A dynamic adjacency-based cross-scale graph bridging decoupling target detection method

Abstract

This invention relates to a cross-scale graph bridging decoupling target detection method based on dynamic adjacency, comprising the following steps: Step 1: Data preprocessing and segmentation; Step 2: Batch sampling and multi-scale feature extraction; Step 3: High-level self-attention enhancement; Step 4: Same-layer dynamic adjacency learning and alignment; Step 5: Cross-scale graph bridging; Step 6: Gated residual steady-state injection; Step 7: Decoupling detection head prediction; Step 8: Training and optimization; Step 9: Inference and evaluation. This invention learns same-layer dynamic graph adjacency on multi-scale paths and performs graph alignment, establishing a binary bridging relationship between adjacent scales for "directional leverage," and then uses gated residual steady-state injection to the decoupling detection head. This improves detection performance in small-scale and short-time scenes while maintaining high IoU accuracy and reducing overall resource consumption.

Description

Technical Field

[0001] This invention relates to a cross-scale graph bridging and decoupling target detection method based on dynamic adjacency, belonging to the field of target detection technology for construction site safety inspection. Background Technology

[0002] Automatic identification of personal protective equipment (PPE) in engineering construction scenarios requires the stable differentiation of multiple targets, such as whether they are worn, type / color of safety helmets, and safety vests, in crowded, dynamic, and complex lighting environments. The common challenges in this scenario are: dense targets, frequent occlusion, large scale range, high proportion of small targets, and interference from highly reflective materials and high-contrast textures.

[0003] Mainstream methods are mostly based on single-stage detectors, employing multi-scale fusion such as FPN / PAFPN / BiFPN to converge features of different resolutions to the detection head. Multi-head self-attention (MHSA) is introduced at high semantic layers (P5 / 32, etc.) to supplement long-range dependencies, and classification / regression decoupling is used at the detection head to alleviate task conflicts. Some studies also employ graph modeling (GNN / GAT) to utilize adjacency relationships for context propagation and noise suppression. However, the benefits of directly stacking these components are not stable, mainly in three aspects: First, there is a mismatch between cross-scale statistics and gradients. The P3, P4, and P5 generated by FPN differ significantly in semantic density and noise distribution. After stacking MHSA at higher levels, the correlation is reorganized and the distribution drifts. When directly fed into the decoupling head, the statistical priors of the two branches are inconsistent, training is highly sensitive to learning rate and regularization, gradients are prone to oscillation, and the improvement in mAP50–95 is limited. Second, pyramid fusion lacks explicit constraints from the "relationship layer." Methods like BiFPN primarily rely on layer-by-layer weighting, lacking fine-grained characterization of the scale from which information is acquired and its weights. This can lead to over-smoothing or erroneous aggregation in densely packed small targets and strong interference backgrounds. Third, graph modeling faces a dilemma: predefined or static adjacencies fail to generalize adequately under dynamic occlusion and crowd flow conditions, while dense global adjacencies introduce high complexity and noise propagation. Without controllable gating, erroneous contexts are easily amplified and propagated to the detection head. The domain characteristics of PPE scenes further exacerbate these contradictions: the bright stripes of reflective vests easily occupy resources in high-level attention; differences in materials and colors from different manufacturers lead to unstable channel selection preferences; and excessive global dependence in densely populated areas can bind together pedestrian features that should be independent, weakening the local accuracy of bounding box regression. Therefore, the simple concatenation of "MHSA × decoupling head" often results in increased parameters without increased efficiency and is sensitive to random initialization. Based on this need, there is an urgent need for an algorithmic mechanism that explicitly learns relationships on multi-scale paths and injects them in a controlled manner: at the same scale level, adaptive graph construction and graph convolution alignment are performed through dynamic adjacency to stabilize statistical and geometric priors; at the cross-scale level, sparse bridging relationships are used to achieve directional information transmission based on position and semantics rather than indiscriminate mixing; at the injection level, gated residuals start from approximate identity and gradually increase in quantity to avoid early training perturbations; at the same time, lightweight design is taken into account to ensure that a stable performance improvement friendly to mAP50–95 is achieved under small-scale configuration.

