Self-adaptive down-sampling aerial material image detection method and system and storage medium

Abstract

The invention relates to the field of computer vision, and discloses a self-adaptive down-sampling aerial material image detection method and system and a storage medium. The method comprises the following steps: acquiring an aerial material image and preprocessing the aerial material image; the method comprises the following steps: extracting features through an improved Backbone module, carrying out adaptive down-sampling by adopting an LAWDS module, and extracting multi-path features by utilizing an RGCSPELAN module; the multi-layer feature map is input into an improved Encoder module, feature transmission and aggregation are performed through an FPN + PAN structure, and channel alignment and bidirectional guide fusion are realized based on a CGFM module; the fusion features are input into a Transform decoder, and a detection result is output; effective detection results are obtained through confidence coefficient screening and NMS processing; carrying out aerial material quantity statistics based on a detection result; and comparing with a warehousing system and updating the inventory. According to the method, the problems of aerial material small target feature loss and poor dense target distinction degree are solved, and the detection precision is improved.

Description

Technical Field

[0001] This application relates to the field of computer vision, and in particular to an adaptive downsampling method, system, and storage medium for detecting aircraft material images. Background Technology

[0002] In aircraft material warehouse management, to achieve accurate inventory counting and real-time status monitoring, drones or fixed cameras are typically used to collect images of aircraft materials, and deep learning object detection models are used for automatic identification and statistics. However, the aircraft material warehouse environment is characterized by high shelves, dense material stacking, small target size, and uneven lighting, posing a severe challenge to the accuracy and robustness of the detection model.

[0003] Currently, deep learning-based object detection models (such as the YOLO series and DEIM-HGNetV2-N) have been attempted for automated detection in aircraft material warehouses. These models typically employ fixed-stride convolution and pooling operations for feature downsampling, achieve multi-scale feature fusion through a feature pyramid network, and finally output the recognition result through a detection head. However, in the specific application scenario of aircraft material warehouses, existing technologies have significant shortcomings. First, traditional downsampling methods lack adaptive mechanisms when compressing feature maps, treating different regions of the image indiscriminately, resulting in severe loss of detailed features and edge information of small-sized aircraft materials in shallow networks. Second, existing feature fusion structures often use simple concatenation or addition operations, failing to fully consider the semantic differences and contextual relationships between features of different scales. When dealing with complex situations where aircraft materials are densely stacked and mutually occluded, it is difficult to achieve effective complementarity between semantic features and detailed features. In addition, existing models often struggle to balance lightweight design with detection accuracy; either the computational complexity is too high to meet real-time requirements, or excessive lightweighting sacrifices the ability to detect small targets. These shortcomings collectively lead to problems in the practical application of existing detection methods in aircraft material warehouses, such as high false negative rates for small targets, poor differentiation of dense targets, and insufficient robustness in complex environments.

[0004] To address the aforementioned issues, this application introduces a lightweight LAWDS module into the backbone network, employs a multi-path aggregation RGCSPELAN module in the feature extraction stage, and adds a CGFM module in the feature fusion stage. This solves the technical challenges of easily losing small target features and poor distinguishability of dense targets in aircraft material images during the acquisition process, thereby improving the detection recall rate of small targets in aircraft materials and the recognition accuracy in complex stacking scenarios. Summary of the Invention

[0005] This application provides an adaptive downsampling method, system, and storage medium for detecting aircraft materials images, which solves the technical problems of easy loss of small target features and poor discrimination of dense targets in aircraft material images, and improves the detection recall rate of small targets in aircraft materials and the recognition accuracy in complex stacking scenarios.

[0006] In a first aspect, this application provides an adaptive downsampling method for detecting aircraft material images, the method comprising: Step S101: Acquire initial images of aviation materials using image acquisition equipment deployed in the aviation material warehouse, preprocess the initial images of aviation materials, and generate standardized images of aviation materials; Step S102: Input the standardized aircraft material image into the improved Backbone module. The improved Backbone module has built-in LAWDS module and RGCSPELAN module. The LAWDS module retains the small target and edge feature information of the aircraft material, and the RGCSPELAN module extracts the multi-path features of the aircraft material. After integration, a multi-layer feature map containing features of different scales of the aircraft material is obtained. Step S103: Input the multi-layer feature map into the improved Encoder module. The improved Encoder module has a built-in FPN+PAN structure and CGFM module. The basic features of the multi-layer feature map are passed and aggregated hierarchically through the FPN+PAN structure. At the same time, based on the CGFM module, the two-way aviation material features in the hierarchical transmission process are aligned, contextual attention modeled and bidirectional guided fusion is performed to obtain the aviation material fusion features with semantic and detail coordination. Step S104: Input the fused features of the aircraft materials into the Transformer decoder, output the prediction information directly related to the aircraft material target, and then accurately match the prediction information with the real annotation to obtain the aircraft material target detection result. The aircraft material target detection result includes bounding box coordinates, category and confidence level. Step S105: Filter the detection results of the aircraft material target according to the preset confidence rules, and perform deduplication processing on the filtered results through NMS operation to obtain valid detection results; Step S106: Based on the effective detection results, obtain the statistical results of the number of aircraft materials by using preset mapping rules and location association rules, combined with the stacking layer calculation method; Step S107: Compare the statistical results of the aircraft material quantity with the inventory data of the warehouse management system, update the warehouse inventory information according to the difference classification processing rules, and generate a report.

[0007] Secondly, this application provides an adaptive downsampling aircraft material image detection system for implementing the aforementioned adaptive downsampling aircraft material image detection method, the system comprising: The image acquisition module is used to acquire initial images of aviation materials through image acquisition equipment deployed in the aviation material warehouse, and to preprocess the initial images of aviation materials to generate standardized images of aviation materials; The feature extraction module is used to process the standardized aircraft material image through the improved Backbone module and output a multi-layer feature map; The feature fusion module is used to input the multi-layer feature map into the improved Encoder module to obtain the semantic and detail-coordinated aircraft material fusion features; The target detection module is used to input the fused features of the aircraft materials into the Transformer decoder, output the prediction information directly related to the aircraft material target, and accurately match the prediction information with the real annotation to obtain the aircraft material target detection result including bounding box coordinates, category and confidence level; The result optimization module is used to deduplicate the detection results of the aircraft material targets through preset confidence threshold rules and NMS operations to obtain effective detection results; The quantity statistics module is used to obtain the quantity statistics results of aviation materials based on the valid detection results, through preset mapping rules and location association rules, and in combination with the stacking layer calculation method; The inventory update module is used to compare the statistical results of the aircraft material quantity with the inventory data of the warehouse management system, update the warehouse inventory information and generate a report based on the difference classification processing rules.

[0008] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, characterized in that the program, when executed by a processor, implements the aforementioned adaptive downsampling method for detecting aircraft material images.

[0009] This application proposes an adaptive downsampling method, system, and storage medium for aircraft material image detection. It solves the technical challenges of easily losing features of small targets and poor distinguishability of dense targets in aircraft material images, improving the recall rate for detecting small targets and the recognition accuracy in complex stacking scenarios. Compared with existing technologies, the beneficial effects of this application's technical solution are at least as follows: First, by introducing a lightweight adaptive weighted downsampling module, dynamic focusing on key areas of the image is achieved, effectively preserving the detailed features and edge information of small-sized aircraft materials, and alleviating the feature loss problem caused by traditional fixed downsampling methods.

