity-constrained method, the method including a feature space orthogonal residual aggregation stage, a geometric space anisotropic offset correction stage and an output space overlapping prototype clamping stage; Feature Space Orthogonal Residual Aggregation Stage: To address the boundary blurring problem caused by texture degradation in agricultural scenarios, orthogonal residual paths with controllable orientations are introduced at the shallow features of the backbone network. This stage calculates the depth-separable directional responses along the row and column directions, constructs symbolic directional imbalance terms, and enhances edge consistency gradients and suppresses false texture noise through Cayley-style orthogonal perturbations and lightweight projections, providing clear boundary cues for subsequent segmentation without changing the feature channel dimensions. Geometric anisotropic offset correction stage: To address the issues of longitudinal structural dependence and lateral sampling jitter in corridor-style agricultural scenarios, the feature sampling offset at the pyramid level is modeled. In this stage, the original offset is decomposed into components along the longitudinal and lateral directions of the corridor. The lateral component is pruned using the tapered aperture parameter, and combined with the low-rank lateral direction shared by the group level, high-frequency lateral aliasing is suppressed while preserving long-distance longitudinal dependence, thus achieving stable sampling of slender crop rows and road boundaries. Output Space Overlapping Prototype Clamping Stage: To address the pixel-level high overlap and ambiguity issues of key categories in agricultural navigation, a proximal clamping update is performed in the Logit space. This stage combines Softmax probability, prototype similarity, and region gating mechanisms to mine overlapping neighborhoods. Single-step proximal updates suppress the fuzzy Logit values ​​of mixed regions, and an exclusive loss is introduced to force semantic separation, thereby improving class separability and security.

[0010] Furthermore, the orthogonal residual aggregation stage of the feature space specifically involves: given a shallow feature tensor First, the directional response is calculated to construct the directional imbalance term. After obtaining the directional imbalance term, orthogonal perturbations of Cayley form and lightweight projection are applied. The output features are obtained.

[0011] Furthermore, the aforementioned imbalance in construction direction specifically refers to: (1) in, This represents the depth-separable convolutional response along the row direction of the feature map. This represents a depth-separable convolutional response along the column direction of the feature map, used to capture horizontal and vertical texture features in agricultural scenes; and These are learnable, channel-wise scaling factors used to adaptively weight edge information in different directions.

[0012] Furthermore, the output feature specifically includes: (2) in, I It is the identity matrix. For the enhanced output features; The hyperparameters used to control the intensity of the disturbance; Item In implementation, it is decomposed into stable rational transformations per channel and per position, avoiding the computational overhead of global matrix inversion; For inclusion Projection layers with convolution, normalization, and nonlinear activation are injected as proximal operators into edge-focusing residuals.

[0013] Furthermore, the geometric spatial anisotropic migration correction stage specifically involves: adjusting the pyramid level migration features. U Perform modeling and output the corrected sampling offset field: (3) in, Used to predict the original dense offset; Let be the basis vectors along the longitudinal direction of the corridor. Let be the basis vectors along the transverse direction of the corridor; Indicates the first g A low-rank lateral direction vector shared by the group. The weight of spatial variation after tapered cutting; It is a differentiable spatial transformer used to apply the calculated anisotropic offset field to the feature map.

[0014] Furthermore, the output space overlap prototype clamping stage specifically involves: constructing an exclusive loss function to penalize highly overlapping categories. (4) in, For network parameters; , Categories a and b The Softmax probability plot; , This is a class prototype similarity map calculated using the lightweight ProtoConv. These are the region-gated weights obtained through global average pooling, used to suppress background noise; This represents element-wise multiplication; For indicator functions; The predefined binarization threshold; For categories mined based on overlap proximity a The neighborhood set of the model is used to perform proximal clamp updates in the Logit space during the forward propagation of model inference and training: (5) in, For the original category Logits, For the updated Logits; The overlapping neighbor weights are calculated based on the IoU mining. This is a stability mask used to prevent over-correction in non-overlapping regions; The proximal step size.

[0015] Furthermore, the overall optimization objective of the method is specifically defined as follows: combining baseline loss and exclusionary constraints, the overall loss function is defined as: (6) in, For classifying losses, For distance regression loss, To mask and predict loss, This is the balancing coefficient for the exclusive loss term; the entire model training process follows the capacity-computation contract. Ensure that parameter drift and frame rate drop are strictly controlled within the allowable range. Inside.