[0004] Based on the aforementioned practical problems and needs, the proposed approach of this invention is as follows: Dynamic graph adjacency is learned on multi-scale feature nodes, and cross-scale binary bridging is used to achieve fine-grained transmission of context "from which layer, to which location, and with what intensity"; then, the aligned information is steadily injected into the decoupled detection head through gating residuals, so that the high-level self-attention and the decoupled head no longer constrain each other, but form an interpretable and controllable cooperative relationship. The starting point of this approach is not "adding another module," but rather eliminating the mismatch problem first, and then discussing gains. Summary of the Invention

[0005] Existing object detection models, while maintaining high IoU accuracy (mAP50–95), still suffer from the following issues: 1) When simply stacking high-level self-attention and decoupled detection heads, the shortcomings of each other are mutually constrained, resulting in insufficient accuracy in short-term (small target / densely occluded scenes); 2) The computational complexity and memory usage are relatively high, and adding parameters does not improve efficiency in small-scale configurations; 3) The statistics of multi-scale features (P3 / P4 / P5) do not match with the gradient flow, leading to unstable training and sensitivity to hyperparameters. To address these issues, this invention proposes a cross-scale graph bridging decoupled object detection method based on dynamic adjacency (DGBD method): It learns the adjacency of the same-layer dynamic graph on multi-scale paths and performs graph alignment, establishes a binary bridging relationship between adjacent scales to "borrow strength", and then uses gated residual steady-state injection to the decoupled detection head. This method improves the detection performance of small-scale and short-term scenes while maintaining high IoU accuracy and reducing overall resource consumption.

[0006] A cross-scale graph bridging decoupling target detection method based on dynamic adjacency includes the following steps:

[0007] Step 1: Data Preprocessing and Segmentation: Obtain the raw data, set the input resolution and class mapping, and standardize the annotation; clean and segment the raw data to form training set, validation set and test set;

[0008] Step 2: Batch Sampling and Multi-Scale Feature Extraction: Randomly select images from the training set as a batch input model, and obtain P3, P4, and P5 three-scale features through the backbone network and pyramid path; record the spatial size and stride of each scale for subsequent bridging alignment;

[0009] Step 3: Enhancement of high-level self-attention: Input P5 into the high-level multi-head self-attention module MHSA-Darknet to obtain global context and long-range dependencies; keep the original statistics of P3 and P4.

[0010] Step 4: Dynamic Adjacency Learning and Alignment at the Same Layer: Perform the following steps on P3, P4, and P5 respectively: 1) Generate node embeddings and scale position embeddings, and determine local windows or k-nearest neighbors as candidate neighborhoods; 2) Adaptively generate sparse adjacencies based on feature compatibility and spatial priors; 3) Perform lightweight graph propagation on these adjacencies and output the same-layer alignment features.

[0011] Step 5: Cross-scale graph bridging: Perform the following between the two pairs of adjacent scales P5-P4 and P4-P3: 1) Align the resolution and channel dimension through upsampling / downsampling and linear mapping; 2) Establish sparse binary bridging relationships based on position and semantic compatibility; 3) Propagate the high-level global semantics down to the low-level along the bridging relationship and propagate the low-level details up to the high-level to obtain cross-layer enhanced features;

[0012] Step Six: Gated Residual Steady-State Injection: The output of "same-layer alignment + cross-layer enhancement" is fused with the original scale features using gated residuals: 1) The gate parameters are initialized negatively, with an approximate identity mapping in the early stages of training; 2) The gate is automatically scaled up during training to achieve controllable and gradual enhancement injection; 3) The output is a set of multi-scale features after alignment and enhancement, which is used by the detection head.