[0010] Second, the RGCSPELAN module based on multi-path aggregation is adopted, which enhances the model's ability to express multi-scale features through channel splitting and multi-path convolution structure, improving the richness and robustness of feature extraction while ensuring lightweight design.

[0011] Third, the context-guided fusion module enables adaptive interaction and complementarity of cross-scale features, enhancing the model's ability to distinguish densely stacked aircraft materials and improving the detection recall and accuracy in complex scenarios. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an adaptive downsampling method for detecting aircraft material images according to this application. Figure 2 This is a schematic diagram of the adaptive downsampling process performed by the LAWDS module in this application; Figure 3 This is a schematic diagram of the process for extracting multi-path features of aircraft materials using the RGCSPELAN module in this application; Figure 4 This is a schematic diagram of the structure of an adaptive downsampling aircraft material image detection system according to this application. Detailed Implementation

[0014] This application provides an adaptive downsampling method, system, and storage medium for aircraft material image detection. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of an adaptive downsampling aircraft material image detection method in this application includes: Step S101: Acquire initial images of aviation materials using image acquisition equipment deployed in the aviation material warehouse, preprocess the initial images of aviation materials, and generate standardized images of aviation materials.

[0016] In one specific embodiment, step S101 may specifically include the following steps: Initial images of the aircraft materials are acquired by using a drone or a fixed camera, and synchronized images of the aircraft materials are obtained by synchronizing the initial images of the aircraft materials with image frames. The synchronized images of the aircraft materials are uniformly converted to JPG format, and the size is adjusted to a preset pixel value to obtain a unified image of the aircraft materials. The uniform image of the aircraft materials is subjected to brightness adjustment and noise reduction processing to obtain an enhanced image of the aircraft materials; Location tags are added to the enhanced image of the aircraft material, and then linked to the location database in the warehouse management system to generate the standardized image of the aircraft material.

[0017] Specifically, the image acquisition equipment deployed in the aircraft material warehouse includes multi-rotor drones and high-definition network cameras. The drones are equipped with a three-axis mechanical stabilization gimbal and a GPS+IMU dual-mode positioning module. Their battery life meets the needs of single-time full-coverage inspections in small and medium-sized aircraft material warehouses. Fixed cameras are deployed at key locations such as warehouse pillars and shelf beams, forming a three-dimensional acquisition network. The image acquisition equipment collects initial images of the aircraft materials according to a preset "Z" shaped flight path or fixed points. After the drone completes shooting of each shelf unit, it hovers to synchronize image frames, eliminating missed or duplicate shots and obtaining synchronized images of the aircraft materials. To address the issues of inconsistent formats and sizes in the synchronized images, all JPG images captured by the drones and PNG images captured by the fixed cameras are converted to JPG format and resized to 640×640 pixels to create unified images of the aircraft materials, ensuring consistent specifications for subsequent model input data. Considering the uneven lighting and noise interference in the aircraft material warehouse, the Retinex algorithm is used to adjust the brightness of the unified image of the aircraft materials, increasing the brightness of the shaded areas of the shelves and suppressing the brightness of the bright areas. Gaussian filtering is then used to remove image noise, generating an enhanced image of the aircraft materials. This solves the problem of lighting changes affecting detection accuracy in existing technologies. Based on the positioning information during image acquisition, location tags are added to the enhanced image of the aircraft materials. The tag information is linked to the location database of the warehouse management system to clarify the storage location of the aircraft materials according to the "shelf partition-shelf number-layer-column" sequence, ultimately generating a standardized image of the aircraft materials. Throughout the data processing, the initial image of the aircraft materials is synchronized to a synchronized image, the synchronized image is standardized in format and size, the unified image is adjusted for brightness and denoising, and the enhanced image is added with location tags and linked to the database to become a standardized image of the aircraft materials. Data at each stage corresponds one-to-one, and key information is preserved during transmission. This process solves the problems of detection errors caused by environmental factors and inconsistent data specifications in existing technologies, such as manual counting or barcode recognition. Through standardized processing, the images of aircraft materials meet the requirements of subsequent model detection, providing high-quality data input for the feature extraction stage. At the same time, the association between location tags and the warehouse database lays the foundation for the subsequent docking of detection results with actual inventory.

[0018] Step S102: Input the standardized aircraft material image into the improved Backbone module. The improved Backbone module has built-in LAWDS and RGCSPELAN modules. The LAWDS module retains the small target and edge feature information of the aircraft material, while the RGCSPELAN module extracts the multi-path features of the aircraft material. After integration, a multi-layer feature map containing features of different scales of the aircraft material is obtained.

[0019] In one specific embodiment, step S102 may specifically include the following steps: Input the standardized aircraft material image into the improved Backbone module; The standardized aircraft material image is adaptively downsampled using the LAWDS module to retain small targets and edge feature information of the aircraft material and obtain a high-fidelity feature map. The RGCSPELAN module is used to extract multi-path features from the fidelity feature map to obtain multi-path output features. By integrating the multi-path output features based on a multi-scale semantic fusion mechanism, a multi-layer feature map containing features of different scales of aviation materials is obtained.

[0020] Specifically, please refer to Figure 2 The standardized aircraft material images are input into the improved Backbone module, which incorporates the LAWDS and RGCSPELAN modules. These modules work together to perform feature processing and output multi-layer feature maps. The standardized aircraft material images first enter the LAWDS module, which performs adaptive downsampling processing in three steps: attention weight generation, feature space reorganization and downsampling, and weighted fusion output. In the attention weight generation stage, average pooling is first performed on the feature map corresponding to the standardized aircraft material image. Then, local context information is extracted through 1×1 convolution to obtain the spatial weight matrix. Subsequently, a rearrange operation divides the feature map into four 2×2 sub-regions. An adaptive weight is assigned to each sub-region using a Softmax operation. During weight allocation, priority is given to strengthening the feature weights corresponding to small targets and edge regions of the aircraft material, addressing the technical problem of information loss for small targets and edges caused by existing fixed downsampling methods. In the feature space reconstruction and downsampling stage, the LAWDS module uses 3×3 grouped convolutions with a stride of 2 to spatially compress the feature map, while simultaneously increasing the number of output channels by a factor of 4 to match the four sub-region divisions. Then, a rearrange operation converts the output into a tensor structure of [batch, channel, height, width, 4], so that each sub-region corresponds to the features at the four sampling positions (top left, top right, bottom left, and bottom right) in the aircraft material image. This grouped convolution design reduces computational overhead while achieving downsampling, adapting to the lightweight requirements of UAVs and warehouse edge devices, and solving the problem of lack of adaptive mechanisms and computational redundancy in existing feature downsampling technologies. In the weighted fusion output stage, the generated attention weights are multiplied by the features of the four sub-regions, and the summation is performed on the last dimension to achieve weighted fusion. If there is a difference in the number of input and output channels, linear projection alignment is performed using 1×1 convolutions. Finally, a high-fidelity feature map is obtained, preserving the small target and edge feature information of the aircraft material. This process "re-stitches" the spatial sampling results through adaptive weights, avoiding information loss caused by the traditional MaxPool method which only takes the maximum value, thus improving feature fidelity.

