An adaptive down-sampling method and system for spare parts image detection, and a storage medium

Abstract

The application relates to the field of computer vision and discloses a kind of self-adapting down-sampling aviation material image detection method, system and storage medium.The method comprises the following steps: collecting aviation material images and pre-processing;extracting features through an improved Backbone module, using a LAWDS module for self-adapting down-sampling, and using an RGCSPELAN module to extract multi-path features;inputting the multi-layer feature map into an improved Encoder module, performing feature transmission and aggregation through an FPN+PAN structure, and realizing channel alignment and bidirectional guided fusion based on a CGFM module;inputting the fused features into a Transformer decoder to output detection results;obtaining effective detection results through confidence screening and NMS processing;counting the number of aviation materials based on the detection results;and comparing with the warehouse system and updating the inventory.The application solves the problems of small target feature loss and poor distinction in dense target area, and improves the detection accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision, in particular to a kind of adaptive down-sampling aviation material image detection method, system and storage medium. BACKGROUND

[0002] In aviation material warehouse management, in order to realize accurate inventory and real-time state perception, unmanned aerial vehicle or fixed camera is usually used to collect aviation material images, and deep learning target detection model is used for automatic identification and statistics. However, the aviation material warehouse environment has the characteristics of high shelves, dense material stacking, small target size and uneven illumination, which poses a serious challenge to the accuracy and robustness of the detection model.

[0003] Currently, deep learning-based target detection models (such as YOLO series and DEIM-HGNetV2-N) have been tried to be applied to the automatic detection of aviation material warehouses. Such models usually use fixed stride convolution and pooling operations for feature down-sampling, implement multi-scale feature fusion through a feature pyramid network, and finally output the recognition results through a detection head. However, in this specific application scenario of aviation material warehouses, the existing technology has obvious shortcomings. First, the traditional down-sampling method lacks an adaptive mechanism when compressing feature maps, and treats all regions of the image equally, resulting in the loss of small-sized aviation material details and edge information in the shallow network. Second, the existing feature fusion structure mostly uses simple concatenation or addition operations, without fully considering the semantic gap and contextual association between different scale features, making it difficult to achieve effective complementarity between semantic features and detailed features when dealing with complex situations such as dense stacking and mutual occlusion of aviation materials. In addition, existing models often struggle to balance between lightweight design and detection accuracy, either having high computational complexity that cannot meet real-time requirements, or being overly lightweight at the expense of small target detection capability. These defects collectively result in the existing detection methods facing high small target miss detection rate, poor dense target discrimination, and insufficient robustness in complex environments in actual application of aviation material warehouses.

[0004] To solve the above problems, the present application introduces a lightweight LAWDS module into the backbone network, uses a RGCSPELAN module with multi-path aggregation in the feature extraction stage, and adds a CGFM module in the feature fusion stage, solving the technical problems of easy loss of small target features and poor discrimination of dense targets in aviation material images during collection, and improving the detection recall rate of small aviation targets and recognition accuracy in complex stacking scenarios. SUMMARY

[0005] The application provides a kind of adaptive downsampling aviation material image detection method, system and storage medium, solve the technical problems that small target feature is easy to lose in aviation material image, and the distinguishing degree of dense target is poor, improve the detection recall rate of small target of aviation material and the identification precision under complex stacking scene.

[0006] In the first aspect, the application provides an adaptive downsampling aviation material image detection method, which comprises:

[0007] Step S101, obtain the initial image of aviation material by the image acquisition device deployed in the aviation material warehouse, preprocess the initial image of aviation material, and generate the standardized aviation material image;

[0008] Step S102, input the standardized aviation material image into the improved Backbone module, the improved Backbone module is built-in LAWDS module and RGCSPELAN module, the small target and edge feature information of aviation material are retained through the LAWDS module, and the multi-path features of aviation material are extracted based on the RGCSPELAN module, and after integration, the multi-layer feature map containing different scale features of aviation material is obtained;

[0009] Step S103, input the multi-layer feature map into the improved Encoder module, the improved Encoder module is built-in FPN+PAN structure and CGFM module, the basic features of the multi-layer feature map are hierarchically transmitted and aggregated through the FPN+PAN structure, and based on the CGFM module, the two-way aviation features in the hierarchical transmission process are aligned in channel, the context attention modeling and bidirectional guided fusion are modeled, and the aviation fusion features with coordinated semantics and details are obtained;

[0010] Step S104, input the aviation fusion features into the Transformer decoder, output the prediction information directly associated with the aviation target, and accurately correspond the prediction information with the real label to obtain the aviation target detection result, the aviation target detection result includes the boundary box coordinates, the category and the confidence;

[0011] Step S105, screen the aviation target detection result through the preset confidence rule, and remove the duplicate processing of the screening result through NMS operation to obtain the effective detection result;

[0012] Step S106, based on the effective detection result, through the preset mapping rule and position association rule, and combining with the stacking layer number calculation method, the aviation quantity statistical result is obtained;

[0013] Step S107, compare the aviation quantity statistical result with the warehouse management system inventory data, update the warehouse inventory information and generate the report combining with the difference classification processing rule.

[0014] In a second aspect, the application provides an adaptive down-sampling aviation material image detection system for implementing the adaptive down-sampling aviation material image detection method, and the system comprises:

[0015] An image acquisition module is configured to acquire aviation material initial images through image acquisition devices deployed in an aviation material warehouse, and to generate standardized aviation material images by preprocessing the aviation material initial images.

[0016] A feature extraction module is configured to process the standardized aviation material images through an improved Backbone module, and to output multi-layer feature maps.

[0017] A feature fusion module is configured to input the multi-layer feature maps into an improved Encoder module, and to obtain aviation material fusion features with coordinated semantics and details.

[0018] A target detection module is configured to input the aviation material fusion features into a Transformer decoder, to output prediction information directly associated with aviation material targets, and to accurately correspond the prediction information with true labels, thereby obtaining aviation material target detection results including boundary box coordinates, categories, and confidence levels.

[0019] A result optimization module is configured to perform deduplication processing on the aviation material target detection results through a preset confidence threshold rule and an NMS operation, thereby obtaining effective detection results.

[0020] A quantity statistics module is configured to obtain aviation material quantity statistics results based on the effective detection results, through a preset mapping rule and a position association rule, and in combination with a stacking layer calculation method.

[0021] An inventory updating module is configured to compare the aviation material quantity statistics results with inventory data of a warehouse management system, to update warehouse inventory information in combination with a difference classification processing rule, and to generate a report.

[0022] In a third aspect, the application provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the adaptive down-sampling aviation material image detection method.

[0023] The application provides an adaptive down-sampling aviation material image detection method, system, and storage medium, solves the technical problems of easy loss of small target features and poor distinguishability of dense target areas in aviation material images, and improves the detection recall rate of small aviation material targets and the recognition accuracy in complex stacking scenarios. Compared with the prior art, the application has at least the following beneficial effects:

[0024] First, by introducing a lightweight adaptive weight downsampling module, dynamic focusing of the key areas of the image is achieved, effectively preserving the small size of the details and edge information of the aviation material, and alleviating the feature loss problem caused by the traditional fixed downsampling method.

