Intelligent goods identification method and system for goods delivery unmanned aerial vehicle, medium and processor

Abstract

The invention discloses a cargo intelligent identification method and system for a material delivery unmanned aerial vehicle, a medium and a processor, and relates to the technical field of image cargo identification. Aiming at the problems of image blurring and cargo edge and texture detail loss caused by insufficient illumination and fog interference when an unmanned aerial vehicle collects images, an original image is processed through an image enhancement algorithm to obtain an enhanced image, cargo edge structure information is extracted and enhanced, and a texture integrity index is calculated after a target area is segmented. Repairing the texture in combination with a preset template library to generate an optimized image; the method comprises the steps of extracting a cargo surface feature set through adaptive filtering by fusing environmental interference parameters, matching a multi-scale judgment index, outputting a preliminary identification tag, verifying the consistency of the tag and actual features through a convolutional neural network, and finally outputting a cargo category and state report. According to the method, image enhancement, deep learning and adaptive filtering are combined, the accuracy and efficiency of cargo identification in a complex environment are improved, and reliable support is provided for cargo conveying management.

Description

Technical Field

[0001] This invention relates to the field of image cargo recognition technology, and in particular to a method, system, medium and processor for intelligent cargo recognition of unmanned aerial vehicles (UAVs) used for cargo delivery. Background Technology

[0002] In the process of transporting goods in harsh environments, freight transport has long faced the dual challenges of efficiency and cost. The freight areas corresponding to these harsh environments are characterized by towering mountains and deep ravines, with narrow paths winding through steep cliffs. Hardened roads that can be used by vehicles often only cover the township level, while village roads are mostly muddy dirt roads that become difficult to travel on during the rainy season and are closed for long periods in winter due to snow and ice. Under extreme weather conditions, transport channels are often in a state of semi-paralysis.

[0003] Against the backdrop of rapid development in modern logistics and drone technology, image-based drone cargo feature accurate identification has become a key area for improving transportation efficiency and safety. Currently, although some methods have attempted to improve cargo identification accuracy through image processing, these methods often struggle to adapt to complex real-world environments. Especially during drone filming, the diversity of environmental factors leads to unstable image quality, thus affecting identification results. Existing solutions often fail to fundamentally address the problems caused by image quality degradation when dealing with these dynamic changes, particularly in terms of detail capture and feature rendering, significantly compromising the reliability of cargo identification.

[0004] Focusing on the technical challenges, image quality fluctuations become the primary issue. When drones capture images under varying lighting and weather conditions, the images often become blurry or have low contrast, making it difficult to clearly present crucial information such as cargo edges and surface textures. This decline in image quality further exacerbates the difficulty of feature extraction, as cargo edges and textures are crucial for identifying cargo type and condition. If these details cannot be effectively preserved or recovered, subsequent analysis becomes like groping in the fog, making it difficult to reach accurate conclusions. Therefore, effectively improving the clarity of cargo edge and texture details in drone-captured images has become a key issue affecting the accurate identification of cargo features. For example, in rainy weather, when a drone captures images of stacked goods from high altitude, insufficient light and fog interference blur the outlines of the goods, making even surface markings unrecognizable. This directly prevents the system from determining the type of goods or whether they are damaged, impacting the efficiency and accuracy of cargo transportation management.

[0005] Therefore, technological exploration to improve image quality will provide a more reliable guarantee for drone cargo identification. In-depth research into environmental adaptability issues can lay a solid foundation for subsequent accurate identification.

[0006] Therefore, there is a need for a method, system, medium, and processor for intelligent cargo identification of drones used for material delivery. Summary of the Invention

[0007] To address the problems of image blurring and loss of cargo edge and texture details caused by insufficient lighting and fog interference when drones acquire images in operational scenarios, this invention provides a method, system, medium, and processor for intelligent cargo identification using drones for material delivery. This solution resolves the issues of image blurring, loss of cargo edge and texture details, and cargo feature extraction challenges in complex environments, significantly improving the accuracy and efficiency of cargo identification in cargo delivery management. The specific technical solution is as follows: A method for intelligent cargo identification for delivery drones includes the following steps: S1: The original image acquired by the drone is processed using an image enhancement algorithm to obtain the enhanced image; S2: Extract cargo edge structure information based on the enhanced image; S3: Segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture-optimized image; S4: Based on texture optimization, image fusion environmental interference parameters are used, and filtering algorithms are employed to suppress interference. Then, the output cargo surface feature descriptor and cargo surface feature set are extracted. S5: Extract cargo surface features from the cargo surface feature set and match them with preset multi-scale cargo judgment indicators to obtain cargo categories and output preliminary identification labels; S6: Obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy; S7: Output cargo category and status report from the final identification accuracy to obtain complete cargo identification results for subsequent cargo transportation management.