[0021] Please see Figure 3The high-fidelity feature map is then input into the RGCSPELAN module, which performs multi-path feature extraction in five steps: dynamic channel splitting, main path feature extraction, additional path expansion, terminal feature refinement, and feature concatenation and fusion. In the dynamic channel splitting stage, the high-fidelity feature map is first split into two parts by a 1×1 convolution. One part retains its original structure for direct connection, while the other part enters the subsequent feature extraction process. Channel splitting reduces the number of channels in the subsequent 3×3 convolution, reducing computational redundancy and addressing the limited multi-scale information capture capability and redundant computation issues of existing HGBlock models. In the main path feature extraction stage, the split features enter the RepConv module. During the training phase, this module enhances feature representation with a multi-branch structure to ensure the capture of semantic information at different scales of the aircraft material. During the inference phase, it folds into a single 3×3 convolution, further reducing computational overhead and balancing model accuracy with lightweight requirements. In the additional path expansion stage, the main path output sequentially passes through three cascaded 3×3 convolutional modules, aggregating multi-scale semantic features of aircraft materials layer by layer. This design borrows from the E-LAN ​​concept but is more lightweight, specifically enhancing feature discriminativeness in densely stacked aircraft materials scenarios and alleviating the problem of decreased recall in dense areas. In the terminal feature refinement stage, the terminal outputs of all paths undergo feature compression and remapping through 1×1 convolutions, enhancing the non-linear expressive power of features and making them more adaptable to subsequent fusion requirements. In the feature concatenation and fusion stage, the original directly connected path features are concatenated with the features extracted by multi-path convolutions, and then integrated through 1×1 convolutions to obtain multi-path output features. This process achieves efficient aggregation of multi-scale features of aircraft materials, avoiding the information bias caused by single-path feature extraction.

[0022] Based on a multi-scale semantic fusion mechanism, multi-path output features are integrated and divided into three levels—stride8, stride16, and stride32—according to differences in feature scale. The stride8 level retains detailed information about small targets in the aircraft materials, the stride32 level contains global semantic information about the materials, and the stride16 level falls between the two. This hierarchical division enables the classification, storage, and association of features at different scales, ultimately resulting in a multi-layer feature map containing features at different scales of the aircraft materials. Throughout the processing, standardized aircraft material images correspond to high-fidelity feature maps, high-fidelity feature maps correspond to multi-path output features, and multi-path output features correspond to multi-layer feature maps. The gradual enhancement and refinement of aircraft material features during data transmission solves the problems of information loss and weak multi-scale information capture capabilities in existing Backbone fixed downsampling methods. Furthermore, the lightweight design meets the real-time operation requirements of warehouse edge devices, providing high-quality multi-scale feature input for subsequent feature fusion in the Encoder module.

[0023] Step S103: Input the multi-layer feature map into the improved Encoder module. The improved Encoder module has a built-in FPN+PAN structure and CGFM module. The basic features of the multi-layer feature map are passed and aggregated hierarchically through the FPN+PAN structure. At the same time, based on the CGFM module, the two-way aviation material features in the hierarchical transmission process are aligned, contextual attention modeled and bidirectional guided fusion is performed to obtain the aviation material fusion features with semantic and detail coordination.

[0024] In one specific embodiment, step S103 may specifically include the following steps: The FPN structure is used to perform top-down semantic enhancement on the multi-layer feature map to obtain a semantically enhanced feature map. The semantic enhancement is based on a hierarchical transfer mechanism to pass high-level semantic information to low-level features. The semantically enhanced feature map is aggregated from bottom to top using a PAN structure to obtain preliminary fused features containing multi-scale information. The path aggregation is based on a feature fusion mechanism that feeds low-level detailed information back to high-level features. The CGFM module performs channel alignment on the two aviation material features during the hierarchical transfer process, and maps the two aviation material features to the same metric space to obtain channel-aligned feature pairs. Furthermore, the CGFM module is used to perform context attention modeling on the channel-aligned feature pairs to obtain context-recalibrated feature representations; Finally, the CGFM module performs bidirectional guided fusion on the context recalibrated feature representation to generate semantic and detail-oriented fused features for aviation materials.

[0025] Specifically, the multi-layer feature map is input into the improved Encoder module, which incorporates an FPN+PAN structure and a CGFM module. These two modules process the data sequentially according to the logic of "basic feature transfer and aggregation - feature optimization and fusion." The multi-layer feature map contains three feature layers: stride8, stride16, and stride32. The stride8 feature layer carries detailed information such as the edges and textures of small targets on the aircraft material. The stride32 feature layer contains global semantic information such as the type and overall distribution of the aircraft material. The stride16 feature layer is a transitional feature between details and semantics, and it first enters the FPN structure to perform top-down semantic enhancement. The FPN structure starts with the stride32 feature layer. First, it adjusts the number of channels to 256 through a 1×1 convolution (to adapt to the subsequent fusion dimension), and then performs a 2x upsampling operation to make the feature map resolution consistent with the stride16 feature layer. Then, it concatenates the stride16 feature layer with the stride16 feature layer in terms of channel dimension. After concatenation, it uses a 3×3 convolution to eliminate the grid artifacts introduced by upsampling, resulting in a semantically enhanced stride16 feature layer. Similarly, the semantically enhanced stride16 feature layer is concatenated with the stride8 feature layer after a 1×1 convolution and a 2x upsampling, and then convolved to obtain a semantically enhanced feature map. This process solves the problem of insufficient semantic information in low-level features in existing technologies, allowing the stride8 feature layer to retain the details of small targets while possessing category semantics, thus avoiding misjudgment of small targets due to insufficient semantics.

[0026] The semantically enhanced feature map enters the PAN structure for bottom-up path aggregation. The PAN structure starts with the stride8 feature layer, performs a 2x downsampling after a 3×3 convolution to match the feature map resolution with the stride16 feature layer, and then concatenates it with the semantically enhanced stride16 feature layer output by the FPN structure. Then, it integrates details and semantic information through a 3×3 convolution to obtain a stride16 feature layer containing dual-view information. The stride16 feature layer is then downsampled by 2x and concatenated with the stride32 feature layer output by the FPN structure, resulting in a preliminary fused feature containing multi-scale information. This process supplements the detailed information of high-level features, solves the problem of blurred contours of small targets by high-level features in existing technologies, and alleviates boundary confusion caused by the lack of details in densely stacked aircraft materials.

[0027] After the initial fusion features enter the CGFM module, the two aviation material features in the hierarchical transmission are processed first (such as the stride8 detail features transmitted by the PAN structure and the stride16 semantic features transmitted by the FPN structure). The number of channels of the two features is first detected. If the number of channels of the two features is 128 and 256 respectively (the number of channels is inconsistent), the 256-channel feature is mapped to 128 channels through 1×1 convolution, so that the two features are in the same channel dimension (128 channels) metric space, and the channel-aligned feature pair is obtained. This step solves the problem of information loss caused by the mismatch of channel dimensions during feature fusion and ensures that the two features can interact effectively.

[0028] After channel alignment, the feature pairs continue to undergo context attention modeling. The CGFM module first performs global average pooling on the feature map (256 channels) after the two features are concatenated, converting the H×W×256 feature map into a 1×1×256 channel descriptor, which reflects the global information of each channel. Then, the channel descriptor is input into a two-layer fully connected network (the first layer compresses 256 channels to 64 channels, and the second layer restores them to 256 channels). After activation by the Sigmoid function, 256 channel weight coefficients are obtained. The weight coefficient values ​​range from 0 to 1. The higher the value, the greater the contribution of the corresponding channel to the detection of aircraft materials. The concatenated feature map is weighted by these weight coefficients to obtain the feature representation of context recalibration. This process solves the problem of background noise interference during feature fusion in the prior art, strengthens the expression of aircraft material feature channels, and suppresses irrelevant background channels.