[0025] Second, the RGCSPELAN module based on multi-path aggregation is adopted, which enhances the model's expression ability of multi-scale features through channel splitting and multi-path convolution structure, while ensuring lightweight, it improves the richness and robustness of feature extraction.

[0026] Third, through the context-guided fusion module, adaptive interaction and complementarity of cross-scale features are realized, which enhances the model's ability to distinguish dense stacked aviation materials and improves the detection recall rate and precision in complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can be obtained based on these drawings without creative labor.

[0028] Figure 1 The flowchart of the aviation material image detection method of the adaptive downsampling in the present application;

[0029] Figure 2 The flowchart of the adaptive downsampling processing of the LAWDS module in the present application;

[0030] Figure 3 The flowchart of the RGCSPELAN module extracting aviation multi-path features in the present application;

[0031] Figure 4 The structure diagram of the aviation material image detection system of the adaptive downsampling in the present application. DETAILED DESCRIPTION

[0032] The embodiment of the present application provides a kind of adaptive downsampling aviation material image detection method, system and storage medium.The terms "first", "second", "third", "fourth" and the like (if exist) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence.It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein.In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0033] For ease of understanding, the specific process of the embodiment of the present application is described below, please refer to Figure 1 One embodiment of the adaptive downsampling aviation material image detection method in the embodiment of the present application includes:

[0034] Step S101, obtain an aviation material initial image by deploying an image acquisition device in an aviation material warehouse, pre-process the aviation material initial image, and generate a standardized aviation material image.

[0035] In a specific embodiment, the step S101 can specifically include the following steps:

[0036] The aviation material initial image is collected by a unmanned aerial vehicle or a fixed camera, and the aviation material initial image is subjected to image frame synchronization to obtain an aviation material synchronization image;

[0037] The aviation material synchronization image is uniformly converted into JPG format, and the size is adjusted to a preset pixel to obtain an aviation material uniform image;

[0038] The aviation material uniform image is subjected to brightness adjustment and denoising processing to obtain an aviation material enhanced image;

[0039] The aviation material enhanced image is added with a position label, and then associated to a position database in a warehouse management system to generate the standardized aviation material image.

[0040] Specifically, the image acquisition device deployed in the aviation material warehouse includes a multi-rotor unmanned aerial vehicle and a high-definition network camera. The unmanned aerial vehicle is equipped with a three-axis mechanical stabilization gimbal and a GPS+IMU dual-mode positioning module. The battery endurance meets the single-time full-coverage inspection needs of small and medium-sized aviation material warehouses. The fixed camera is deployed at key positions such as the warehouse upright column and the shelf beam, and the two form a three-dimensional acquisition network. The image acquisition device acquires initial images of aviation materials according to a preset "Z" shaped route or fixed point position. After the unmanned aerial vehicle completes the shooting of each shelf unit, it hovers to realize image frame synchronization, eliminating the problem of missing or re-shooting, and obtaining synchronized images of aviation materials. In view of the problem of non-uniform format and size difference of the synchronized images of aviation materials, the JPG format images shot by the unmanned aerial vehicle and the PNG format images shot by the fixed camera are all converted to JPG format, and the size is adjusted to 640*640 pixels, forming unified images of aviation materials, and ensuring that the subsequent model input data specifications are consistent. Considering the problem of uneven illumination and noise interference in the aviation material warehouse, the Retinex algorithm is used to adjust the brightness of the unified images of aviation materials, improve the brightness of the shelf shadow area, suppress the brightness of the strong light area, and then remove the image noise through Gaussian filtering to generate enhanced images of aviation materials, solving the problem of illumination change affecting detection accuracy in the prior art. Based on the positioning information during image acquisition, location labels are added to the enhanced images of aviation materials. The label information is associated with the location database of the warehouse management system, and the corresponding "shelf partition-shelf number-layer-column" storage location of the aviation material is clear. Finally, the standardized aviation material images are generated. In the whole data processing process, the initial images of aviation materials correspond to the synchronized images of aviation materials after frame synchronization, the synchronized images of aviation materials correspond to the unified images of aviation materials after format and size unification, the unified images of aviation materials correspond to the enhanced images of aviation materials after brightness adjustment and noise removal, and the enhanced images of aviation materials correspond to the standardized aviation material images after adding location labels and database association. The data of each link corresponds one by one and the key information is preserved during the transmission process. This process solves the problem that the manual counting or barcode recognition in the prior art is greatly affected by environmental factors, and the detection error caused by non-uniform data specifications. Through standardization processing, the aviation material images meet the subsequent model detection needs, provide high-quality data input for the feature extraction link, and lay a foundation for the subsequent detection results to be connected with the actual inventory through the association of the location label and the warehouse database.

[0041] Step S102, input the standardized aviation material images into the improved Backbone module, the improved Backbone module is built-in with LAWDS module and RGCSPELAN module, through the LAWDS module, the small target and edge feature information of the aviation material are retained, and based on the RGCSPELAN module, the multi-path features of the aviation material are extracted, and after integration, a multi-layer feature map containing different scale features of the aviation material is obtained.

[0042] In a specific embodiment, the step S102 can specifically include the following steps:

[0043] inputting the standardized aviation material image into an improved Backbone module;

[0044] performing adaptive down-sampling processing on the standardized aviation material image through the LAWDS module to retain aviation material small target and edge feature information, and obtaining a fidelity feature map;

[0045] performing multi-path feature extraction on the fidelity feature map through the RGCSPELAN module to obtain multi-path output features;

[0046] integrating the multi-path output features based on a multi-scale semantic fusion mechanism to obtain a multi-layer feature map containing aviation material features of different scales.

[0047] Specifically, referring to Figure 2 , the standardized aviation material image is input into an improved Backbone module, which is internally provided with a LAWDS module and a RGCSPELAN module, and the two modules cooperatively complete feature processing in sequence to output a multi-layer feature map. The standardized aviation material image first enters the LAWDS module, which performs adaptive down-sampling processing in three steps of attention weight generation, feature space reorganization and down-sampling, and weighted fusion output. In the attention weight generation step, first, average pooling is performed on the feature map corresponding to the standardized aviation material image, then spatial weight matrix is obtained by extracting local context information through 1x1 convolution, and then the feature map is divided into 2x2 four sub-regions through rearrange operation, and adaptive weight is allocated to each sub-region combined with Softmax operation. In the weight allocation process, the feature weight corresponding to the small target and edge region of the aviation material is preferentially strengthened, solving the technical problem of loss of small target and edge information caused by the existing fixed down-sampling method. In the feature space reorganization and down-sampling step, the LAWDS module uses 3x3 grouped convolution with a stride of 2 to compress the feature map in space, and at the same time expands the output channel number by 4 times to match the four sub-regions, and then converts the output to a tensor structure of [batch, channel, height, width, 4] through rearrange operation, so that each sub-region corresponds to the features of the left upper, right upper, left lower and right lower four sampling positions in the aviation material image. The design of grouped convolution reduces the computational overhead while achieving down-sampling, adapts to the lightweight needs of unmanned aerial vehicles and edge devices, and solves the problem of lack of adaptive mechanism and computational redundancy in feature down-sampling in the prior art. In the weighted fusion output step, the generated attention weight is multiplied by the features of the four sub-regions, and the sum of the last dimension is calculated to realize weighted fusion. If there is a difference between the input and output channel numbers, linear projection alignment is performed through 1x1 convolution, and finally the fidelity feature map retaining the small target and edge feature information of the aviation material is obtained. This process "re-stitches" the spatial sampling results through adaptive weight, avoids the information loss caused by traditional MaxPool which only takes the maximum value, and improves the feature fidelity.