[0008] Furthermore, in step S1, the original image acquired by the drone is processed using an image enhancement algorithm to obtain an enhanced image, specifically including the following steps: S11: The raw images acquired by the drone are subjected to preliminary format unification and noise filtering to obtain the preliminarily processed images; S12: Detect the illumination of the pre-processed image, and perform global illumination compensation on the insufficient image using brightness adjustment technology to obtain the illumination-corrected image; S13: For the image after illumination correction, the histogram equalization method in the image enhancement algorithm is used to adjust the contrast in order to highlight the edge information of the goods in the image, and obtain the image with improved contrast. S14: Detect fog interference areas in the image after contrast enhancement. If fog interference areas are detected, apply defogging technology to clarify the corresponding areas and obtain the defogging image. S15: Extract cargo edge information from the dehazed image to enhance cargo edge details, resulting in a cargo edge-enhanced image; S16: Obtain texture feature information from the image, perform detail reconstruction processing on the enhanced image of the cargo edge to restore texture details, and obtain the final enhanced image.

[0009] Furthermore, in step S2, extracting the cargo edge structure information based on the enhanced image includes the following steps: S21: Obtain initial cargo edge information of the enhanced image through a pre-established cargo edge detection operator to obtain preliminary cargo edge distribution data; S22: Calculate the edge strength value of the cargo based on the preliminary cargo edge distribution data. If the calculated strength value is lower than the preset threshold, trigger the deep processing flow to determine the area that needs further processing. S23: For areas requiring further processing, use a convolutional neural network to extract deep features from the image data and obtain fine feature information; S24: The extracted fine feature information is reorganized to construct an enhanced cargo edge structure, resulting in an optimized cargo edge representation; S25: Apply smoothing technology to the optimized cargo edge representation to eliminate noise interference and obtain continuous cargo edge lines; S26: Structure and integrate continuous cargo edge lines to generate the final cargo edge structure information.

[0010] Furthermore, in step S3, the process of segmenting the target cargo region based on cargo edge structure information, extracting the texture of the target cargo region and calculating the texture integrity index, and then performing similarity matching and repair with a preset texture template to generate a texture-optimized image includes the following steps: S31: Based on the cargo edge structure information, a contour lookup and filling algorithm is used to determine the closed region where the cargo body is located, which serves as the cargo target region mask; based on the cargo target region mask, the corresponding cargo region image is cropped from the enhanced image, denoted as... The local binary mode or gray-level co-occurrence matrix algorithm is used to process the cargo area image. Processing is performed to generate the corresponding texture feature map. ; S32: Texture feature map Adaptive thresholding is performed to obtain a binarized texture saliency map. ; for binarized texture saliency maps Morphological closing operations are performed to fill tiny holes and connect neighboring texture pixels. Subsequently, connected component analysis is performed, and regions with a pixel value of 1 and which are interconnected are marked as a single "texture contiguous region". The texture contiguous region is defined as one with an area greater than a preset minimum area threshold. Connected components; calculate texture integrity index : ; In the above formula, The total number of pixels in all contiguous texture regions; Masking the target area of ​​the cargo The total number of pixels within.

[0011] S33: Set integrity threshold ;like If the texture is complete, the cargo area image is directly processed. As a texture-optimized image Output; if If the texture is incomplete, the incomplete area is recorded, triggering the next step of the texture matching and repair process based on the template library; S34: The texture matching and repair process based on the template library includes: establishing a preset texture template library; each template in the preset texture template library contains two parts of data: texture feature vector. and corresponding high-quality texture image blocks For each incomplete region identified, feature vectors are extracted from the surrounding complete texture regions. ; Calculate the eigenvectors Compared with all texture feature vectors in the preset texture template library The cosine similarity is used; the template k with the highest similarity is selected as the matching template; high-quality texture image patches of the matching template are used. Histogram matching and adaptive affine transformation are performed based on the local statistical characteristics of the incomplete region to generate a repair patch. Then, using Poisson image editing or a deep learning-based image inpainting model, the repair patch is seamlessly integrated into the incomplete region to obtain the repaired cargo area image. ; Image of the repaired cargo area The image is blended with the original image background to output the final texture-optimized image. .