[0013] Step 7: Decoupled Detector Prediction: The enhanced features of P3, P4, and P5 are fed into the decoupled detector: 1) The classification branch and the regression branch are modeled independently. If necessary, lightweight channel / spatial attention micro-units are inserted before the branches to enhance discriminability; 2) The last layer is kept as a convolutional output to be compatible with the detection bias and decoding process; 3) The outputs of each scale are summarized to complete the unified decoding and merging of the bounding boxes and categories.

[0014] Step 8: Training and Optimization: An end-to-end training paradigm based on single-stage detection is adopted: 1) The loss includes classification, localization, and target-oriented components (IoU series and confidence strategies); 2) The AdamW optimizer is used with an appropriate learning rate plan and weight decay, combined with EMA for stable training; 3) It is used in conjunction with conventional data augmentation to cover scenarios such as small targets, occlusion, and reflections. In this setup, gated residuals and graph bridging progressively align statistics and gradient flow, improving convergence stability and performance in high IoU intervals.

[0015] Step 9: Inference and Evaluation: Perform inference on the test set, and use confidence thresholding and NMS / improved NMS for post-processing; calculate mAP50 and mAP50-95, parameter count and memory usage, and compare with the baseline model; when the evaluation metrics no longer improve or the convergence condition is met, end the training and solidify the model.

[0016] The batch input model in step two consists of three main parts: Backbone+Neck, Graph FusionCore, and Dual-Branch Head. Backbone+Neck generates multi-scale features P3 / P4 / P5 from bottom to top. MHSA-Darknet is connected in series at P5 to introduce global context, and the multi-scale channels are unified through Concat & Conv. GraphFusion Core consists of two paths: DGNAdapter and Cross-scale. The former performs dynamic adjacency graph construction and graph alignment within each scale, while the latter establishes sparse bridges between adjacent scales to directionally transmit complementary information. The outputs of both are fused through Gate-Residual-Fusion gated residual fusion to obtain aligned and enhanced multi-scale features. Dual-Branch Head sets up classification branches and bounding box regression branches at each scale, and decodes and merges multi-scale predictions through Decode & Merge, finally outputting detection boxes and category results.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] This invention employs dynamic adjacency adaptation and cross-scale graph bridging to simultaneously extract the spatial global features and local details of PPE targets in construction scenarios. A high-level multi-head self-attention (MHSA-Darknet) approach is introduced at the P5 level to capture long-range dependencies. Before entering the decoupled detection head, gated residuals are used for steady-state fusion to reduce mismatch and propagation errors caused by cross-layer information transmission. This scheme significantly improves detection accuracy in small target / densely occluded scenarios while maintaining high IoU accuracy (mAP50–95) and reducing model size and memory usage.

[0019] In PPE detection, the ability to acquire global and local features, robustness in dense scenes, and lightweight design are key factors affecting performance. This invention improves overall accuracy and reduces memory usage without significantly increasing parameters through DGNAdapter's same-layer "mapping + diffusion" and cross-scale binary bridging directional transfer. Simultaneously, gated residuals suppress noise and over-smoothing, providing better fitting capabilities for reflective interference, fine edges, and minor local changes, thereby improving detection performance for small targets and highly crowded scenes, achieving stable gains even with small target sizes.

[0020] This invention addresses challenges such as the easy submersion of small targets in complex environments, mutual interference among dense targets, distortion in cross-scale semantic transmission, and real-time constraints at the edge. It proposes a decoupled detection method based on dynamic adjacency learning and cross-scale graph bridging on multi-scale features: Topological alignment is achieved within the same layer through adaptive graph construction based on appearance and geometric priors; cross-layer semantics are injected into higher levels through constrained bridging and gated residual steady-state injection; finally, classification and regression are optimized separately in the decoupled head. This significantly improves recall and accuracy under conditions of strong reflection, occlusion, and compressed noise, while maintaining lightweight and real-time performance. This technology is particularly suitable for construction site safety inspections (safety helmets / reflective vests and other protective gear).