[0048] Referring to Figure 3 , the 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, end feature refinement, and feature splicing and fusion. In the dynamic channel splitting stage, the fidelity feature map is first split into two parts by 1x1 convolution, one part is kept in the original structure for direct connection, and the other part enters the subsequent feature extraction process. Channel splitting reduces the number of channels in the subsequent 3x3 convolution processing, reduces computational redundancy, and solves the problem of limited multi-scale information capture and redundant computation in existing HGBlock. In the main path feature extraction stage, the split features enter the RepConv module, which enhances feature expression in a multi-branch structure during training to ensure that it can capture semantic information of different scales of aviation materials. In the inference stage, it is folded into a single 3x3 convolution to further reduce computational overhead, balancing model accuracy and lightweight requirements. In the additional path expansion stage, the main path output successively passes through three 3x3 convolution modules in series, layer by layer aggregating multi-scale semantic features of aviation materials. This design draws on the E-LAN idea but is more lightweight, can specifically strengthen feature discrimination in dense stacking scenarios of aviation materials, and alleviate the problem of declining recall rate in dense areas. In the end feature refinement stage, the end output of all paths is compressed and remapped by 1x1 convolution to enhance the non-linear expression ability of the features, making them more suitable for subsequent fusion needs. In the feature splicing and fusion stage, the original direct connection path features and the multi-path convolution extracted features are spliced, and then integrated by 1x1 convolution to output, obtaining multi-path output features. This process realizes efficient aggregation of multi-scale features of aviation materials and avoids the information bias caused by single path feature extraction.

[0049] Based on the multi-scale semantic fusion mechanism, the multi-path output features are integrated and divided into stride8, stride16, and stride32 levels according to the feature scale difference. The stride8 level feature retains the detail information of small target aviation materials, the stride32 level feature contains the global semantic information of aviation materials, and the stride16 level feature is between the two. Through hierarchical division, different scale features are classified and stored, and finally multi-layer feature maps containing different scale features of aviation materials are obtained. In the entire processing process, the standardized aviation image corresponds to the fidelity feature map, the fidelity feature map corresponds to the multi-path output feature, and the multi-path output feature corresponds to the multi-layer feature map. During data transmission, the aviation material features are gradually strengthened and refined, solving the problems of information loss in existing Backbone fixed downsampling and weak multi-scale information capture ability, and meeting the real-time operation requirements of warehouse edge devices through lightweight design, providing high-quality multi-scale feature input for the subsequent feature fusion of the Encoder module.

[0050] Step S103, input the multi-layer feature map into an improved Encoder module, which is built-in with a FPN+PAN structure and a CGFM module, pass and aggregate the base features of the multi-layer feature map through the FPN+PAN structure, meanwhile, align the channels of the two-way aviation material features in the hierarchical transmission process, model the context attention, and conduct bidirectional guided fusion based on the CGFM module, to obtain aviation material fusion features with coordinated semantics and details.

[0051] In a specific embodiment, the step S103 can specifically include the following steps:

[0052] perform semantic enhancement on the multi-layer feature map from top to bottom through the FPN structure to obtain a semantic-enhanced feature map, wherein the semantic enhancement is based on a hierarchical transmission mechanism to transmit high-level semantic information to low-level features;

[0053] perform path aggregation on the semantic-enhanced feature map from bottom to top through the PAN structure to obtain preliminary fusion features containing multi-scale information, wherein the path aggregation is based on a feature fusion mechanism to feed back low-level detailed information to high-level features;

[0054] align the channels of the two-way aviation material features in the hierarchical transmission process through the CGFM module, and map the two-way aviation material features to the same metric space to obtain a pair of channel-aligned features;

[0055] further model the context attention of the pair of channel-aligned features through the CGFM module to obtain context-realigned feature representations;

[0056] finally, conduct bidirectional guided fusion on the context-realigned feature representations through the CGFM module to generate aviation material fusion features with coordinated semantics and details.

[0057] Specifically, the multi-layer feature map is input into an improved Encoder module, which is built-in with an FPN+PAN structure and a CGFM module, both of which sequentially process data according to the logic of “basic feature transmission aggregation-feature optimization fusion”. The multi-layer feature map contains three levels of features of stride8, stride16 and stride32, among which the stride8 feature layer carries details such as edges and textures of small target materials, the stride32 feature layer contains global semantic information such as material categories and overall distribution, and the stride16 feature layer is a transitional feature between details and semantics, which first enters the FPN structure to perform semantic enhancement from top to bottom. The FPN structure takes the stride32 feature layer as the starting point, first adjusts its channel number to 256 (adapted to the subsequent fusion dimension) through 1x1 convolution, then performs 2 times upsampling operation to make the feature map resolution consistent with the stride16 feature layer, and then splices the channel dimension with the stride16 feature layer. After splicing, 3x3 convolution is performed to eliminate the grid artifacts introduced by upsampling, and the stride16 feature layer with enhanced semantics is obtained; in the same way, the stride16 feature layer with enhanced semantics is spliced and convolved with the stride8 feature layer after 1x1 convolution and 2 times upsampling, and the semantic enhanced feature map is obtained. This process solves the problem of lack of semantic information of low-level features in the prior art, allowing the stride8 feature layer to retain small target details while having category semantics, avoiding misjudgment of small targets due to lack of semantics.

[0058] The semantic enhanced feature map enters the PAN structure to perform path aggregation from bottom to top. The PAN structure takes the stride8 feature layer as the starting point, performs 2 times downsampling after 3x3 convolution to make the feature map resolution match the stride16 feature layer, and then splices the semantic enhanced stride16 feature layer output by the FPN structure, and then integrates the details and semantic information through 3x3 convolution to obtain the stride16 feature layer containing double perspective information; continue to perform 2 times downsampling on the stride16 feature layer, splice and convolve with the stride32 feature layer output by the FPN structure to obtain the preliminary fusion feature containing multi-scale information. This process supplements the details of high-level features, solves the problem of blurred small target outline description of high-level features in the prior art, and alleviates the boundary confusion of dense stacked materials caused by lack of details.

[0059] After the preliminary fusion features enter the CGFM module, first process the two-way aviation material features in the hierarchical transmission (such as the stride8 detail features transmitted by the PAN structure and the stride16 semantic features transmitted by the FPN structure), detect the channel numbers of the two-way features first, if the channel numbers of the two-way features are 128 and 256 respectively (the channel numbers are inconsistent), then map the 256-channel features to 128 channels through 1x1 convolution, so that the two-way features are in the same channel dimension (128 channels) of the measurement space, and the channel-aligned feature pair is obtained. This step solves the information loss problem caused by the inconsistent channel dimensions during feature fusion, and ensures that the two-way features can interact effectively.