[0029] The feature representation for context recalibration is split into two paths (128 channels each) along the channel dimension, which serve as guiding weights for the two original feature paths (channel-aligned feature pairs). Let the guiding weights for the two paths be... and The original features of the two paths are and Then calculate the enhanced features , Execute again and The bidirectional residual fusion allows the two features to inject each other's contextual information. If the number of feature channels after fusion does not match the requirements of the downstream Decoder module (e.g., 256 channels), the number of channels is adjusted to 256 through 1×1 convolution to generate fusion features of aviation materials with semantic and detail synergy. This process solves the problem of poor synergy caused by static feature fusion in the existing technology, allowing detail features to carry semantic constraints and semantic features to carry detail support, thereby improving the feature discrimination of occluded aviation materials.

[0030] Throughout the process, multi-layer feature maps correspond to semantically enhanced feature maps, semantically enhanced feature maps correspond to preliminary fused features, preliminary fused features correspond to channel-aligned feature pairs, channel-aligned feature pairs correspond to context-recalibrated feature representations, and context-recalibrated feature representations correspond to aviation material fusion features. Each step in the data transmission process is optimized to address the technical defects of the previous step, ensuring that the final output aviation material fusion features meet the requirements for dense, small-target aviation material detection.

[0031] Step S104: Input the fused features of the aircraft materials into the Transformer decoder, output the prediction information directly related to the aircraft material target, and then accurately match the prediction information with the real annotation to obtain the aircraft material target detection result. The aircraft material target detection result includes bounding box coordinates, category and confidence level.

[0032] In one specific embodiment, step S104 may specifically include the following steps: The fused features of the aircraft materials are input into the Transformer decoder, and initial prediction information containing bounding box coordinates and category distribution is generated through a query mechanism; The initial prediction information is optimized by using a cross-attention mechanism to enhance the feature differentiation between aircraft material targets and obtain optimized prediction information. The Hungarian matching algorithm is used to match the optimized prediction information with the real annotations one-to-one, establish the correspondence between the prediction results and the annotation data, and obtain the matching detection results. The matched detection results are filtered by confidence level. Valid detection targets are selected based on a preset confidence threshold, and the final aircraft material target detection results, including bounding box coordinates, category, and confidence level, are generated.

[0033] Specifically, after the fused features of the aircraft materials are input into the Transformer decoder, the decoder first generates initial prediction information through a query mechanism. This mechanism includes fixed learnable queries and dynamic queries. The learnable queries are set to 100, serving as preset learnable parameters to capture the global structural features of the aircraft materials. Dynamic queries are generated based on potential target regions of the fused features, with their number adaptively adjusted according to the target density within the fused features. This is used to accurately capture the local features of small-sized, densely stacked aircraft materials. Both types of queries jointly perform multiple rounds of query operations on the fused features, which are then processed through a fully connected layer and a Softmax activation function to output the initial prediction information. This information includes the bounding box coordinates corresponding to each query (in the format of...). , representing the pixel coordinates of the top left and bottom right corners of the bounding box respectively) and the category distribution (the dimension is equal to the total number of aviation material categories in the aviation material warehouse, such as 60 dimensions, and the value of each dimension corresponds to the prediction probability under the category). This process solves the problem that the single query mechanism in the existing technology leads to insensitivity to small target capture, so that the initial prediction information can cover all potential aviation material areas in the warehouse and focus on the detailed features of small-sized aviation materials.

[0034] The initial prediction information then enters the cross-attention mechanism. Each DecoderLayer of the Transformer decoder contains self-attention and cross-attention modules. The self-attention module uses the formula:

[0035] Calculate the attention weights among different queries in the initial prediction information. ,in, For the first A query vector for each query. and The first The and the first A key vector for each query. For vector dimensions, The total number of queries is used to extract the relative positional relationships between aircraft material targets (such as the vertical hierarchy of stacked aircraft materials and the horizontal spacing between adjacent aircraft materials). This avoids feature confusion caused by the proximity of dense targets. The cross-attention module interacts with the key-value pairs of queries and the fused features of aircraft materials to further extract detailed features such as surface texture and shape edges from the fused features. This corrects the coordinate offset of the bounding box and the deviation of the class probability distribution in the initial prediction information. After cross-attention iteration optimization through 8 layers of DecoderLayer, the optimized prediction information is obtained. This step solves the problem of low feature discrimination between targets in the existing technology, so that densely stacked aircraft material targets form clear feature boundaries in the prediction information, reducing class misjudgment and bounding box overlap.

[0036] The optimized predictions and ground truth annotations are input into the Hungarian matching algorithm. The algorithm uses a weighted sum of the two types of losses as the matching cost. The ground truth annotations include the actual bounding box coordinates and actual category labels of the aircraft material targets. The bounding box regression loss uses the GIoU loss function, with the specific formula as follows: , Where A is the predicted bounding box, B is the ground truth bounding box, C is the smallest rectangle containing both A and B, and IoU is the intersection-union ratio of A and B. The classification loss uses the cross-entropy loss function (formula: Where M is the total number of categories, For real category labels, To predict class probabilities, the weighting coefficients for the two types of losses are set to 0.6 and 0.4, respectively, and the total loss value is... The algorithm constructs a bipartite graph and performs a one-to-one matching between the optimized prediction information and the real labels based on the principle of minimizing the total loss value. Redundant prediction boxes without corresponding real targets are eliminated to obtain the matched detection results. This step solves the problem of duplicate detection or missed detection caused by the chaotic allocation of detection boxes in the existing technology, ensuring that each aviation material target corresponds to only one prediction result and improving the stability of the detection results.

[0037] After matching, the detection results are then filtered based on confidence. The system has a preset confidence threshold of 0.7. The system extracts the maximum probability value from the category distribution of each detection result as the confidence value of that result. If the confidence value is ≥0.7, its bounding box coordinates, corresponding category (the category to which the maximum probability value belongs), and confidence value are retained, and it is determined to be a valid detection target. If the confidence value is in the range of 0.5 to 0.7, it is marked as "awaiting manual review" and is not included in the final result. If the confidence value is <0.5, the detection result is directly filtered. Finally, the aviation material target detection result containing the bounding box coordinates, category, and confidence value of the valid detection target is generated. This process solves the problem of low confidence prediction results interfering with inventory statistics in the existing technology, ensuring that the output detection results have high reliability and can be directly used for subsequent aviation material quantity statistics and warehouse system inventory updates.

[0038] Throughout the entire data processing chain, the fusion features of aviation materials correspond to the initial prediction information, the initial prediction information corresponds to the optimized prediction information, the optimized prediction information corresponds to the matched detection results, and the matched detection results correspond to the aviation material target detection results. Data from each link is transmitted and optimized step by step to meet the technical requirements of dense aviation material warehouses and small target detection scenarios.

[0039] Step S105: Filter the detection results of the aircraft material target using a preset confidence rule, and perform deduplication processing on the filtered results using NMS operation to obtain valid detection results.

[0040] In one specific embodiment, step S105 may specifically include the following steps: The detection results of aircraft material targets are screened according to the preset confidence level rules, and the target detection boxes with confidence levels lower than the confidence level threshold are filtered out to obtain the initial screening detection results; Traverse the detection boxes corresponding to each aviation material target in the initial screening test results, and calculate the crossover ratio between any two detection boxes; Set the overlap threshold for NMS operation, and determine whether the intersection-union ratio is greater than the overlap threshold. If so, retain the detection boxes with high confidence scores and remove the detection boxes with low confidence scores. If not, both detection boxes are retained. By integrating the detection frames and corresponding aircraft material target information after NMS operation, the deduplicated detection results are obtained; The deduplicated detection results are mapped and matched with the aircraft material codes in the warehouse management system to generate valid detection results containing aircraft material category, quantity, and location information.