[0012] Furthermore, in step S4, the step of optimizing the image fusion environmental interference parameters based on texture, using a filtering algorithm to suppress interference, and then extracting the output cargo surface feature descriptor and cargo surface feature set includes the following steps: S41: Define the set of environmental interference parameters This set includes the following parameters: light interference factor Fog interference factors Motion fuzziness factor Input texture optimization image, based on interference parameter set Adaptive filtering is performed on the texture-optimized image to obtain a preliminary filtered image. ; S42: In the initial filtered image Within the target area of ​​the goods, calculate the variance of the grayscale gradient magnitude for each local block; if the variance value of a certain local block is lower than the preset grayscale gradient variance threshold... If the area is identified as a residual blurred region, soft thresholding based on wavelet transform is applied to denoise these regions to obtain an optimized surface image. ; S43: Optimize surface image Keypoint detection and feature description are performed, and the coordinates and descriptor vectors of all keypoints are organized together into a set of cargo surface features, as shown in the following formula: ; in, The pixel coordinates of the i-th key point; Let be the feature descriptor vector of the i-th keypoint; n is the total number of keypoints. S44: Analyze the set of cargo surface features The recurrence rate and descriptor consistency of feature points across multiple consecutive frames; if the stability score of the current feature point is lower than the feature stability threshold. If the condition is not met, it will be removed from the set; finally, a clear set of cargo surface features will be obtained and output.

[0013] Furthermore, in step S5, the step of extracting cargo surface features from the cargo surface feature set and matching them with preset multi-scale cargo judgment indicators to obtain cargo categories and outputting preliminary identification labels specifically includes the following steps: S51: Extract a clear cargo surface image from the cargo surface feature set, and perform segmentation and enhancement processing on the cargo surface image using image processing technology to obtain clear surface information data; S52: Use multi-scale analysis to extract features from surface information data at different levels and determine the multi-scale feature set; S53: Construct a corresponding judgment index system based on the multi-scale feature set, perform index calculation and normalization on the multi-scale feature set, and obtain standardized index data; S54: Compare the standardized indicator data with the category parameters in the preset model, output the matching data as the initial category label, and record the non-matching data as pending and transfer it to subsequent processing; S55: Perform a secondary comparison between the surface parameters associated with the initial category label and the surface information data to determine the final identification result; S56: Generate preliminary identification labels based on the final identification results.

[0014] Furthermore, in step S6, the step of obtaining associated preset cargo status data based on the preliminary identification tag, verifying the consistency between the cargo status data and the cargo surface features, and determining the final identification accuracy includes the following steps: S61: Extract cargo status information related to the initial identification tag from the database, and sort it according to preset rules to obtain the initial sorted status dataset; S62: Obtain the actual feature data of the cargo surface features corresponding to each preliminary identification tag and associate and map it with the data in the preliminary sorting status dataset to form a feature mapping set; S63: Use a convolutional neural network to process the feature mapping set, analyze the degree of matching between the preliminary identification labels and the actual feature data, and determine the feature consistency result; S64: If the feature consistency result is lower than the preset threshold, the feature mapping set is extracted a second time to obtain supplementary actual feature data and determine whether the consistency has been improved. S65: By classifying and organizing the results after consistency improvement, the final identified label data is labeled to obtain the basis for accuracy evaluation; S66: Based on the accuracy assessment criteria, a logical comparison tool is used to verify the final identification results and determine the final matching conclusion between the surface features of the goods and the preliminary identification label; S67: By storing the final matching results, an index structure is built for subsequent query needs to obtain traceable identification records.

[0015] A cargo intelligent identification system for a material delivery drone, used to implement the above-described cargo intelligent identification method for a material delivery drone, comprising: The enhancement processing module is used to process the original images acquired by the UAV using image enhancement algorithms to obtain enhanced images; The edge enhancement module is used to extract the edge structure information of the cargo based on the enhanced image; The optimization processing module is used to segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture optimization image; based on the texture optimization image, environmental interference parameters are fused, and a filtering algorithm is used to suppress interference, and then the cargo surface feature descriptor and cargo surface feature set are extracted and output. The preliminary label recognition module is used to extract cargo surface features from the cargo surface feature set, match them with preset multi-scale cargo judgment indicators to obtain cargo categories, and output preliminary recognition labels. The verification module is used to obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy. The report output module is used to output cargo category and status reports from the final identification accuracy, obtaining complete cargo identification results for subsequent cargo transportation management.