[0021] The proposed method for cross-scale graph bridging and decoupling target detection based on dynamic adjacency is based on a three-stage detection architecture of Backbone–Neck–Head. Backbone & Neck output three-scale feature maps P3, P4, and P5; a high-level self-attention module (MHSA-Darknet) is connected at P5 to obtain global semantics; dynamic adjacency adapters (DGNAdapter) are connected at each scale of P3 / P4 / P5 to complete same-layer graph construction and alignment; a cross-scale graph bridging module is connected between adjacent scales to "directively" transmit complementary information through a binary relationship; the enhancement results are steadily fused back to the features of each scale through gated residuals; finally, they are sent to the decoupled detection head (Dual-Branch) to complete classification and regression prediction respectively. The overall process emphasizes "alignment first, bridging second, and injection third" to avoid mismatch and oscillation caused by simple stacking. Attached Figure Description

[0022] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a system structure diagram of the cross-scale graph bridging and decoupling target detection method based on dynamic adjacency of the present invention.

[0024] Figure 2 This is a structural diagram of the cross-scale graph bridging and decoupling target detection method based on dynamic adjacency of the present invention.

[0025] Figure 3 This is a structural diagram of the MHSA-Darknet backbone network in the cross-scale graph bridging decoupling target detection method based on dynamic adjacency of the present invention.

[0026] Figure 4 This is a structural diagram of the Dynamic Graph Adjacency Adapter in the cross-scale graph bridging and decoupling target detection method based on dynamic adjacency of the present invention.

[0027] Figure 5 This is a comparative analysis diagram of the present invention. Detailed Implementation

[0028] 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.

[0029] Reference Figure 1 A cross-scale graph bridging decoupling target detection method based on dynamic adjacency (DynaBridge-YOLO) includes the following steps:

[0030] Step 1: Data Preprocessing and Segmentation: Obtain the raw data, set the input resolution and category mapping according to the PPE task requirements, and standardize the annotation; clean and segment the raw data to form training set, validation set and test set; the data processing and segmentation part is at the entrance of the structure of this invention, responsible for performing preliminary processing on the raw data to form the data structure required by the prediction model.

[0031] Step 2: Batch Sampling and Multi-Scale Feature Extraction (Backbone–Neck): 16 images are randomly selected from the training set each time as a batch input model. The images are then processed through the backbone network and pyramid path to obtain P3, P4, and P5 three-scale features. The spatial dimensions and stride of each scale are recorded for subsequent bridging and alignment.

[0032] like Figure 2 The overall structure of the algorithm of this invention is shown in the figure. The model consists of three main parts: Backbone+Neck, Graph Fusion Core, and Dual-Branch Head. The Backbone+Neck generates multi-scale features P3 / P4 / P5 from bottom to top. MHSA-Darknet is connected in series at the P5 end to introduce global context, and the multi-scale channels are unified through Concat & Conv. The Graph Fusion Core consists of two paths: DGNAdapter and Cross-scale. The former performs dynamic adjacency graph construction and graph alignment within each scale, while the latter establishes sparse bridges between adjacent scales to directionally transmit complementary information. The outputs of both are fused through Gate-Residual-Fusion gated residual fusion to obtain aligned and enhanced multi-scale features. The Dual-Branch Head sets up classification branches and bounding box regression branches at each scale, and decodes and merges multi-scale predictions through Decode & Merge, finally outputting detection boxes and category results.