[0041] Specifically, the results of aircraft material target inspection are first screened using preset confidence level rules. These rules are not a single threshold, but rather a multi-layered screening logic built by combining the inspection characteristics of different types of aircraft materials in the aircraft material warehouse with inventory management needs. The rules set differentiated confidence level ranges for different types of aircraft materials: For small-sized and easily stacked aircraft materials such as bolts and gaskets, due to the high inspection difficulty, a basic confidence level threshold of 0.65 is set, with the additional condition that "if the height of the inspection frame boundary deviates from the preset aircraft material size by ≤10%, the confidence level threshold can be lowered to 0.6"; For large-sized aircraft materials such as engine blades and fuselage frames, due to their obvious features and high inspection stability, a basic confidence level threshold of 0.75 is set, without any downward adjustment condition; For aircraft materials such as wires and hoses, which are prone to inspection deviations due to shape changes, in addition to the basic threshold of 0.7, an additional requirement is that "the intersection-union ratio (CIU) of the inspection frame and the inspection frames of adjacent aircraft materials is ≤0.3" is required for retention. When iterating through the aircraft material target detection results, the aircraft material category corresponding to each detection box is first identified, and then the filtering conditions of that category in the rule are matched. If the detection box meets both the basic threshold and the additional conditions, its bounding box coordinates and category information are retained; if it only meets the basic threshold but not the additional conditions, it is marked as "awaiting manual review"; if it does not meet the basic threshold, the detection box is directly filtered. This process yields the initial screening results. This confidence rule solves the problem that a single confidence threshold cannot be adapted to different types of aircraft material detection scenarios, avoiding the situation where small-sized aircraft materials are falsely screened due to excessively high thresholds and large-sized aircraft materials are falsely detected due to excessively low thresholds, ensuring that the initial screening results achieve a balance between accuracy and recall.

[0042] After obtaining the initial screening results, the detection boxes corresponding to each aviation material target are traversed, and the intersection-over-union (IoU) ratio between any two detection boxes is calculated. The IoU calculation is based on the bounding box coordinates of the detection boxes, and the formula is as follows: ,in The area of ​​the intersection of the two detection boxes is calculated by... Calculate (D and E represent two detection boxes, respectively) Let D be the coordinates of its top-left corner. (The coordinates of the lower right corner of D are given, and the coordinate parameters of E are similarly given). The area of ​​the union region of the two detection boxes is obtained by... Calculate, where, and This represents the area of ​​the two detection boxes. , For example, in a small area of ​​the aircraft parts shelf, the bounding box coordinates of two M8 bolt detection frames are (180, 420, 210, 450) and (185, 425, 215, 455), respectively. The preset standard height of an M8 bolt is 30 pixels. Both detection frames are 30 pixels high, and the deviation meets the supplementary rule conditions. The confidence levels are 0.63 and 0.61, respectively, both reaching the 0.6 reduction threshold for this type of aircraft parts. Therefore, both are included in the initial screening results. The intersection area of ​​the two is calculated as (210-185)×(450-425)=25×25=625. The area of ​​the first detection frame is 30×30=900, and the area of ​​the second detection frame is 30×30=900. The union area is 900+900-625=1175, and the intersection-union ratio is 625 / 1175≈0.532.

[0043] Subsequently, the overlap threshold for the NMS operation was set to 0.5. This threshold was determined based on the common requirements of densely stacked aircraft materials scenarios, avoiding duplicate counting due to overly lenient overlap judgment and missed detection due to overly strict judgment. The calculated crossover ratio (CUR) was compared with the overlap threshold. If the CUR > 0.5, it means that the two detection boxes correspond to the same aircraft material target, and the detection box with the higher confidence score is retained, while the detection box with the lower confidence score is removed. If the CUR ≤ 0.5, it means that the two detection boxes correspond to different aircraft material targets, and both are retained. The CUR of the two M8 bolt detection boxes is ≈ 0.532 > 0.5, and the detection box with a confidence score of 0.63 has a higher score, so this detection box is retained, and the detection box with a confidence score of 0.61 is removed. Through this NMS operation, the detection boxes in the initial screening results are deduplicated, and the deduplicated detection boxes and their corresponding aircraft material target information are integrated to obtain the deduplicated detection results. This process solves the problem of duplicate counting caused by overlapping detection boxes in existing technologies, and is especially suitable for storage scenarios with multiple layers of stacked aviation materials, ensuring that each actual aviation material target corresponds to only one detection box.

[0044] Finally, the deduplicated inspection results are mapped and matched with the aircraft material codes in the warehouse management system. The system pre-establishes a "model inspection category - aircraft material code" mapping table, stored in the warehouse management system database, which can be manually modified according to the addition or specification change of aircraft materials (e.g., mapping the model inspection category "M8 bolt" to the aircraft material code "HC-BLT-002", and mapping "aviation high-pressure hose" to "HC-HOS-015"). The deduplicated inspection results are traversed, the category information of each inspection box is extracted, the corresponding aircraft material code is found according to the mapping table, and the actual storage location of the aircraft material is determined by combining the bounding box coordinates of the inspection box with the location label at the time of image acquisition (e.g., "shelf B-2 layer-5 column"). The number of inspection boxes corresponding to each aircraft material code is counted, and a valid inspection result containing aircraft material category, quantity, and location information is generated. This mapping and matching step solves the problem of data disconnect between inspection results and warehouse management system data, achieving accurate connection between "inspection results - inventory information", and providing structured data support for subsequent aircraft material inventory updates and inventory checks.

[0045] Throughout the data processing, the target detection results of aviation materials correspond to the initial screening detection results, the initial screening detection results correspond to the deduplicated detection results, and the deduplicated detection results correspond to the valid detection results. Data is passed sequentially and optimized step by step in each stage. Each technical operation is designed to address the differentiated needs and pain points of the aviation material warehouse detection scenario, ensuring that the final output of valid detection results meets the accuracy and practicality requirements of inventory management.

[0046] Step S106: Based on the effective detection results, obtain the statistical results of the number of aircraft materials by using preset mapping rules and location association rules, combined with the stacking layer calculation method.

[0047] In one specific embodiment, step S106 may specifically include the following steps: Based on preset mapping rules, the aircraft material categories in the valid detection results are matched with the aircraft material codes in the system database to obtain standardized aircraft material identification information; Based on the standardized aircraft material identification information, combined with the location tags added during image acquisition, the detected aircraft material targets are associated and bound with the actual storage locations in the warehouse, generating a correspondence between location information and aircraft material identification. Based on the aforementioned correspondence, the number of stacked layers of aviation materials is calculated by analyzing the ratio of the bounding box height of the aviation material target at the same location to the preset standard size of the aviation material. Based on the number of stacked aviation materials and the corresponding relationship, the actual quantity of each type of aviation material in each storage location is counted, and triplet data containing location information, aviation material identification and quantity statistics is generated. The triplet data is the aviation material quantity statistics result.