[0060] The channel-aligned feature pair continues to perform context attention modeling, the CGFM module first performs global average pooling on the feature map (channel number 256) after splicing the two-way features, converts the HxWx256 feature map to a 1x1x256 channel descriptor, which reflects the global information of each channel; then input the channel descriptor into two fully connected networks (the first layer compresses 256 channels to 64 channels, and the second layer restores to 256 channels), and obtains 256 channel weight coefficients after Sigmoid function activation. The weight coefficient value ranges from 0 to 1, and the higher the value, the greater the contribution of the corresponding channel to the aviation material detection. The channel-weighted spliced feature map is obtained by using the weight coefficient, and the context recalibrated feature representation is obtained. This process solves the problem of background noise interference in the prior art during feature fusion, strengthens the expression of aviation material feature channels, and suppresses irrelevant background channels.

[0061] The context recalibrated feature representation is split into two paths (each with 128 channels) according to the channel dimension, and is used as the guide weight of the two-way original features (channel-aligned feature pair) respectively. Let the two-way guide weights be and , the two-way original features be and , then calculate the enhanced features , , and perform bidirectional residual fusion of and to inject the context information of the two-way features into each other; if the channel number of the fused features does not match the requirement of the downstream Decoder module (such as 256 channels), adjust the channel number to 256 through 1x1 convolution, and generate aviation material fusion features with coordinated semantics and details. This process solves the poor coordination problem caused by the static feature fusion in the prior art, so that the detail features carry semantic constraints and the semantic features carry detail support, and improves the feature discrimination of the occluded aviation material.

[0062] Throughout the process, the multi-layer feature map corresponds to the semantic enhanced feature map, the semantic enhanced feature map corresponds to the preliminary fused feature, the preliminary fused feature corresponds to the channel aligned feature pair, the channel aligned feature pair corresponds to the context recalibrated feature representation, and the context recalibrated feature representation corresponds to the aviation material fusion feature. Each link in the data transmission process is optimized for the technical defects of the previous link to ensure that the final output aviation material fusion feature meets the needs of dense and small target aviation material detection.

[0063] In step S104, the aviation material fusion feature is input into the Transformer decoder, and prediction information directly related to the aviation material target is output. The prediction information is accurately matched with the true label to obtain an aviation material target detection result, which includes a bounding box coordinate, a category, and a confidence.

[0064] In one specific embodiment, the step S104 can specifically include the following steps:

[0065] The aviation material fusion feature is input into the Transformer decoder to generate initial prediction information containing a bounding box coordinate and a category distribution through a query mechanism;

[0066] The initial prediction information is optimized using a cross-attention mechanism to enhance the feature discrimination between aviation material targets and obtain optimized prediction information;

[0067] The optimized prediction information is matched one-to-one with the true label using the Hungarian matching algorithm to establish a correspondence between the prediction result and the labeled data, and a matched detection result is obtained;

[0068] The matched detection result is filtered by confidence, and valid detection targets are selected according to a pre-set confidence threshold to generate a final aviation material target detection result containing a bounding box coordinate, a category, and a confidence.

[0069] Specifically, after the aviation material fusion feature is input into the Transformer decoder, the decoder first generates initial prediction information through a query mechanism. The query mechanism includes fixed learnable queries and dynamic queries. The learnable queries are set to 100 as pre-set learnable parameters to capture the global structural features of the aviation material. The dynamic queries are generated based on the potential aviation material target regions in the aviation material fusion feature, and the number is adaptively adjusted according to the target density in the fusion feature, which is used to accurately capture the local features of small size and densely stacked aviation materials. The two types of queries perform multiple rounds of query operations on the aviation material fusion feature, and then the initial prediction information is output through full connection layer mapping and Softmax activation function processing. This information contains the bounding box coordinate corresponding to each query (format is , 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.

[0070] 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:

[0071]

[0072] 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.

[0073] 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:

[0074] ,

[0075] Wherein, A is a prediction bounding box, B is a real bounding box, C is a minimum rectangular frame containing A and B, and IoU is the intersection over union of A and B. The class classification loss adopts a cross-entropy loss function (the formula is Wherein, M is the total number of classes, is a real class label, is a predicted class probability), and the weighting coefficients of the two types of losses are respectively set to 0.6 and 0.4. The total loss value The algorithm matches the optimized prediction information and the real label one by one according to the principle of minimum total loss value, eliminates the redundant prediction frame corresponding to no real target, and obtains the matched detection result. This step solves the problem of repeated detection or missed detection caused by chaotic detection frame allocation in the prior art, ensures that each aviation material target corresponds to only one prediction result, and improves the stability of the detection result.

[0076] The matched detection result is then subjected to confidence filtering. The system pre-sets the confidence threshold to be 0.7, extracts the maximum value of the probability in the class distribution of each detection result as the confidence of the result, and if the confidence is greater than or equal to 0.7, the bounding box coordinates, corresponding class (the class to which the maximum value belongs) and confidence of the result are retained, and the result is determined as an effective detection target. If the confidence is in the interval of 0.5-0.7, the result is marked as “to be manually reviewed” and is not included in the final result. If the confidence is less than 0.5, the detection result is directly filtered out. Finally, the aviation material target detection result including the bounding box coordinates, class and confidence of the effective detection target is generated. This process solves the problem of low-confidence prediction results interfering with inventory statistics in the prior art, and ensures that the output detection result has high reliability and can be directly used for subsequent aviation material quantity statistics and warehouse system inventory updating.

[0077] In the entire data processing link, the aviation material fusion features 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 result, and the matched detection result corresponds to the aviation material target detection result. The data is transmitted and optimized step by step to adapt to the technical requirements of the aviation material warehouse dense and small target detection scene.

[0078] In step S105, the aviation material target detection result is screened through a pre-set confidence rule, and the screening result is de-duplicated through an NMS operation to obtain an effective detection result.

[0079] In a specific embodiment, the step S105 can specifically include the following steps:

[0080] The aviation material target detection result is screened according to a pre-set confidence rule, and the target detection frame with a confidence lower than a confidence threshold is filtered out to obtain a preliminary screening detection result;

[0081] Traverse each detection frame corresponding to each aviation material target in the preliminary screening detection result, and calculate the intersection-over-union between any two detection frames;

[0082] Set the overlap threshold of the NMS operation, judge whether the intersection-over-union is greater than the overlap threshold, if yes, keep the detection frame with the higher confidence score and eliminate the detection frame with the lower confidence score, if not, keep both detection frames;

[0083] Integrate the detection frame and the corresponding aviation material target information after the NMS operation to obtain the de-duplicated detection result;

[0084] Map and match the de-duplicated detection result with the aviation material code in the warehouse management system to generate an effective detection result containing aviation material category, quantity and location information.