[0033] Step 3: High-level self-attention enhancement: Input P5 into the high-level multi-head self-attention module MHSA-Darknet to obtain the global context and long-range dependencies, such as... Figure 3 As shown; P3 and P4 retain the original statistics;

[0034] The formula for the high-level multi-head self-attention module in step three is as follows:

[0035]

[0036]

[0037] Where X5 is the input feature of the high-level (P5) stage in the third step; W Q To map X5 to a trainable matrix for the query, W K To map X5 to a trainable matrix of keys, W V Let X5 be a trainable matrix that maps X5 to values; Q is the training matrix of X5 mapped by W. Q The query obtained from the mapping, K is X5 via W K The mapped key, V is X5 via W V The value obtained from the mapping; d is the channel dimension of a single head; softmax is the normalization operation on the key dimension after scaling QKᵀ; Concat is the concatenation of the outputs of each attention head on the channel dimension; W O is the linear remapping matrix of the splicing result; MHSA(X5) is the output after applying multi-head self-attention to X5;

[0038] Wherein, the output is denoted as ; , Without global attention, keep the original statistics and record... = , = .

[0039] like Figure 3 As shown, the MHSA-Darknet backbone network structure of this invention is presented. The model adopts a bottom-up approach, consisting of...

[0040] And connected in series at the highest level To introduce a global context.

[0041] Step 4: Dynamic Adjacency Learning and Alignment at the Same Layer: Perform the following steps on P3, P4, and P5 respectively: 1) Generate node embeddings and scale position embeddings, and determine local windows or k-nearest neighbors as candidate neighborhoods; 2) Adaptively generate sparse adjacencies based on feature compatibility and spatial priors; 3) Perform lightweight graph propagation on these adjacencies and output the same-layer alignment features.

[0042] The formula for dynamic adjacency learning and alignment in step four is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] in, The feature flattening matrix at this scale ( ), For the i-th position feature, Embed it; Normalized coordinates prior; Indicates node similarity. Temperature coefficient; Let i be the neighborhood of i; In order to within this neighborhood Sparse adjacency obtained by performing softmax; For degree matrix, For symmetric normalized adjacency; , For learnable projection, It is a non-linear activation.

[0048] Step 5: Cross-scale graph bridging: Perform the following between the two pairs of adjacent scales P5-P4 and P4-P3: 1) Align the resolution and channel dimension through upsampling / downsampling and linear mapping; 2) Establish sparse binary bridging relationships based on position and semantic compatibility; 3) Propagate the high-level global semantics down to the low-level along the bridging relationship and propagate the low-level details up to the high-level to obtain cross-layer enhanced features;

[0049] How to write cross-level aggregation:

[0050]

[0051] in, This refers to the enhanced output features of the l-th layer; Add a self-loop and normalize the dynamic graph adjacency matrix to the l-th layer; The input features are those of the l-th layer; γ is the linear mapping / 1×1 convolution weight of the same layer message; γ is the gating coefficient of cross-layer injection; IB(l↔l+1) is the matching / gating matrix of cross-scale binary bridging; X' is the adjacent scale (l+1) feature after alignment with the l-th layer; Linear mapping / 1×1 convolution weights injected across layers.

[0052] Step Six: Gated Residual Steady-State Injection: The output of "same-layer alignment + cross-layer enhancement" is fused with the original scale features using gated residuals: 1) The gate parameters are initialized negatively, with an approximate identity mapping in the early stages of training; 2) The gate is automatically scaled up during training to achieve controllable and gradual enhancement injection; 3) The output is a set of multi-scale features after alignment and enhancement, which is used by the detection head.

[0053]

[0054] in Negative initialization, approximate identity mapping in the early training phase, followed by gradual scaling; output is a multi-scale feature set after alignment enhancement. .