[0048] Specifically, after obtaining standardized aircraft material identification information, the location tags added during image acquisition are used to associate and bind the aircraft material targets with their actual storage locations. Location tags are generated during image preprocessing and contain four levels of location information: "shelf zone - shelf number - layer - column" (e.g., "East Zone - Shelf 05 - Layer 3 - Column 2"). These tags are associated with each detection box in the valid detection results through image frame number and pixel coordinate range (i.e., each detection box can be traced back to the location tag in the corresponding acquired image). The detection results carrying aircraft material identification information are traversed, and their associated location tags are extracted. The "aircraft material code - location tag - detection box" are then bound together to generate a correspondence between location information and aircraft material identification (e.g., "HC-BLT-001-East Zone - Shelf 05 - Layer 3 - Column 2 - Detection Box F1"). This association process solves the problem in existing technologies where detection results cannot pinpoint the actual storage location, preventing a disconnect between aircraft material quantity statistics and physical inventory locations, and providing data support for subsequent regional inventory checks and inventory location.

[0049] Based on the above correspondence, the number of stacked layers of aviation materials is calculated by the ratio of the bounding box height to the preset standard dimensions of the aviation materials. The preset standard dimensions of aviation materials are stored in the system database, categorized by aviation material code (e.g., the standard height corresponding to "HC-BLT-001" is 20mm, and the standard height corresponding to "HC-GSK-011" is 5mm). Corrections based on batch differences are supported (e.g., if the actual height of a batch of "HC-BLT-001" is 21mm, the standard dimensions of that code can be manually updated). Detection boxes for the same aviation material code under the same label in the correspondence are extracted, and the bounding box height (in pixels) of the detection box is read. Simultaneously, the bounding box height is converted to the actual physical height (in mm) using parameters from the image acquisition device (e.g., drone camera focal length, shooting distance). The conversion formula is as follows: (in This refers to the actual physical height. This represents the height of the bounding box in pixels. Camera focal length (unit: mm). Shooting distance (unit: mm) Where S is the horizontal pixel resolution of the image and S is the horizontal dimension of the camera sensor (unit: mm). For example, if the bounding box height of a "HC-BLT-001" detection box is 85 pixels, the camera focal length is 12mm, the shooting distance is 800mm, the horizontal pixel resolution of the image is 1280, and the horizontal dimension of the sensor is 1 / 2.3 inch (approximately 10.67mm), then the actual physical height is... If the actual measured physical height of the stacked aircraft materials corresponding to this detection frame is 60mm, and the preset standard height of "HC-BLT-001" is 20mm, then the number of stacked layers is... (round is the rounding function,) (for standard height), that is This calculation process solves the problem of quantity statistics deviation caused by the inability to identify the stacking of aviation materials in existing technologies. It is especially suitable for multi-layer stacking scenarios on shelves, ensuring the accuracy of quantity statistics for densely stacked aviation materials.

[0050] Based on the stacking layers of aviation materials and the above correspondence, the actual quantity of each type of aviation material in each storage location is counted. All correspondences are traversed, and groups are formed by "location label - aviation material code". If multiple detection frames exist within the same group (e.g., the same stacked aviation material photographed from different angles at the same location), the average stacking layer number is taken as the final layer number; if only a single detection frame exists, its calculated stacking layer number is used directly. For example, in the location "East Zone - Shelf 05 - 3rd Layer - 2nd Column", "HC-BLT-001" has two detection frames, with calculated layer numbers of 3 and 4 respectively. The average value of 3.5 is taken and rounded to 4, meaning the actual quantity of this aviation material at this location is 4 pieces. The quantity of all locations and all aviation material codes is counted in this way, generating a triplet data containing "location information - aviation material identifier - quantity" (e.g., "East Zone - Shelf 05 - 3rd Layer - 2nd Column - HC-BLT-001 - 4"). This triplet data is the aviation material quantity count result. This statistical process solves the problems of low efficiency and easy omissions in manual counting in existing technologies, realizes the automated and accurate counting of aircraft materials, and provides a data foundation for real-time inventory updates.

[0051] Throughout the data processing, valid detection results correspond to standardized aircraft material identification information. The standardized aircraft material identification information is combined with the location tag to correspond to the aircraft material identification. This correspondence is combined with the bounding box height and standard size to correspond to the number of aircraft material stacking layers. The number of stacking layers and the correspondence finally correspond to the aircraft material quantity statistics. The data of each link is closely linked, which is adapted to the statistical needs of complex storage scenarios in aircraft material warehouses.

[0052] Step S107: Compare the statistical results of the aircraft material quantity with the inventory data of the warehouse management system, update the warehouse inventory information according to the difference classification processing rules, and generate a report.

[0053] In one specific embodiment, step S107 may specifically include the following steps: The quantity statistics of the aircraft materials are compared with the inventory data in the warehouse management system to obtain the quantity difference items; The quantity difference items are classified based on the preset difference classification rules. Items with quantity differences not exceeding a first threshold are marked as normal fluctuations, and items with quantity differences exceeding the first threshold generate a difference warning notice. The inventory information of the warehouse management system is updated based on the classification processing results to obtain the updated inventory data; Based on the updated inventory data and comparison process data, an inventory comparison report containing details of discrepancies and processing results is generated.

[0054] Specifically, the aircraft material quantity statistics are first compared with the inventory data in the warehouse management system. The inventory data in the warehouse management system has a core structure of "aircraft material code - location tag - book quantity" (e.g., "HC-BLT-001-East Zone-Shelf 05-3rd Floor-2 Columns-5 Items"). The aircraft material quantity statistics are triplet data containing "location information - aircraft material identification - actual quantity" (e.g., "HC-BLT-001-East Zone-Shelf 05-3rd Floor-2 Columns-4 Items"). The system uses "aircraft material code + location tag" as the unique matching key, traverses the two types of data and performs item-by-item comparison, calculates the quantity difference value and difference rate for each matching item, and the formula for calculating the difference value is as follows: (where P is the quantity difference value, The actual quantity in the statistical results. (The quantity recorded in the system), the formula for calculating the difference rate is: The system integrates "part code - location tag - difference value - difference rate" to form a quantity difference item (e.g., "HC-BLT-001-East Zone-Shelf 05-3rd Floor-2 Columns-C=1 Item-K=20%)". If a "part code - location tag" only exists in the statistical results or system inventory data (e.g., system omission or statistical omission), it is marked as a "one-sided difference item" and included in the quantity difference item. This comparison process solves the problems of time-consuming manual inventory verification and easy errors due to visual fatigue in existing technologies, achieving automated and accurate matching of detection data and accounting data, avoiding omissions or calculation errors during manual verification.

[0055] The quantity difference items are classified and processed based on the preset difference classification rules. The first threshold in the preset difference classification rules is set to 5% (this threshold is based on the inventory fluctuation data of the aviation materials warehouse in the past year and is suitable for the normal difference range caused by handling loss of small-sized aviation materials such as bolts and gaskets). The system iterates through the quantity discrepancies. If the discrepancy rate K ≤ 5%, it is marked as normal fluctuation. For example, the discrepancy rate of “HC-GSK-011-West Zone-Shelf No. 02-1st Floor-4th Column” is 2.8%, which is judged as normal fluctuation. The system automatically records the fluctuation details to the temporary log for confirmation during weekly manual spot checks. If the discrepancy rate K > 5%, a discrepancy warning notice is generated. The warning notice includes the material code, location tag, system book quantity, statistical actual quantity, discrepancy value, discrepancy rate, and the storage path of the corresponding detection image (e.g., “HC-HYD-023-North Zone-Shelf No. 08-2nd Floor-1st Column-System Quantity 8 Pieces-Statistical Quantity 5 Pieces-P=3 Pieces-K=37.5%-Image Path / 20251113 / Img0825”). After the warning notice is generated, it is pushed to the terminal device (such as PDA or computer client) of the warehouse management personnel in real time. This tiered processing method addresses the problem of misallocation of management resources caused by the lack of prioritization of inventory discrepancies in existing technologies. It allows managers to prioritize high-discrepancy items, reduces ineffective manual checks, and improves the efficiency of handling inventory anomalies.