[0085] Specifically, the aviation material target detection result is first screened through a preset confidence rule. The rule is not a single threshold, but a multi-layer screening logic constructed in combination with the detection characteristics of different types of aviation materials in the warehouse and the inventory management requirements. Different confidence intervals are set for different types of aviation materials in the rule: for small-sized and easily stacked aviation materials such as bolts and gaskets, the basic confidence threshold is set to 0.65 due to the higher detection difficulty, and the confidence threshold is additionally lowered to 0.6 if the height deviation of the detection frame bounding box from the preset aviation material size is ≤10%; for large-sized aviation materials such as engine blades and fuselage frames, the basic confidence threshold is set to 0.75 due to the obvious features and high detection stability, and no additional lowering condition is added; for aviation materials such as wires and hoses that are prone to detection deviation due to changes in shape, in addition to the basic threshold of 0.7, the intersection-over-union between the detection frame and the adjacent aviation material detection frame must be ≤0.3 to be kept. When traversing the aviation material target detection result, the aviation material category corresponding to each detection frame is first identified, and then the screening condition of the category in the rule is matched. If the detection frame meets both the basic threshold and the additional condition, the bounding box coordinates and category information of the detection frame are kept; if only the basic threshold is met but the additional condition is not met, the detection frame is marked as “to be manually reviewed”; if the basic threshold is not met, the detection frame is directly filtered. The preliminary screening detection result is obtained through this process. The confidence rule solves the problem that a single confidence threshold cannot adapt to different types of aviation material detection scenarios, avoids the situation that small-sized aviation materials are mis-screened due to a too high threshold and large-sized aviation materials introduce false positives due to a too low threshold, and ensures that the preliminary screening detection result achieves a balance between precision and recall rate.

[0086] After obtaining the preliminary screening detection result, traverse each detection frame corresponding to each aviation material target, calculate the intersection-over-union (IoU) between any two detection frames, and the intersection-over-union calculation is based on the bounding box coordinates of the detection frame, with the formula being wherein is the area of the intersection region of the two detection frames, which is calculated by D, E represent two detection boxes respectively, is the upper left corner coordinate of D, is the lower right corner coordinate of D, and the coordinate parameters of E are the same. is the area of the union region of the two detection boxes, calculated by wherein, and represent the areas of the two detection boxes, , For example, in the small size area of the aviation material shelf, the bounding box coordinates of the two M8 bolt detection boxes are (180, 420, 210, 450) and (185, 425, 215, 455) respectively, the preset standard height of M8 bolt is 30 pixels, the height of the two detection boxes is 30 pixels, the deviation meets the supplementary condition, and the confidence is 0.63 and 0.61 respectively, both of which reach the down threshold of 0.6 for this type of aviation material, so both of them enter the preliminary screening result. The intersection area of the two is (210-185) x (450-425) = 25 x 25 = 625, the area of the first detection box is 30 x 30 = 900, the area of the second detection box is 30 x 30 = 900, and the union area is 900 + 900 - 625 = 1175, and the intersection ratio is 625 / 1175≈0.532.

[0087] Then set the overlap threshold of the NMS operation to 0.5, which is determined according to the common needs of the aviation material dense stacking scene, to avoid repeated counting due to loose overlap judgment and missing detection due to strict overlap judgment. Compare the calculated intersection ratio with the overlap threshold, if the intersection ratio > 0.5, it means that the two detection boxes correspond to the same aviation material target, keep the detection box with higher confidence score and eliminate the detection box with lower confidence score; if the intersection ratio ≤ 0.5, it means that the two detection boxes correspond to different aviation material targets, both of which are kept. The intersection ratio of the above two M8 bolt detection boxes is ≈0.532>0.5, and the detection box with confidence 0.63 has higher score, so the detection box is kept and the detection box with confidence 0.61 is eliminated. Through this NMS operation, the detection boxes in the preliminary screening detection result are de-duplicated, the de-duplicated detection boxes and the corresponding aviation material target information are integrated, and the de-duplicated detection result is obtained. This process solves the problem of repeated counting caused by overlapping detection boxes in the prior art, especially adapts to the storage scene of multi-layer stacking of aviation materials, and ensures that each actual aviation material target corresponds to only one detection box.

[0088] Finally, the de-duplicated detection results are mapped and matched with the aviation material codes in the warehouse management system. The system pre-establishes a "model detection category-aviation material code" mapping table and stores it in the warehouse management system database. Manual modification is supported according to the addition of aviation materials or changes in specifications (e.g., mapping the model detection category "M8 bolt" to the aviation material code "HC-BLT-002" and "aviation high-pressure hose" to "HC-HOS-015"). The de-duplicated detection results are traversed, and the category information of each detection box is extracted. According to the mapping table, the corresponding aviation material code is found, and the boundary box coordinates of the detection box and the position label (e.g., "shelf B-2 layer-5 column") during image acquisition are combined to determine the actual storage location of the aviation material. The number of detection boxes corresponding to each aviation material code is counted, and effective detection results containing aviation material categories, quantities, and location information are generated. This mapping and matching link solves the problem of disconnection between detection results and warehouse management system data, realizes the precise connection of "detection results-inventory information", and provides structured data support for subsequent aviation material inventory updating and checking.

[0089] During the entire data processing process, the aviation material target detection result corresponds to the preliminary screening detection result, the preliminary screening detection result corresponds to the de-duplicated detection result, and the de-duplicated detection result corresponds to the effective detection result. The data is transmitted and gradually optimized in sequence, and each technical operation is designed for the differentiated needs and pain points of the aviation material warehouse detection scene to ensure that the final output of the effective detection result meets the accuracy and practicality requirements of inventory management.

[0090] Step S106, based on the effective detection result, through a preset mapping rule and a position association rule, and combining a stacking layer calculation method, an aviation material quantity statistical result is obtained.

[0091] In a specific embodiment, the step S106 can specifically include the following steps:

[0092] Based on the preset mapping rule, the aviation material categories in the effective detection result are matched with the aviation material codes in the system database to obtain standardized aviation material identification information;

[0093] Based on the standardized aviation material identification information, the detected aviation material targets are associated and bound with the actual storage location of the warehouse by combining the position label added during image acquisition, and a corresponding relationship containing position information and aviation material identification is generated;

[0094] Based on the corresponding relationship, the number of stacked layers of aviation materials is calculated by analyzing the ratio of the boundary box height of the aviation material target at the same position to the preset aviation material standard size;

[0095] Based on the number of stacked layers of the aviation material and the corresponding relationship, the actual number of each type of aviation material at each storage location is counted, and a triple data containing location information, aviation material identification and quantity statistics is generated, which is the aviation material quantity statistics result.

[0096] Specifically, after obtaining the standardized aviation material identification information, the association binding of the aviation material target and the actual storage location is completed in combination with the position label added during image acquisition. The position label is generated in the image preprocessing stage and contains four-level position information of "shelf partition-shelf number-layer-column" (such as "East Zone-05 Shelf-3 Layer-2 Column"), and is associated with each detection box in the valid detection result through image frame number and pixel coordinate range (i.e. each detection box can be traced back to the position label of the corresponding acquisition image). The detection result carrying the aviation material identification information is traversed, the associated position label is extracted, and the "aviation material code-position label-detection box" is bound to generate a corresponding relationship containing location information and aviation material identification (such as "HC-BLT-001-East Zone-05 Shelf-3 Layer-2 Column-Detection Box F1"). This association process solves the problem that the detection result cannot be positioned in the actual storage location in the prior art, avoids the disconnection between aviation material quantity statistics and physical inventory location, and provides data support for subsequent regional inventory and inventory positioning.