[0016] This invention also proposes a capacity-constrained projection-proximal agricultural instance segmentation system, which includes a feature space orthogonal residual aggregation module, a geometric space anisotropic offset correction module, and an output space overlapping prototype clamping module. Feature space orthogonal residual aggregation module: Configures shallow features for receiving backbone network output, calculates depth-separable directional responses along row and column directions, constructs directional imbalance terms, and generates output features with enhanced edge consistency through local Cayley transform and lightweight projection; Geometric spatial anisotropic offset correction module: configured to predict the sampling offset and decompose it into vertical and horizontal components in the feature pyramid decoder, perform cone pruning and group-level low-rank sharing on the horizontal component, and resample the feature map through a differentiable spatial transformer. Output Space Overlapping Prototype Clamping Module: Configured to mine overlapping neighborhoods at the classification head output by combining Softmax probability, prototype similarity and region gating, calculate exclusive loss, and perform near-end clamping update in Logit space to correct fuzzy predictions.

[0017] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the capacity-constrained projection-proximal agriculture instance segmentation method.

[0018] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the capacity-constrained projection-proximal agriculture instance segmentation method.

[0019] The beneficial effects of this invention are: This invention aims to address the challenge of existing agricultural autonomous driving perception systems struggling to balance robustness and high-precision boundary localization under complex and adverse weather conditions, especially given the limited resources of edge computing devices. The method and system described in this invention effectively solve the perception failure problem of agricultural autonomous driving under adverse weather conditions (fog, snow, strong light, low illumination) by introducing projection-near-end correction operators into the feature, geometry, and output spaces. Through explicit constraints of the capacity-computation contract, this model (AURORA-AgriSeg) significantly improves the boundary IoU and structural integrity of crop rows, drivable areas, and small obstacles while maintaining the real-time inference speed (49.9 FPS) of edge devices (such as Jetson Orin NX). Compared to existing heavy Transformer models, this invention achieves superior cross-domain generalization ability and resistance to environmental interference while significantly reducing computational load. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0021] Figure 1 This is a diagram illustrating the overall architecture of the capacity-constrained projection-proximal agriculture instance segmentation system (AURORA-AgriSeg) described in this invention.

[0022] Figure 2 This is a schematic diagram of the internal structure of the Feature Space Orthogonal Residual Aggregation Module (OREA) in this invention.

[0023] Figure 3 This is a schematic diagram of the internal structure of the geometric space anisotropic offset correction module (ACORN) in this invention.

[0024] Figure 4 This is a schematic diagram of the internal structure of the Output Space Overlap Prototype Clamp Module (OPPC) in this invention. Detailed Implementation

[0025] 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.

[0026] Combination Figures 1-4 This invention proposes a capacity-constrained projection-near-end agricultural instance segmentation method, which includes a feature space orthogonal residual aggregation stage, a geometric space anisotropic offset correction stage, and an output space overlapping prototype clamping stage. The system aims to establish a capacity-computation contract while meeting the real-time inference requirements of edge devices, and achieve accurate perception of agricultural scenarios through three lightweight correction modules. The Orthogonal Residual Edge Aggregation (OREA) stage addresses the boundary blurring problem caused by texture degradation in agricultural scenarios by introducing orthogonal residual paths with controllable orientations at the shallow features of the backbone network. This stage first receives the shallow features output by the backbone network, calculates the depth-separable directional responses along the row and column directions, constructs symbolic directional imbalance terms, and enhances edge consistency gradients and suppresses false texture noise through Cayley-style orthogonal perturbations and lightweight projections, providing clear boundary cues for subsequent segmentation without changing the feature channel dimensions. The geometrical anisotropic offset correction stage (ACORN: Anisotropic Cone-Projected Rank-Normalized Offsets) addresses the issues of longitudinal structural dependence and lateral sampling jitter in corridor-style agricultural scenarios by modeling the feature sampling offsets at the pyramid level. This stage decomposes the original offsets into longitudinal and lateral components along the corridor, uses tapered aperture parameters to prune the lateral components, and combines group-shared low-rank lateral directions to suppress high-frequency lateral aliasing while preserving long-distance longitudinal dependence, thus achieving stable sampling of slender crop rows and road boundaries. Output Space Overlap-Proximity Prototype Clamping (OPPC): To address the pixel-level high overlap and ambiguity issues of key agricultural navigation categories (such as crops, drivable areas, and obstacles), a proximal clamping update is performed in the Logit space. This stage combines Softmax probability, prototype similarity, and region gating mechanisms to mine overlapping neighborhoods. Single-step proximal updates suppress the fuzzy Logit values ​​of mixed regions, and an exclusive loss is introduced to force semantic separation, improving inter-class separability and safety.