[0056] Based on the classification and processing results, the inventory information in the warehouse management system is updated. For items marked as normal fluctuations, the system does not automatically modify the book quantity, but adds statistical quantity and difference information to the inventory change log. Manual spot checks are required to confirm that there are no abnormalities, and the system approval process determines whether to retain the original data or make adjustments. For items that generate difference warnings, the system locks the editing permissions of the corresponding inventory data. Managers conduct on-site verification based on the warning (such as on-site counting of aviation materials and checking whether there are misjudgments in the detection images). If the verification confirms that the difference is caused by actual inventory changes (such as aviation materials not being registered), the manager corrects the book quantity to the statistical actual quantity through the system terminal. If the verification finds that the difference is caused by detection errors (such as deviation in stacking layer calculation), the warning is rejected and the stacking calculation parameters of the detection model are updated. At the same time, the local detection at that location is retried. For "one-sided difference items", if it is a system omission, the statistical quantity is supplemented and entered. If it is a statistical omission, the image is re-collected and the detection is performed to finally obtain the updated inventory data. This update process addresses the problems of existing technologies where inventory updates rely on manual input and are prone to data corruption due to operational errors. By employing a "tiered processing + manual review" mechanism, it balances update efficiency with data accuracy, ensuring that inventory information is consistent with the actual situation.

[0057] An inventory comparison report is generated based on updated inventory data and comparison process data. The report includes three parts: a detailed list of discrepancies, a summary of processing results, and data traceability. The detailed list of discrepancies includes the part code, location tag, original system quantity, actual quantity, discrepancy value, discrepancy rate, and processing status (normal fluctuation / under warning / corrected / rejected) for all quantity discrepancies. The summary of processing results shows the number of items in different processing statuses (e.g., "15 items of normal fluctuation - 6 items of warning - 4 items of corrected - 2 items of rejected") and calculates the overall discrepancy rate (the ratio of the sum of the discrepancy values ​​of all discrepancies to the total inventory quantity in the system). The data traceability section links the original data generated during the comparison process (e.g., detection image links, discrepancy calculation logs). The report is automatically generated as a PDF file using a fixed template, synchronously stored in the warehouse management system database and backed up in the cloud, and sent to the designated management personnel's email address. It supports retrieval by "date-part code-processing status". This report generation process solves the problem of insufficient written records of inventory discrepancies and lack of evidence for subsequent traceability in existing technologies, providing complete data support for inventory management auditing and detection model optimization.

[0058] Throughout the data processing process, the quantity difference items between the aviation material quantity statistics results and the system inventory data correspond to the quantity difference items, the classification processing results correspond to the updated inventory data, and the updated inventory data and the comparison process data correspond to the inventory comparison report. The data of each link are transmitted sequentially and are interconnected, which is adapted to the precise and traceable inventory management needs of aviation material warehouses.

[0059] Please see Figure 4 The following describes an adaptive downsampling aircraft material image detection system according to an embodiment of this application. The adaptive downsampling aircraft material image detection system 400 includes: The image acquisition module 401 is used to acquire initial images of aviation materials through image acquisition equipment deployed in the aviation material warehouse, and to preprocess the initial images of aviation materials to generate standardized images of aviation materials; Feature extraction module 402 is used to process the standardized aircraft material image through the improved Backbone module and output a multi-layer feature map; The feature fusion module 403 is used to input the multi-layer feature map into the improved Encoder module to obtain the semantic and detail-coordinated aircraft material fusion features; The target detection module 404 is used to input the fused features of the aircraft materials into the Transformer decoder, output the prediction information directly related to the aircraft material target, and accurately match the prediction information with the real annotation to obtain the aircraft material target detection result including the bounding box coordinates, category and confidence level; The result optimization module 405 is used to perform deduplication processing on the detection results of the aircraft material targets through preset confidence threshold rules and NMS operation to obtain effective detection results; The quantity statistics module 406 is used to obtain the quantity statistics results of aviation materials based on the effective detection results, through preset mapping rules and location association rules, and in combination with the stacking layer calculation method; The inventory update module 407 is used to compare the statistical results of the aircraft material quantity with the inventory data of the warehouse management system, update the warehouse inventory information and generate a report in accordance with the difference classification processing rules.

[0060] Through the collaborative efforts of the aforementioned components, the system constructs a complete closed loop for intelligent aircraft material inspection and management, from image acquisition to inventory updates. Specifically: Image acquisition module 401 acquires images of aircraft materials according to a preset flight path using a stereo acquisition network composed of a UAV and a fixed camera. It performs preprocessing such as image synchronization, format unification, illumination correction, and position labeling, providing standardized 640×640 pixel images of the aircraft materials for the feature extraction module. Feature extraction module 402 utilizes an improved Backbone, first performing adaptive downsampling through the LAWDS module to preserve small target details, and then performing multi-path feature extraction through the RGCSPELAN module, outputting multi-layer feature maps with strides of 8, 16, and 32, providing feature inputs containing rich detailed information for the feature fusion module. Feature fusion module 403 uses an improved Encoder, first constructing a basic feature pyramid using an FPN+PAN structure, and then performing channel alignment, contextual attention modeling, and bidirectional guided fusion through the CGFM module, providing semantically and detail-oriented fused features of the aircraft materials for the target detection module. Target detection module 404, based on the fused features, uses a Transformer decoder to generate initial predictions through a query mechanism, and then performs cross-attention optimization and Hungarian matching, outputting detection results containing bounding box coordinates, category, and confidence, providing detections to be filtered for the result optimization module. The data and results optimization module 405 first performs initial screening based on differentiated confidence rules, then removes overlapping detection boxes through NMS operations, and finally maps and matches them with the aircraft material codes in the warehousing system, providing accurate and effective structured detection results for the quantity statistics module. The quantity statistics module 406, based on the effective detection results, generates triplet data containing location information, aircraft material identification, and quantity statistics through aircraft material code mapping, location association binding, and stacking layer calculation, providing accurate quantity statistics results for the inventory update module. The inventory update module 407 compares the quantity statistics results with the inventory, processes normal fluctuations and abnormal differences according to the difference grading rules, and finally completes the inventory data update and generates a detailed comparison report. Through standardized data interfaces and streamlined business processes, these modules form a complete collaborative system for aircraft material inspection and inventory management, achieving automation of inventory management while ensuring inspection accuracy, significantly improving the inventory counting efficiency and inventory accuracy of the aircraft material warehouse.

[0061] This application also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the adaptive downsampling aircraft material image detection method.