[0055] like Figure 4 As shown, the Dynamic Graph Adjacency Adapter (DGNAdapter) consists of Dynamic Graph Attention and Adaptive Graph Convolutional. The DGNAdapter is the "relationship modeling and alignment" unit within the same scale of this method. Based on the current features and simple geometric priors, it automatically selects the most relevant neighborhood for each location and assigns weights, essentially building a more reliable local relationship network on the feature map temporarily. This relationship network is then used to perform a lightweight update and calibration of the original features, making features at the same level more consistent semantically and statistically, and easier for subsequent modules to utilize. The goal of this adapter is not to increase the fusion strength, but rather to first clarify the relationships within the same level and align the features, laying a solid foundation for cross-scale information transfer. The connection between the nine steps is as follows: In step three, the DGNAdapter generates a relationship enhancement output within the same layer. This output serves as both input and matching prior for the cross-scale binary bridging in steps four / five, helping the bridging layer to only transmit complementary information from the "matching" layers. Furthermore, it acts as a gated residual steady-state injection into step six as a branch within the same layer, undergoing controlled fusion with cross-layer information to reduce statistical mismatch and training oscillations. In subsequent steps seven / eight, the decoupled detection head directly consumes this stable, aligned feature to complete classification and regression. Overall, the DGNAdapter strings together "same-layer alignment → cross-scale bridging → gated injection" into a closed loop: first stabilizing the same layer, then aligning cross-layers, and finally fusing, thereby improving accuracy and convergence stability with low computational cost, especially beneficial for small targets and complex backgrounds. The formula for Dynamic Graph Attention is shown below:

[0056]

[0057] in, For input features, Embedded for nodes, For normalized coordinate priors, For the neighborhood, For temperature coefficient, For learnable parameters, This is the output of attention for a single-layer dynamic graph.

[0058] The formula for Adaptive Graph Convolutional is as follows:

[0059]

[0060] in, , Embed two sets of independent nodes. For the adaptive matrix generated based on the embedding, To normalize adjacency, For learnable parameters, This is the output of the convolution.

[0061] Step 7: Decoupled Detection Head Prediction: The enhanced features of P3, P4, and P5 are fed into the decoupled detection head: 1) The classification branch and the regression branch are modeled independently. If necessary, lightweight channel / spatial attention micro-units are inserted before the branches to enhance discriminability; 2) The last layer is kept as a convolutional output to be compatible with the detection bias and decoding process; 3) The outputs of each scale are summarized to complete the unified decoding and merging of the bounding boxes and categories; the unified decoding and merging of multi-scale outputs yields the final detection result.

[0062] Step 8: Training and Optimization: An end-to-end training paradigm based on single-stage detection is adopted: 1) The loss includes classification, localization, and target-oriented components (IoU series and confidence strategies); 2) The AdamW optimizer is used with an appropriate learning rate plan and weight decay, combined with EMA for stable training; 3) It is used in conjunction with conventional data augmentation to cover scenarios such as small targets, occlusion, and reflections. In this setup, gated residuals and graph bridging progressively align statistics and gradient flow, improving convergence stability and performance in high IoU intervals.

[0063] like Figure 5 The results of eight methods on two real-world PPE datasets are shown under the same training configuration, with metrics including Precision (P), Recall (R), mAP50, mAP50–95, and number of parameters. For each metric, the best-performing results are highlighted in bold, with the best result from the present invention, DynaBridge-YOLO, highlighted in bold red.

[0064] Dataset 1: DynaBridge-YOLO is comparable to YOLOv8-L at mAP50=0.587, but has only 3.44M parameters, a reduction of approximately 92.1% compared to YOLOv8-L and approximately 69.1% compared to YOLOv8-S; it shows a 0.024 improvement (+4.26%) in mAP50 relative to YOLOv8-N (0.563). At mAP50–95, it is roughly on par with the lightweight baseline (0.282), slightly lower than YOLOv8-L (0.289).

[0065] Dataset 2: DynaBridge-YOLO achieved the best mAP50 score of 0.698 across the entire dataset, a 0.003 improvement (+0.43%) compared to the best YOLO baseline (YOLOv8-S, 0.695). Simultaneously, its parameter count was only 3.52M, significantly smaller than models with 11.13M (–approximately 68%), 43.61M (–approximately 92%), and 68.13M (–approximately 95%). The mAP50–95 scores were roughly equivalent (0.489 vs 0.494). It achieved the highest precision score of 0.626, while its recall was slightly lower than some larger models (0.692 vs 0.75±).