[0027] Furthermore, the orthogonal residual aggregation stage of the feature space specifically involves: given a shallow feature tensor First, the directional response is calculated to construct the directional imbalance term. After obtaining the directional imbalance term, orthogonal perturbations of Cayley form and lightweight projection are applied. (Include Convolution, normalization, and nonlinear activation are applied to the input features to obtain the output features.

[0028] Furthermore, the aforementioned imbalance in construction direction specifically refers to: (1) in, This represents the depth-separable convolutional response along the row direction of the feature map. This represents a depth-separable convolutional response along the column direction of the feature map, used to capture horizontal and vertical texture features in agricultural scenes; and These are learnable, channel-wise scaling factors used to adaptively weight edge information in different directions.

[0029] Furthermore, the output feature specifically includes: (2) in, I It is the identity matrix. For the enhanced output features; To control the hyperparameters of disturbance intensity; in specific implementation, the project... In implementation, it is decomposed into channel-wise and position-wise stable rational transformations (local Cayley factor), avoiding the computational overhead of global matrix inversion; For inclusion Projection layers with convolution, batch normalization, and SiLU activation are used as proximal operators to inject edge-focusing residuals.

[0030] Furthermore, the geometric spatial anisotropic migration correction stage specifically involves: adjusting the pyramid level migration features. U Perform modeling and output the corrected sampling offset field: (3) in, Used to predict the original dense offset; Let be the basis vectors along the longitudinal direction of the corridor. Let be the basis vectors along the transverse direction of the corridor; Indicates the first g A low-rank lateral direction vector shared by the group. The spatial variation weights after cone clipping; Differentiable spatial transformers (such as GridSample) are used to apply the calculated anisotropic offset field to the feature map; this process ensures that the sampling points are distributed along the crop row direction by constraining the lateral degrees of freedom, thereby reducing lateral jitter.

[0031] Furthermore, the output space overlap prototype clamping stage specifically involves: constructing an exclusive loss function to penalize highly overlapping categories. (4) in, For network parameters; , Categories a and b The Softmax probability plot; , This is a class prototype similarity map calculated using the lightweight ProtoConv. These are the region-gated weights obtained through global average pooling, used to suppress background noise; This represents element-wise multiplication; For indicator functions; The predefined binarization threshold; For categories mined based on overlap proximity a The neighborhood set is used, and the loss is calculated only for spatially highly overlapping class pairs; during the forward propagation of model inference and training, a proximal clamp update in the Logit space is performed: (5) in, For the original category Logits, For the updated Logits; The overlapping neighbor weights, calculated based on mined IoU, reflect the category. a and b The degree of confusion; This is a stability mask used to prevent over-correction in non-overlapping regions; This is the proximal step size. This update part directly suppresses fuzzy predictions in the mixed region and increases the inter-class Logit distance.

[0032] Furthermore, the overall optimization objective of the method is specifically defined as follows: combining baseline loss and exclusionary constraints, the overall loss function is defined as: (6) in, For classifying losses, For distance regression loss, To mask and predict loss, This is the balancing coefficient for the exclusive loss term; the entire model training process follows the capacity-computation contract. Ensure that parameter drift and frame rate drop are strictly controlled within the allowable range. Inside.