[0097] Based on the above corresponding relationship, the number of stacked layers of the aviation material is calculated by the ratio of the bounding box height to the preset aviation material standard size. The preset aviation material standard size is stored in the system database and is recorded according to the aviation material code classification (such as the standard height of "HC-BLT-001" is 20mm, and the standard height of "HC-GSK-011" is 5mm), and can be corrected according to the batch difference of the aviation material (such as the actual height of a batch of "HC-BLT-001" is 21mm, which can be manually updated to the standard size of this code). The detection boxes of the same aviation material code under the same position label in the corresponding relationship are extracted, the bounding box height (unit: pixel) of the detection box is read, and the bounding box height is converted into the actual physical height (unit: mm) through the parameters of the image acquisition device (such as the focal length of the drone camera, the shooting distance). The conversion formula is (wherein is the actual physical height, is the bounding box height pixel value, is the camera focal length (unit: mm), is the shooting distance (unit: mm), is the image horizontal pixel resolution, and S is the horizontal size of the camera sensor (unit: mm)). For example, the bounding box height of a certain "HC-BLT-001" detection box is 85 pixels, the camera focal length is 12mm, the shooting distance is 800mm, the image horizontal pixel resolution is 1280, and the sensor horizontal size is 1 / 2.3 inch (about 10.67mm), then the actual physical height is If the detection frame corresponds to the stacked state of the aviation material, the actual measured physical height is 60mm, and the preset standard height of "HC-BLT-001" is 20mm, then the number of stacked layers (round is a rounding function, is the standard height), that is This calculation process solves the problem of number deviation caused by the inability to identify aviation material stacking in the prior art, especially adapts to the multi-layer stacking scene of the shelf, and ensures the number statistical accuracy of densely stacked aviation materials.

[0098] Based on the number of stacked layers of aviation materials and the above correspondence, the actual number of each type of aviation material at each storage location is counted. All correspondence is traversed, grouped by "location label-aviation material code", and if there are multiple detection frames in the same group (such as the same stacked aviation material photographed at different angles), the average value of the number of stacked layers is taken as the final number of layers; if there is only a single detection frame, the number of stacked layers calculated by it is directly used as the standard. For example, under the "East-05 shelf-3 layers-2 columns" location, "HC-BLT-001" has 2 detection frames, the number of layers is calculated as 3 layers and 4 layers respectively, and the average value 3.5 is taken and rounded to 4 layers, that is, the actual number of this aviation material at this location is 4 pieces. Count the number of all locations and all aviation material codes in this way to generate a triple data containing "location information-aviation material identification-number" (such as "East-05 shelf-3 layers-2 columns-HC-BLT-001-4"), which is the aviation material number statistical result. This statistical process solves the problem of low efficiency and easy to miss in the prior art, realizes the automatic and accurate statistics of the number of aviation materials, and provides a data basis for real-time updating of inventory.

[0099] In the entire data processing process, the effective detection result corresponds to the standardized aviation material identification information, the standardized aviation material identification information combines the position label to correspond to the corresponding relationship between the position and the aviation material identification, the corresponding relationship combines the boundary box height and the standard size to correspond to the number of stacked layers of aviation materials, and the number of stacked layers and the corresponding relationship finally correspond to the number of aviation material statistics results. The data of each link is connected, which adapts to the statistical needs of the complex storage scene of the aviation material warehouse.

[0100] Step S107, compare the number of aviation materials 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.

[0101] In a specific embodiment, the step S107 can specifically include the following steps:

[0102] Compare the number of aviation materials with the inventory data in the warehouse management system to obtain the number difference item;

[0103] The quantity difference items are classified based on preset difference classification rules, wherein an item with a quantity difference less than a first threshold is marked as normal fluctuation, and an item with a quantity difference greater than the first threshold generates a difference early warning single;

[0104] The inventory information of the warehouse management system is updated according to the classification processing result, to obtain updated inventory data;

[0105] Based on the updated inventory data and the comparison process data, an inventory comparison report containing difference item details and processing results is generated.

[0106] Specifically, the aviation material quantity statistics result is compared with the inventory data in the warehouse management system. The inventory data in the warehouse management system has a core structure of “aviation material code-position label-accounting quantity” (such as “HC-BLT-001-East Zone-05 shelf-3 layers-2 columns-5 pieces”), and the aviation material quantity statistics result is a three-tuple data containing “position information-aviation material identification-actual quantity” (such as “HC-BLT-001-East Zone-05 shelf-3 layers-2 columns-4 pieces”). The system takes “aviation material code+position label” 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 of each matching item, the difference value calculation formula is (wherein P is the quantity difference value, is the actual quantity in the statistics result, is the accounting quantity in the system), and the difference rate calculation formula is “aviation material code-position label-difference value-difference rate” is integrated to form a quantity difference item (such as “HC-BLT-001-East Zone-05 shelf-3 layers-2 columns-C=1 piece-K=20%”). If a certain “aviation material code-position label” exists only in the statistics result or the system inventory data (such as system recording or statistics missing), it is marked as a “one-sided difference item” and included in the quantity difference item. This comparison process solves the problem of time-consuming manual inventory checking and easy errors due to visual fatigue in the prior art, realizes automatic and accurate matching of detected data and accounting data, and avoids missing items or calculation deviations during manual checking.

[0107] The quantity difference items are classified based on preset difference classification rules. The first threshold value in the preset difference classification rules is set to 5% (the threshold value is referenced to the inventory fluctuation data of the aviation material warehouse in the past year, and is suitable for the normal difference range of small-size aviation materials such as bolts and gaskets caused by transportation loss). If the difference rate K is less than or equal to 5%, the item is marked as normal fluctuation, for example, the difference rate of “HC-GSK-011- West Zone-02 Rack-1 Layer-4 Column” is 2.8%, which is determined as normal fluctuation, and the system automatically records the fluctuation details to the temporary log, which is confirmed by manual inspection every week. If the difference rate K is greater than 5%, a difference warning sheet is generated, which includes the aviation material code, location label, system account quantity, statistical actual quantity, difference value, difference rate and storage path of the corresponding detection image of the difference aviation material (such as “HC-HYD-023-North Zone-08 Rack-2 Layer-1 Column-System Quantity 8-Pieces-Statistical Quantity 5-Pieces-P=3-Pieces-K=37.5%-Image Path / 20251113 / Img0825”), and the warning sheet is pushed to the terminal device (such as PDA or computer client) of the warehouse management personnel in real time after being generated. The classification process solves the problem of mismatching of management resources caused by the lack of priority in the prior art, so that the management personnel can prioritize high difference rate items, reduce invalid manual verification, and improve the efficiency of inventory exception processing.

[0108] The inventory information of the warehouse management system is updated according to the classification processing result. For the items marked as normal fluctuation, the system does not automatically modify the account quantity, but adds the statistical quantity and difference information in the inventory change log. After manual inspection confirms that there is no exception, whether to keep the original data or adjust is determined through the system approval process. For the items that generate the difference warning sheet, the system locks the editing permission of the corresponding inventory data. The management personnel rechecks the site according to the warning sheet (such as counting the aviation materials on site and checking whether the detection image is misjudged). If it is confirmed that the difference is caused by actual inventory change (such as aviation material taking without registration), the management personnel corrects the account quantity to the statistical actual quantity through the system terminal. If it is found that the difference is caused by detection error (such as calculation deviation of stacking layer), the warning is rejected, the stacking calculation parameters of the detection model are updated, and the local detection of the position is retriggered. For the “single-side difference item”, if it is a system omission, the statistical quantity is supplemented, and if it is a statistical omission, the image is reacquired for detection. Finally, the updated inventory data is obtained. The updating process solves the problem that the inventory updating in the prior art relies on manual input and is prone to data confusion due to operation errors. Through the mechanism of “classification processing + manual rechecking”, the updating efficiency and data accuracy are balanced to ensure that the inventory information is consistent with the actual situation.