[0062] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the methods and systems described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0063] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. An adaptive downsampling method for detecting aircraft material images, characterized in that, The process includes the following steps: Step S101: Acquire initial images of aircraft materials using image acquisition equipment deployed in the aircraft material warehouse; preprocess the initial images to generate standardized aircraft material images; Step S102: Input the standardized aircraft material images into an improved Backbone module. The improved Backbone module has built-in LAWDS and RGCSPELAN modules. The LAWDS module retains small target and edge feature information of the aircraft materials, while the RGCSPELAN module extracts multi-path features of the aircraft materials. After integration, a multi-layer feature map containing features of different scales of the aircraft materials is obtained; Step S103: Input the multi-layer feature map into an improved Encoder module. The improved Encoder module has built-in FPN+PAN structure and CGFM module. The FPN+PAN structure performs hierarchical transfer and aggregation of the basic features of the multi-layer feature map. At the same time, the CGFM module performs channel alignment, contextual attention modeling, and bidirectional guided fusion on the two-way aircraft material features during the hierarchical transfer process to obtain semantically and detail-oriented fused features of the aircraft materials. Step S104: Input the fused features of the aircraft materials into the Transformer decoder, output the predicted information directly related to the aircraft material target, and then accurately match the predicted information with the real annotation to obtain the aircraft material target detection result. The aircraft material target detection result includes bounding box coordinates, category, and confidence level. Step S105: Filter the aircraft material target detection result through preset confidence level rules, and perform deduplication processing on the filtered result through NMS operation to obtain the effective detection result. Step S106: Based on the effective detection result, obtain the aircraft material quantity statistics result through preset mapping rules and location association rules, combined with the stacking layer calculation method. Step S107: Compare the aircraft material quantity statistics result with the inventory data of the warehouse management system, update the warehouse inventory information and generate a report by combining the difference classification processing rules.

2. The method according to claim 1, characterized in that, Step S101 includes: acquiring initial images of the aircraft materials using a drone or a fixed camera; performing image frame synchronization on the initial images to obtain synchronized images; converting the synchronized images into JPG format and adjusting their size to a preset pixel value to obtain a unified image; performing brightness adjustment and noise reduction on the unified image to obtain an enhanced image; adding location tags to the enhanced image and linking it to the location database in the warehouse management system to generate the standardized aircraft material image.

3. The method according to claim 1, characterized in that, Step S102 includes: inputting the standardized aircraft material image into the improved Backbone module; performing adaptive downsampling processing on the standardized aircraft material image through the LAWDS module to retain small target and edge feature information of the aircraft material and obtain a high-fidelity feature map; extracting multi-path features from the high-fidelity feature map through the RGCSPELAN module to obtain multi-path output features; and integrating the multi-path output features based on a multi-scale semantic fusion mechanism to obtain a multi-layer feature map containing features of different scales of the aircraft material.

4. The method according to claim 1, characterized in that, Step S103 includes: performing top-down semantic enhancement on the multi-layer feature map using an FPN structure to obtain a semantically enhanced feature map, wherein the semantic enhancement is based on a hierarchical transfer mechanism to transfer high-level semantic information to low-level features; performing bottom-up path aggregation on the semantically enhanced feature map using a PAN structure to obtain preliminary fused features containing multi-scale information, wherein the path aggregation is based on a feature fusion mechanism to feed back low-level detail information to high-level features; performing channel alignment on the two aviation material features in the hierarchical transfer process using the CGFM module, and mapping the two aviation material features to the same metric space to obtain channel-aligned feature pairs; further performing context attention modeling on the channel-aligned feature pairs using the CGFM module to obtain context-recalibrated feature representations; and finally performing bidirectional guided fusion on the context-recalibrated feature representations using the CGFM module to generate semantically and detail-coordinated aviation material fused features.

5. The method according to claim 1, characterized in that, Step S104 includes: inputting the fused features of the aircraft materials into the Transformer decoder, generating initial prediction information containing bounding box coordinates and category distribution through a query mechanism; optimizing the initial prediction information using a cross-attention mechanism to enhance the feature discrimination between aircraft material targets and obtain optimized prediction information; using the Hungarian matching algorithm to perform one-to-one matching between the optimized prediction information and the real annotations, establishing a correspondence between the prediction results and the annotation data, and obtaining the matched detection results; performing confidence filtering on the matched detection results, selecting valid detection targets according to a preset confidence threshold, and generating the final aircraft material target detection results containing bounding box coordinates, categories, and confidence levels.

6. The method according to claim 1, characterized in that, Step S105 includes: screening the aircraft material target detection results according to a preset confidence rule, filtering out target detection boxes with confidence scores lower than a confidence threshold, and obtaining preliminary screening detection results; traversing the detection boxes corresponding to each aircraft material target in the preliminary screening detection results, and calculating the intersection-union ratio (CIU) between any two detection boxes; setting an overlap threshold for the NMS operation, and determining whether the CIU is greater than the overlap threshold. If so, retaining the detection boxes with higher confidence scores and removing the detection boxes with lower confidence scores; otherwise, retaining both detection boxes; integrating the detection boxes after the NMS operation and the corresponding aircraft material target information to obtain deduplicated detection results; mapping and matching the deduplicated detection results with the aircraft material codes in the warehouse management system to generate valid detection results containing aircraft material category, quantity, and location information.

7. The method according to claim 1, characterized in that, Step S106 includes: matching the aircraft material categories in the valid detection results with the aircraft material codes in the system database based on preset mapping rules to obtain standardized aircraft material identification information; based on the standardized aircraft material identification information, combined with the location tags added during image acquisition, associating and binding the detected aircraft material targets with the actual storage locations in the warehouse to generate a correspondence relationship containing location information and aircraft material identification; based on the correspondence relationship, calculating the number of aircraft material stacking layers by analyzing the ratio of the bounding box height of the aircraft material targets at the same location to the preset standard size of aircraft materials; based on the number of aircraft material stacking layers and the correspondence relationship, counting the actual quantity of each type of aircraft material at each storage location, generating triplet data containing location information, aircraft material identification, and quantity statistics, wherein the triplet data is the aircraft material quantity statistics result.

8. The method according to claim 1, characterized in that, Step S107 includes: comparing the statistical results of the aircraft material quantity with the inventory data in the warehouse management system to obtain quantity difference items; classifying the quantity difference items according to a preset difference classification rule, wherein items with quantity differences not exceeding a first threshold are marked as normal fluctuations, and items with quantity differences exceeding the first threshold generate a difference warning notice; updating the inventory information of the warehouse management system according to the classification results to obtain updated inventory data; and generating an inventory comparison report containing details of the difference items and processing results based on the updated inventory data and the comparison process data.

9. An adaptive downsampling aircraft material image detection system, used to implement the adaptive downsampling aircraft material image detection method as described in any one of claims 1 to 8, characterized in that, The adaptive downsampling aircraft material image detection system includes: an image acquisition module for acquiring initial images of aircraft materials through image acquisition equipment deployed in an aircraft material warehouse, and preprocessing the initial images to generate standardized aircraft material images; a feature extraction module for processing the standardized aircraft material images through an improved Backbone module to output multi-layer feature maps; a feature fusion module for inputting the multi-layer feature maps into an improved Encoder module to obtain semantically and detail-coordinated aircraft material fusion features; a target detection module for inputting the aircraft material fusion features into a Transformer decoder to output predicted information directly associated with aircraft material targets, and accurately matching the predicted information with the ground truth annotations to obtain aircraft material target detection results including bounding box coordinates, category, and confidence level; a result optimization module for deduplicating the aircraft material target detection results through preset confidence level threshold rules and NMS operations to obtain effective detection results; and a quantity statistics module for obtaining aircraft material quantity statistics results based on the effective detection results, through preset mapping rules and position association rules, combined with a stacking layer calculation method. The inventory update module is used to compare the statistical results of the aircraft material quantity with the inventory data of the warehouse management system, update the warehouse inventory information and generate a report based on the difference classification processing rules.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements an adaptive downsampling method for detecting aircraft material images as described in any one of claims 1 to 8.

Images