[0033] This invention also proposes a capacity-constrained projection-proximal agricultural instance segmentation system, which includes a feature space orthogonal residual aggregation module, a geometric space anisotropic offset correction module, and an output space overlapping prototype clamping module. The orthogonal residual aggregation module in the feature space is configured to receive shallow features from the backbone network output, calculate depth-separable directional responses along the row and column directions, construct directional imbalance terms, and generate output features with enhanced edge consistency through local Cayley transform and lightweight projection. This module is configured to address the boundary blurring problem caused by texture degradation in agricultural scenes by introducing directionally controllable orthogonal residual paths at the shallow features of the backbone network. Specifically, this module provides shallow feature tensors... U Calculate the depth-separable directional response along the row and column directions. Constructing directional imbalance terms ; and through Cayley-style orthogonal perturbations and lightweight projections Obtain output features This enhances edge consistency gradients and suppresses false texture noise; Geometric Spatial Anisotropic Offset Correction Module: Configured to predict sampling offsets and decompose them into longitudinal and lateral components in the feature pyramid decoder, perform cone pruning and group-level low-rank sharing on the lateral components, and resample the feature map through a differentiable spatial transformer; this module is configured to model the feature sampling offsets at the pyramid level to address the longitudinal structure dependency and lateral sampling jitter issues present in corridor-style agricultural scenarios; specifically, this module decomposes the original offset into components along the corridor longitudinal direction. With horizontal The components are trimmed using tapered aperture parameters, and combined with the low-rank lateral direction shared by the group level. Through formula Generate a corrected offset field to achieve stable sampling of slender crop rows and road boundaries; Output Space Overlapping Prototype Clamping Module: Configured to mine overlapping neighborhoods at the classification head output by combining Softmax probability, prototype similarity, and region gating, calculate the exclusivity loss, and perform near-end clamping updates in the Logit space to correct for fuzzy predictions. This module is configured to address the pixel-level high overlap and fuzziness issues of key categories in agricultural navigation by performing near-end clamping updates in the Logit space. Specifically, this module combines Softmax probability, prototype similarity, and region gating mechanisms to mine overlapping neighborhoods. Calculate the exclusive loss Simultaneously, updates are performed during inference. By updating the fuzzy Logit value of the mixed region in a single step near-end update, the separability and security between classes are improved.

[0034] Example Combination Figure 1 This invention proposes a capacity-constrained projection-near-end agricultural instance segmentation method, which includes a feature space orthogonal residual aggregation stage, a geometric space anisotropic offset correction stage, and an output space overlapping prototype clamping stage. The system aims to establish a capacity-computation contract while meeting the real-time inference requirements of edge devices, and achieve accurate perception of agricultural scenarios through three lightweight correction modules.

[0035] Feature Space Orthogonal Residual Aggregation Stage: To address the boundary blurring problem caused by texture degradation in agricultural scenarios, orthogonal residual paths with controllable orientations are introduced at the shallow features of the backbone network. This stage calculates the depth-separable directional responses along the row and column directions, constructs symbolic directional imbalance terms, and enhances edge consistency gradients and suppresses false texture noise through Cayley-style orthogonal perturbations and lightweight projections, providing clear boundary cues for subsequent segmentation without changing the feature channel dimensions.

[0036] Furthermore, the Orthogonal Residual Aggregation (OREA) stage in the feature space specifically involves: combining... Figure 2 Given shallow feature tensors The OREA module first calculates two depth-separable directional responses, one along the row direction and the other along the row direction. and column direction And introduce learnable channel-wise scaling factors. and Define the symbolic direction imbalance term: (1) This project utilizes Cayley-style orthogonal perturbations and lightweight projections. (Include Convolution, normalization, and nonlinear activation are applied to the input features to obtain the output features. : (2) in, Controlling the intensity of disturbances. In specific implementation, the project... It is implemented as a channel-wise, position-wise stable rational transformation (local Cayley factor) rather than a global matrix inversion, to ensure that OREA injects edge-focusing residuals as a proximal operator while maintaining feature energy stability.

[0037] Geometric space anisotropic offset correction stage: To address the issues of vertical structure dependence and lateral sampling jitter in corridor-style agricultural scenarios, the feature sampling offset of the pyramid level is modeled.

[0038] Furthermore, the geometrical anisotropic migration correction stage (ACORN) specifically involves: combining... Figure 3 This stage aims to stabilize the sampling offset in the Feature Pyramid Network (FPN) decoder. It involves processing the input pyramid-level features. U The module outputs the corrected sampled offset field. This process first predicts the original offset. Then decompose it into longitudinal sections along the corridor. and horizontal The components are truncated using tapered aperture parameters, and combined with the low-rank lateral direction shared by the group level. The final offset field generation process is as follows: (3) in, The spatial variation weights after tapered cutting. This is used to apply a differentiable spatial transformer (such as GridSample) to the offset field of the results. The process constrains the lateral degrees of freedom, ensuring that the sampling points are distributed along the crop row direction, preserving long-distance longitudinal dependence while suppressing high-frequency lateral aliasing.