[0066] On two datasets, this invention achieves performance comparable to or better than that of large models with a significantly smaller parameter scale while maintaining or improving the main metrics (especially mAP50). This verifies that the collaborative design of "same-layer dynamic graph adaptation + cross-scale bridging + gating fusion" has stable gains and good engineering usability under small-scale conditions.

[0067] Step 9: Inference and Evaluation: Perform inference on the test set, and use confidence thresholding and NMS / improved NMS for post-processing; calculate mAP50 and mAP50-95, parameter count and memory usage, and compare with the baseline model; when the evaluation metrics no longer improve or the convergence condition is met, end the training and solidify the model.

[0068]

[0069]

[0070]

[0071]

[0072] Where mAP@0.5:0.95 is the average of multiple IoU thresholds (0.50 to 0.95, with a step size of 0.05).

[0073] This invention proposes a collaborative mechanism of "same-layer dynamic graph adaptation + cross-scale binary bridging + gated residual steady-state injection" to align feature statistics and gradient flow across multiple scales (P3 / P4 / P5), transforming the high-level self-attention (MHSA-Darknet) and decoupled detection head (Dual-Branch) from "simple superposition" to "controlled collaboration." Specifically: 1) The DGNAdapter adaptively learns sparse adjacencies at each scale and completes same-layer graph alignment; 2) Cross-scale binary bridging transmits complementary information based on position and semantic orientation; 3) Gated residuals are negatively initialized, starting from approximate identity and gradually increasing in size to avoid early perturbations. This mechanism significantly improves mAP50 and mAP50-95 with small datasets and reduces memory / computational overhead; and its stable performance is verified by comparing training budgets and parameter sizes on datasets with small PPE targets and dense occlusion scenes.

[0074] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A dynamic-adjacency-based cross-scale graph bridging decoupled object detection method, characterized in that: Comprising the following steps: Step one: data preprocessing and segmentation: obtain the original data, set the input resolution and category mapping, and label normalization; clean and segment the original data to form the training set, validation set and test set; Step two: batch sampling and multi-scale feature extraction: randomly select images from the training set as a batch input model, get P3, P4, P5 three scale features through the main network and pyramid path; record the spatial size and step of each scale; Step three: high-level self-attention enhancement: input P5 into the high-level multi-head self-attention module MHSA-Darknet to obtain global context and long-range dependence; P3, P4 keep the original statistics; Step four: same layer dynamic adjacent learning and alignment: respectively on P3, P4, P5: Generate node embedding and scale position embedding, determine the local window or k-nearest neighbor as the candidate neighborhood; Adaptively generate sparse adjacency according to feature compatibility and spatial priori; Lightweight graph propagation on the adjacency, output the same layer alignment feature; Step five: cross-scale graph bridging: between P5-P4, P4-P3 two pairs of adjacent scales: Align the resolution and channel dimension through up / down sampling and linear mapping; Establish a sparse two-part bridge relationship according to the position and semantic compatibility; Downlink the high-level global semantics to the low layer, uplink the low-level details to the high layer along the bridge relationship, and get the cross-layer enhanced feature; Step six: gated residual steady-state injection: fuse the output of "same layer alignment + cross layer enhancement" with the original scale feature through gated residual: The gating parameter is initialized in a negative direction, and it is approximately an identity mapping in the early stage of training; With the training, the gating automatically increases, realizing controllable and gradual enhancement injection; The output is a set of multi-scale features after alignment and enhancement, which is used for the detection head; Step seven: decoupled detection head prediction: send the enhanced features of P3, P4, P5 into the decoupled detection head respectively: The classification branch and the regression branch are independently modeled, and lightweight channel / space attention micro units are inserted before the branch to enhance the discriminability; Keep the last layer as a convolutional output, compatible with detection bias and decoding process; Collect the outputs of each scale to complete the unified decoding and merging of boxes and categories; Step eight: training and optimization: use the end-to-end training paradigm of single-stage detection: The loss includes classification, positioning and target parts; Use AdamW optimizer and appropriate learning rate plan, weight decay, and EMA to stabilize training; Combined with conventional data augmentation, covering small targets, occlusion and reflection scenarios; Step nine: reasoning and evaluation: perform reasoning on the test set, use confidence threshold and NMS / improved NMS for post-processing; statistics mAP50 and mAP50-95, parameter quantity and memory occupation, and compare with baseline model; when the evaluation index no longer improves or meets the convergence condition, end the training and solidify the model.