[0016] A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the above-described intelligent cargo identification method for unmanned aerial vehicles (UAVs) used for material delivery.

[0017] A processor for running a program, wherein the program executes the above-described intelligent cargo identification method for unmanned aerial vehicles (UAVs) used for material delivery.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application discloses a cargo recognition method based on UAV image processing. Addressing the business scenario problem of image blurring and loss of cargo edge and texture details caused by insufficient lighting and fog interference when UAVs acquire images, it achieves accurate cargo recognition through a series of technical means. Specifically, firstly, the original image acquired by the UAV is processed using an image enhancement algorithm to remove interference caused by insufficient lighting and fog, resulting in an enhanced image used to restore edge and texture details of the cargo features. Based on the enhanced image, preliminary cargo edge information is extracted. If the cargo edge intensity is lower than a preset threshold, a convolutional neural network is applied for deep feature learning to determine the enhanced cargo edge structure and output an enhanced cargo edge binary image. Based on the enhanced cargo edge structure, the cargo target region is segmented, and its texture detail distribution is extracted and analyzed. By calculating a texture integrity index and performing similarity matching and repair with a preset texture template library, a texture-optimized image is generated. Based on texture-optimized images, this invention integrates environmental interference parameters obtained from image metadata and sensors, employs a parameter-adaptive filtering algorithm to suppress specific interference, and then extracts clear cargo surface feature descriptors and outputs a cargo surface feature set. From this set, clear cargo surface features are extracted, and multi-scale cargo judgment indicators are derived. If the indicators match a preset cargo category model, a preliminary identification label is output. Based on the preliminary identification label, associated cargo status data is obtained, and the consistency between the label and actual features is verified using a convolutional neural network to determine the final identification accuracy. From the final identification accuracy, a cargo category and status report are output, yielding a complete cargo identification result. Analysis of the above execution scheme shows that this invention first processes the original image using an image enhancement algorithm to restore cargo edges and texture details. Then, it utilizes deep learning with a convolutional neural network to strengthen the cargo edge structure and optimizes texture distribution by combining texture template similarity comparison. Simultaneously, for the remaining blurred areas, this invention integrates environmental interference parameters, employs adaptive filtering technology to extract clear cargo surface features, and ensures consistency between the identification label and actual features through multi-scale indicator matching and status data verification, ultimately outputting a cargo category and status report. The core innovation of this invention lies in combining image enhancement, deep learning, and adaptive filtering, which solves the problems of image quality in complex environments and the difficulty of cargo feature extraction, and significantly improves the accuracy and efficiency of cargo identification in cargo transportation management. Attached Figure Description

[0019] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0020] Figure 1 This is a flowchart illustrating a method for intelligent cargo identification in drones used for material delivery. Figure 2 This is a schematic diagram of a cargo intelligent identification system for use with drones for material delivery. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be understood that, when used in this application, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof.

[0023] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0024] It should also be further understood that the term “and / or” as used in this application refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0025] Example 1 like Figure 1 The diagram shows a flowchart of a smart cargo identification method for unmanned aerial vehicles (UAVs) used for cargo delivery, which includes the following steps: S1: The original image is acquired by the drone, and the image enhancement algorithm is used to process the interference caused by insufficient lighting and weather fog to obtain the enhanced image, which is used to restore the edge details and texture details of the cargo features.

[0026] In the context of "supply delivery," specific examples of edge and texture details of cargo features can be found in the following scenario: Assume a drone is capturing an image of a plastic storage box containing relief supplies. Edge details could include the outer contour edge or structural edge, and the boundary line between the box and the ground, grass, or ruins. This is the most prominent edge—the outer contour edge. The structural edge is the seam line between the lid and the box body.

[0027] Specific examples in the "material distribution" scenario: Taking plastic storage boxes as an example again, the texture details can be as follows: Texture details of a smooth plastic surface: A uniform, high-gloss texture, possibly with slight reflective patterns. Texture details of a frosted plastic surface: A fine, granular, irregular texture. Texture details of a canvas bag or burlap sack surface: A rough, interwoven fiber mesh texture. Texture details of a corrugated cardboard box surface: A regular, parallel striped wavy texture.