[0039] Output Space Overlap Prototype Clamping Stage: To address the pixel-level high overlap and blurring issues of key agricultural navigation categories (such as crops, drivable areas, and obstacles), near-end clamping updates are performed in the Logit space.

[0040] Furthermore, the Output Space Overlap Prototype Clamping (OPPC) stage specifically comprises: combining Figure 4 This stage first constructs an exclusive loss function to penalize highly overlapping categories: (4) in, p For Softmax probability, s This represents the prototype similarity calculated using the lightweight ProtoConv. The region-gated weights are obtained through global pooling. For a predefined threshold, For categories mined based on overlap proximity a The neighborhood set. During inference and training, a near-end clamp update of the Logit space is performed: (5) Where Z represents the original category Logits. For the updated Logits, To determine the overlapping neighbor weights based on IoU mining, For stability mask, This is the proximal step size. This single-step update directly suppresses fuzzy predictions in mixed regions, improving inter-class separability.

[0041] Furthermore, the overall optimization objective of the method of this invention is: Combining baseline loss and exclusiveness constraints, the overall loss function is defined as: (6) in, These are classification loss, distance regression loss, and mask prediction loss, respectively. This is the balancing coefficient for the exclusivity loss term. The entire model follows the capacity-computational contract. This ensures that parameter drift and frame rate drop are within acceptable limits.

[0042] This invention also proposes a capacity-constrained projection-proximal agricultural instance segmentation system, which includes a feature space orthogonal residual aggregation module, a geometric space anisotropic offset correction module, and an output space overlapping prototype clamping module. Feature space orthogonal residual aggregation module: Configures shallow features for receiving backbone network output, calculates depth-separable directional responses along row and column directions, constructs directional imbalance terms, and generates output features with enhanced edge consistency through local Cayley transform and lightweight projection; Geometric spatial anisotropic offset correction module: configured to predict the sampling offset and decompose it into vertical and horizontal components in the feature pyramid decoder, perform cone pruning and group-level low-rank sharing on the horizontal component, and resample the feature map through a differentiable spatial transformer. Output Space Overlapping Prototype Clamping Module: Configured to mine overlapping neighborhoods at the classification head output by combining Softmax probability, prototype similarity and region gating, calculate exclusive loss, and perform near-end clamping update in Logit space to correct fuzzy predictions.

[0043] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the capacity-constrained projection-proximal agriculture instance segmentation method.

[0044] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the capacity-constrained projection-proximal agriculture instance segmentation method.

[0045] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0046] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0047] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0048] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0049] The present invention has been described in detail above as a method and system for segmenting projection-proximal agricultural instances based on capacity-constrained projection. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for segmenting agricultural instances based on capacity-constrained projection-proximal agriculture, characterized in that, The method includes a feature space orthogonal residual aggregation stage, a geometric space anisotropic offset correction stage, and an output space overlapping prototype clamping stage. Feature Space Orthogonal Residual Aggregation Stage: To address the boundary blurring problem caused by texture degradation in agricultural scenarios, orthogonal residual paths with controllable orientations are introduced at the shallow features of the backbone network. This stage calculates the depth-separable directional responses along the row and column directions, constructs symbolic directional imbalance terms, and enhances edge consistency gradients and suppresses false texture noise through Cayley-style orthogonal perturbations and lightweight projections, providing clear boundary cues for subsequent segmentation without changing the feature channel dimensions. Geometric anisotropic offset correction stage: To address the issues of longitudinal structural dependence and lateral sampling jitter in corridor-style agricultural scenarios, the feature sampling offset at the pyramid level is modeled. In this stage, the original offset is decomposed into components along the longitudinal and lateral directions of the corridor. The lateral component is pruned using the tapered aperture parameter, and combined with the low-rank lateral direction shared by the group level, high-frequency lateral aliasing is suppressed while preserving long-distance longitudinal dependence, thus achieving stable sampling of slender crop rows and road boundaries. Output Space Overlapping Prototype Clamping Stage: To address the pixel-level high overlap and ambiguity issues of key categories in agricultural navigation, a proximal clamping update is performed in the Logit space. This stage combines Softmax probability, prototype similarity, and region gating mechanisms to mine overlapping neighborhoods. Single-step proximal updates suppress the fuzzy Logit values ​​of mixed regions, and an exclusive loss is introduced to force semantic separation, thereby improving class separability and security.