[0109] Based on the updated inventory data and the comparison process data, an inventory comparison report is generated, which includes three parts of contents: difference item details, processing result summary and data traceability. The difference item details list all quantity difference items of the aviation material code, location label, system original account quantity, statistical actual quantity, difference value, difference rate and processing state (normal fluctuation / early warning processing / has been corrected / has been rejected); the processing result summary statistics the number of items in different processing states (such as “normal fluctuation items 15-early warning items 6-corrected items 4-rejected items 2”), and calculates the overall difference rate (the sum of the difference values of all difference items / the total inventory quantity of the system); the data traceability part is associated with the original data generated in the comparison process (such as detection image link, difference calculation log). The report automatically generates a PDF file in a fixed template, synchronously stores it in the warehouse management system database and the cloud backup, and sends it to the designated management personnel mailbox, supporting retrieval and query by “date-aviation material code-processing state”. The report generation process solves the problem of lack of written records of inventory differences and no traceable data in the prior art, and provides complete data support for inventory management audit and detection model optimization.

[0110] During the entire data processing process, the aviation material quantity statistics result corresponds to the quantity difference item of the system inventory data, the quantity difference item corresponds to the classified processing result, the classified processing result corresponds to the updated inventory data, the updated inventory data corresponds to the inventory comparison report of the comparison process data, and the data of each link is transmitted and associated with each other, which meets the precise and traceable inventory management requirements of the aviation material warehouse.

[0111] Please refer to Figure 4 , the following describes a self-adaptive down-sampling aviation material image detection system in an embodiment of the present application. The self-adaptive down-sampling aviation material image detection system 400 includes:

[0112] An image acquisition module 401 is configured to acquire an initial aviation material image through an image acquisition device deployed in an aviation material warehouse, and to preprocess the initial aviation material image to generate a standardized aviation material image.

[0113] A feature extraction module 402 is configured to process the standardized aviation material image through an improved Backbone module to output a multi-layer feature map.

[0114] A feature fusion module 403 is configured to input the multi-layer feature map into an improved Encoder module to obtain aviation fusion features with coordinated semantics and details.

[0115] The target detection module 404 is configured to input the aviation material fusion feature into a Transformer decoder, output prediction information directly associated with an aviation material target, accurately correspond the prediction information with a true label, and obtain an aviation material target detection result including a bounding box coordinate, a category, and a confidence level.

[0116] The result optimization module 405 is configured to perform a deduplication processing on the aviation material target detection result by using a preset confidence threshold rule and an NMS operation, and obtain an effective detection result.

[0117] The quantity statistics module 406 is configured to obtain an aviation material quantity statistics result by using a preset mapping rule and a position association rule, and combining a stacking layer number calculation method, based on the effective detection result.

[0118] The inventory updating module 407 is configured to compare the aviation material quantity statistics result with warehouse management system inventory data, update warehouse inventory information in combination with a difference classification processing rule, and generate a report.

[0119] Through the cooperation of the above-mentioned components, the system builds a complete aviation material intelligent detection and management closed loop from image acquisition to inventory updating. Among them:

[0120] The image acquisition module 401 acquires the aerial material image according to the preset route through the stereoscopic acquisition network composed of the unmanned aerial vehicle and the fixed camera, and completes image synchronization, format unification, illumination correction and position labeling and the like pretreatment, to provide a standardized aerial material image of 640*640 pixels for the feature extraction module; the feature extraction module 402 uses the improved Backbone, first performs adaptive down-sampling to reserve small target details through the LAWDS module, and then performs multi-path feature extraction through the RGCSPELAN module, to output multi-layer feature maps with strides of 8, 16 and 32, and provide feature input containing rich detail information for the feature fusion module; the feature fusion module 403 uses the improved Encoder, first uses the FPN+PAN structure to complete the basic feature pyramid construction, and then performs channel alignment, context attention modeling and bidirectional guided fusion through the CGFM module, to provide the aerial material fusion features with semantic and detail cooperation for the target detection module; the target detection module 404 uses the Transformer decoder to generate the initial prediction through the query mechanism based on the fusion features, and then performs cross-attention optimization and Hungarian matching, to output the detection results containing the boundary box coordinates, the category and the confidence, and provide the detection data to be screened for the result optimization module; the result optimization module 405 first performs preliminary screening according to the differentiated confidence rules, then removes the overlapping detection boxes through the NMS operation, and finally maps and matches with the warehouse system aerial material code, to provide the accurate and effective structured detection results for the quantity statistics module; the quantity statistics module 406 generates the triple data containing the position information, the aerial material identification and the quantity statistics based on the effective detection results through the aerial material code mapping, the position correlation binding and the stack layer calculation, to provide the accurate quantity statistics results for the inventory updating module; the inventory updating module 407 compares the quantity statistics results with the inventory, respectively processes the normal fluctuation and the abnormal difference according to the difference grading rules, and finally completes the inventory data updating and generates the detailed comparison report. The modules form a complete aerial material detection and inventory management collaborative system through the standardized data interface and the process-based business connection, realize the automation of the inventory management while ensuring the detection accuracy, and significantly improve the inventory efficiency and the inventory accuracy of the aerial material warehouse.

[0121] The 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, and instructions are stored in the computer readable storage medium, and when the instructions are run on a computer, the computer executes the steps of the aerial material image detection method with adaptive down-sampling.

[0122] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described method and system can refer to the corresponding process in the foregoing method embodiment, which will not be described here.

[0123] The above examples are only used to illustrate the technical solutions of the present application, but not to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some technical features can be replaced by equivalent features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An adaptive down-sampling method for spare parts image detection, characterized in that, The method comprises the following steps: step S101, obtaining an initial image of the aviation material through an image acquisition device arranged in an aviation material warehouse, preprocessing the initial image of the aviation material, and generating a standardized aviation material image; step S102, inputting the standardized aviation material image into an improved Backbone module, the improved Backbone module being internally provided with a LAWDS module and an RGCSPELAN module, retaining small target and edge feature information of the aviation material through the LAWDS module, extracting multi-path features of the aviation material based on the RGCSPELAN module, and obtaining a multi-layer feature map containing different scale features of the aviation material after integration; step S103, inputting the multi-layer feature map into an improved Encoder module, the improved Encoder module being internally provided with an FPN+PAN structure and a CGFM module, performing hierarchical transmission and aggregation of basic features of the multi-layer feature map through the FPN+PAN structure, and performing channel alignment, context attention modeling and bidirectional guided fusion on two paths of aviation material features in the hierarchical transmission process based on the CGFM module, to obtain aviation material fusion features with coordinated semantics and details; step S104, inputting the aviation material fusion features into a Transformer decoder, outputting prediction information directly related to the aviation material target, accurately corresponding the prediction information with the real label, and obtaining an aviation material target detection result, the aviation material target detection result comprising a bounding box coordinate, a category and a confidence; step S105, screening the aviation material target detection result through a preset confidence rule, performing a de-duplication process on the screening result through an NMS operation, and obtaining an effective detection result; step S106, based on the effective detection result, obtaining an aviation material quantity statistical result through a preset mapping rule and a position association rule, and combining a stack layer calculation method; step S107, comparing the aviation material quantity statistical result with warehouse management system inventory data, updating warehouse inventory information in combination with a difference classification processing rule, and generating a report.