2. The method of claim 1, wherein: The batch input model in the second step is composed of three parts: Backbone+Neck, Graph Fusion Core and Dual-Branch Head; wherein Backbone+Neck generates multi-scale features P3 / P4 / P5 from bottom to top, introduces global context through MHSA-Darknet at the P5 end, and unifies the multi-scale channels through Concat&Conv; Graph Fusion Core is composed of DGNAdapter and Cross-scale two paths, the former builds a dynamic adjacent graph and graph alignment in each scale, and the latter builds a sparse bridge between adjacent scales to direct the transmission of complementary information, and the two outputs are aligned and enhanced through Gate-Residual-Fusion gate residual fusion; Dual-Branch Head sets classification branch and bounding box regression branch at each scale, and finally outputs the detection frame and class results through Decode&Merge unified decoding and merging multi-scale prediction.

3. The method of claim 1, wherein: The formula of the high-level multi-head self-attention module in the third step is: ； ； wherein X5 is the high-level input feature in the third step; W Q is a trainable matrix for mapping X5 to the query; K is a trainable matrix for mapping X5 to the key; V is a trainable matrix for mapping X5 to the value; Q is the query obtained by mapping X5 through W Q ; K is the key obtained by mapping X5 through W K ; V is the value obtained by mapping X5 through W V ; d is the channel dimension of the single head; softmax is the normalization operation on the key dimension after scaling QKᵀ; Concat is the concatenation of the outputs of each attention head in the channel dimension; W O is the linear remapping matrix of the concatenation result; and MHSA(X5) is the output obtained by applying multi-head self-attention to X5. Wherein, the output is denoted as ; , Without global attention, keep the original statistics, record = , = .

4. The method of claim 1, wherein: The formula of dynamic adjacent learning and alignment in the fourth step is: ； ； ； ； where, is the flattened matrix of features for this scale, , is the i-th position feature, is its embedding; is the normalized coordinate prior; denotes the node similarity, is the temperature coefficient; is the neighborhood of i; is the sparse adjacency obtained by doing softmax on in this neighborhood; is the degree matrix, is the symmetric normalized adjacency; , is the learnable projection, is the non-linear activation.

5. The method of claim 1, wherein: The formula of cross-layer enhanced features in the fifth step is: ； wherein, is the enhanced output feature of the l-th layer; is the dynamic graph adjacency matrix with self-loop and normalization for the l-th layer; is the input feature of the l-th layer; is the linear mapping / lxl convolution weight for the same layer message; γ is the gating coefficient for cross-layer injection; IB(l↔l+1) is the matching / gating matrix for cross-scale binary bridging; is the adjacent scale L+1 feature aligned with the l-th layer; is the linear mapping / lxl convolution weight for cross-layer injection.

6. The method of claim 1, wherein: The formula of gate residual fusion in the sixth step is: ; wherein negative initialization, approximate identity mapping at the beginning of training, followed by gradual increase The output is a set of multi-scale features with alignment enhancement .

7. The method of claim 1, wherein: The first step specifically comprises: obtaining original data, setting input resolution and class mapping according to PPE task requirements, and labeling normalization; cleaning and segmenting the original data to form a training set, a validation set and a test set; the data processing and segmentation part is at the entrance of the structure of the application, responsible for the preliminary processing of the original data to form the data structure required by the prediction model.

Images