2. The method according to claim 1, characterized in that, The orthogonal residual aggregation stage of the feature space specifically involves: given a shallow feature tensor First, the directional response is calculated to construct the directional imbalance term. After obtaining the directional imbalance term, orthogonal perturbations of Cayley form and lightweight projection are applied. The output features are obtained.

3. The method according to claim 2, characterized in that, The aforementioned imbalance in construction direction specifically refers to: (1) in, This represents the depth-separable convolutional response along the row direction of the feature map. This represents a depth-separable convolutional response along the column direction of the feature map, used to capture horizontal and vertical texture features in agricultural scenes; and These are learnable, channel-wise scaling factors used to adaptively weight edge information in different directions.

4. The method according to claim 3, characterized in that, The output features are specifically: (2) in, I It is the identity matrix. For the enhanced output features; The hyperparameters used to control the intensity of the disturbance; Item In implementation, it is decomposed into stable rational transformations per channel and per position, avoiding the computational overhead of global matrix inversion; For inclusion Projection layers with convolution, normalization, and nonlinear activation are injected as proximal operators into edge-focusing residuals.

5. The method according to claim 4, characterized in that, The geometric spatial anisotropic migration correction stage specifically involves: adjusting the pyramid level migration features. U Perform modeling and output the corrected sampling offset field: (3) in, Used to predict the original dense offset; Let be the basis vectors along the longitudinal direction of the corridor. Let be the basis vectors along the transverse direction of the corridor; Indicates the first g A low-rank lateral direction vector shared by the group. The weight of spatial variation after tapered cutting; It is a differentiable spatial transformer used to apply the calculated anisotropic offset field to the feature map.

6. The method according to claim 5, characterized in that, The output space overlap prototype clamping stage specifically involves: constructing an exclusive loss function to penalize highly overlapping categories. (4) in, For network parameters; , Categories a and b The Softmax probability plot; , This is a class prototype similarity map calculated using the lightweight ProtoConv. These are the region-gated weights obtained through global average pooling, used to suppress background noise; This represents element-wise multiplication; For indicator functions; The predefined binarization threshold; For categories mined based on overlap proximity a The neighborhood set of the model is used to perform proximal clamp updates in the Logit space during the forward propagation of model inference and training: (5) in, For the original category Logits, For the updated Logits; The overlapping neighbor weights are calculated based on the IoU mining. This is a stability mask used to prevent over-correction in non-overlapping regions; The proximal step size.

7. The method according to claim 6, characterized in that, The overall optimization objective of the method is as follows: combining baseline loss and exclusionary constraints, the overall loss function is defined as: (6) in, For classifying losses, For distance regression loss, To mask and predict loss, This is the balancing coefficient for the exclusive loss term; the entire model training process follows the capacity-computation contract. Ensure that parameter drift and frame rate drop are strictly controlled within the allowable range. Inside.

8. A capacity-constrained projection-proximal agricultural instance segmentation system, characterized in that, The system includes a feature space orthogonal residual aggregation module, a geometric space anisotropic offset correction module, and an output space overlapping prototype clamping module. Feature space orthogonal residual aggregation module: Configures shallow features for receiving backbone network output, calculates depth-separable directional responses along row and column directions, constructs directional imbalance terms, and generates output features with enhanced edge consistency through local Cayley transform and lightweight projection; Geometric spatial anisotropic offset correction module: configured to predict the sampling offset and decompose it into vertical and horizontal components in the feature pyramid decoder, perform cone pruning and group-level low-rank sharing on the horizontal component, and resample the feature map through a differentiable spatial transformer. Output Space Overlapping Prototype Clamping Module: Configured to mine overlapping neighborhoods at the classification head output by combining Softmax probability, prototype similarity and region gating, calculate exclusive loss, and perform near-end clamping update in Logit space to correct fuzzy predictions.

9. An electronic 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 method according to any one of claims 1-7.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images