2. The method of claim 1, wherein, The step S101 comprises: collecting an initial image of the aviation material through a drone or a fixed camera, obtaining an aviation material synchronization image through image frame synchronization of the initial image of the aviation material; uniformly converting the aviation material synchronization image into a JPG format, adjusting the size to a preset pixel, and obtaining an aviation material unified image; performing brightness adjustment and denoising processing on the aviation material unified image, and obtaining an aviation material enhanced image; adding a position label to the aviation material enhanced image, and associating it to a position database in a warehouse management system, to generate the standardized aviation material image.

3. The method of claim 1, wherein, The step S102 comprises: inputting the standardized aviation material image into an improved Backbone module; performing adaptive down-sampling processing on the standardized aviation material image through the LAWDS module, retaining aviation material small target and edge feature information, and obtaining a fidelity feature map; performing multi-path feature extraction on the fidelity feature map through the RGCSPELAN module, and obtaining multi-path output features; and integrating the multi-path output features based on a multi-scale semantic fusion mechanism, and obtaining a multi-layer feature map containing aviation material features of different scales.

4. The method of claim 1, wherein, The step S103 comprises: performing top-down semantic enhancement on the multi-layer feature map through an FPN structure, and obtaining a semantic-enhanced feature map, wherein the semantic enhancement is based on a hierarchical transmission mechanism to transmit high-level semantic information to low-level features; performing bottom-up path aggregation on the semantic-enhanced feature map through a PAN structure, and obtaining preliminary fusion features containing multi-scale information, wherein the path aggregation is based on a feature fusion mechanism to feed back low-level detailed information to high-level features; performing channel alignment on two aviation material features in the hierarchical transmission process through the CGFM module, and simultaneously mapping the two aviation material features to the same metric space to obtain a pair of channel-aligned features; further modeling context attention of the pair of channel-aligned features through the CGFM module to obtain context-realigned feature representations; and finally performing bidirectional guided fusion on the context-realigned feature representations through the CGFM module to generate aviation material fusion features with coordinated semantics and details.

5. The method of claim 1, wherein, The step S104 comprises: inputting the aviation material fusion features into a Transformer decoder to generate initial prediction information containing bounding box coordinates and class distribution through a query mechanism; performing optimization processing on the initial prediction information by using a cross-attention mechanism to enhance feature discrimination between aviation material targets, and obtaining optimized prediction information; adopting a Hungarian matching algorithm to one-to-one match the optimized prediction information with real labels, establishing a corresponding relationship between prediction results and labeled data, and obtaining matched detection results; performing confidence filtering on the matched detection results, screening effective detection targets according to a pre-set confidence threshold, and generating final aviation material target detection results containing bounding box coordinates, classes and confidence. The step S104 comprises: inputting the aviation material fusion features into a Transformer decoder to generate initial prediction information containing bounding box coordinates and class distribution through a query mechanism; performing optimization processing on the initial prediction information by using a cross-attention mechanism to enhance feature discrimination between aviation material targets, and obtaining optimized prediction information; adopting a Hungarian matching algorithm to one-to-one match the optimized prediction information with real labels, establishing a corresponding relationship between prediction results and labeled data, and obtaining matched detection results; performing confidence filtering on the matched detection results, screening effective detection targets according to a pre-set confidence threshold, and generating final aviation material target detection results containing bounding box coordinates, classes and confidence.

6. The method of claim 1, wherein, The step S105 comprises: screening the spare parts target detection result according to a preset confidence rule, filtering out the target detection frame with a confidence lower than a confidence threshold, and obtaining a preliminary screening detection result; traversing the detection frame corresponding to each spare parts target in the preliminary screening detection result, calculating the intersection-over-union between any two detection frames; setting an overlap threshold of the NMS operation, judging whether the intersection-over-union is greater than the overlap threshold, if yes, retaining the detection frame with a higher confidence score and eliminating the detection frame with a lower confidence score, if not, both detection frames are retained; integrating the detection frame and the corresponding spare parts target information after the NMS operation, obtaining a deduplicated detection result; mapping and matching the deduplicated detection result with the spare parts code in the warehouse management system, and generating an effective detection result containing spare parts category, quantity and location information.

7. The method of claim 1, wherein, The step S106 comprises: matching the spare parts category in the effective detection result with the spare parts code in the system database based on a preset mapping rule, and obtaining standardized spare parts identification information; based on the standardized spare parts identification information, combining the position label added during image acquisition, associating and binding the detected spare parts target with the actual storage location of the warehouse, and generating a corresponding relationship containing location information and spare parts identification; based on the corresponding relationship, calculating the number of spare parts stacking layers by analyzing the ratio of the boundary box height of the spare parts target at the same position to the preset spare parts standard size; based on the number of spare parts stacking layers and the corresponding relationship, statistically obtaining the actual quantity of each type of spare parts at each storage location, and generating a triple data containing location information, spare parts identification and quantity statistics, wherein the triple data is the spare parts quantity statistical result.

8. The method of claim 1, wherein, The step S107 comprises: comparing the spare parts quantity statistical result with the inventory data in the warehouse management system, and obtaining a quantity difference item; classifying the quantity difference item based on a preset difference classification rule, wherein the item with a quantity difference not exceeding a first threshold is marked as normal fluctuation, and the item with a quantity difference exceeding the first threshold generates a difference early warning single; updating the inventory information of the warehouse management system according to the classification processing result, and obtaining updated inventory data; generating an inventory comparison report containing difference item details and processing results based on the updated inventory data and comparison process data.

9. An adaptive down-sampling aerial material image detection system for implementing the adaptive down-sampling aerial material image detection method according to any one of claims 1 to 8, characterized in that, The adaptive down-sampling aircraft material image detection system comprises: an image acquisition module configured to acquire an aircraft material initial image by an image acquisition device deployed in an aircraft material warehouse, and to generate a standardized aircraft material image by preprocessing the aircraft material initial image; a feature extraction module configured to process the standardized aircraft material image by an improved Backbone module, and to output a multi-layer feature map; a feature fusion module configured to input the multi-layer feature map into an improved Encoder module, and to obtain aircraft material fusion features with coordinated semantics and details; a target detection module configured to input the aircraft material fusion features into a Transformer decoder, to output prediction information directly associated with aircraft material targets, and to accurately correspond the prediction information with true labels, and to obtain aircraft material target detection results including boundary box coordinates, categories and confidence; a result optimization module configured to remove redundant aircraft material target detection results by a preset confidence threshold rule and an NMS operation, and to obtain effective detection results; a quantity statistics module configured to obtain aircraft material quantity statistics results based on the effective detection results, by a preset mapping rule and a position association rule, and in combination with a stacking layer number calculation method; and a stock updating module configured to compare the aircraft material quantity statistics results with warehouse management system stock data, to update warehouse stock information in combination with a difference classification processing rule, and to generate a report.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the adaptive down-sampling aircraft material image detection method according to any one of claims 1 to 8.

Images