[0028] Furthermore, in step S1, the original image acquired by the drone is processed using an image enhancement algorithm to obtain an enhanced image, specifically including the following steps: S11: The raw images acquired by the drone are processed using a pre-defined image preprocessing procedure to perform preliminary format unification and noise filtering, resulting in a pre-processed image. S12: Detect the illumination of the pre-processed image. If insufficient illumination is detected, perform global illumination compensation on the image using brightness adjustment technology to obtain an illumination-corrected image. S13: For the image after illumination correction, the histogram equalization method in the image enhancement algorithm is used to adjust the contrast in order to highlight the edge information of the goods in the image, and obtain the image with improved contrast. S14: Detect fog interference areas in the image after contrast enhancement. If fog interference areas are detected, obtain the possible fog interference areas and apply defogging technology to clarify the corresponding areas to obtain the defogging image. S15: To address the need for cargo edge restoration, cargo edge detection technology is used to extract cargo edge information from the dehazed image to enhance cargo edge details, resulting in an enhanced cargo edge image. S16: Obtain texture feature information from the image, perform detail reconstruction processing on the enhanced image of the cargo edge to restore texture details, and obtain the final enhanced image.

[0029] For example, in scenarios where drones acquire raw image data, assuming a terrain survey is being conducted in a mountainous area, the acquired images may have various problems due to factors such as weather and lighting. For image preprocessing, the raw images can first be formatted uniformly, for example, adjusting images of different resolutions to 1920x1080 pixels. Simultaneously, Gaussian filtering can be applied to filter noise and reduce image blurring caused by drone vibration, resulting in pre-processed image data. This preprocessing effectively improves the accuracy of subsequent analysis.

[0030] For example, in situations with insufficient lighting, assuming a mountain image is generally dark due to overcast skies, brightness adjustment techniques can be used to increase the average brightness value of the image from 50 to around 80, performing global illumination compensation. This method makes image details easier to identify, especially in shadowed areas of the terrain, ensuring that subsequent processing does not lose crucial information due to lighting issues.

[0031] For example, by adjusting the contrast of an image after illumination correction using histogram equalization, the edge information of goods in mountainous images can be highlighted. Assuming the original image has low contrast and the boundaries between mountains and valleys are not clear, equalization processing results in a more uniform grayscale distribution, revealing the details of the goods' edges. For instance, if fog interference is detected in the image, such as blurring of valleys due to morning fog, dark channel prior dehazing techniques can be applied to sharpen these foggy areas. Assuming the pixel values ​​in foggy areas are high, dehazing restores the details of roads in the valleys. This technique significantly improves the usability of images in harsh environments. In the dehazed image, for the need to restore goods edges, Canny goods edge detection technology can be used to extract the goods' outline information. By adjusting the detection threshold, the details of the goods' edges are enhanced, such as by thickening the boundary lines between the goods and the background. This enhanced goods edge effect helps to more accurately identify goods features.

[0032] S2: Extract cargo edge structure information based on the enhanced image.

[0033] Furthermore, in step S2, extracting the cargo edge structure information based on the enhanced image includes the following steps: S21: Obtain initial cargo edge information of the enhanced image by using pre-established cargo edge detection operators (such as Canny, Sobel, etc.) to obtain preliminary cargo edge distribution data; S22: Calculate the edge strength value of the cargo based on the preliminary cargo edge distribution data. If the calculated strength value is lower than the preset threshold, trigger the deep processing flow to determine the area that needs further processing. S23: For areas requiring further processing, use a convolutional neural network to extract deep features from the image data to obtain more refined feature information; S24: The extracted fine feature information is reorganized to construct an enhanced cargo edge structure, resulting in an optimized cargo edge representation; S25: Apply smoothing technology to the optimized cargo edge representation to eliminate noise interference and obtain more continuous cargo edge lines; S26: Perform structured integration on continuous cargo edge lines to generate final cargo edge structure information data, complete the image cargo edge enhancement processing, and output the cargo edge structure (i.e., the enhanced cargo edge binary image).

[0034] For example, in the field of image processing acquired by drones, the extraction and optimization of cargo edge information for enhanced images can be analyzed and implemented in detail from multiple perspectives. Firstly, regarding the acquisition of initial cargo edge information, a pre-designed cargo edge detection operator can be used to scan the image and identify areas with significant grayscale changes as preliminary cargo edge distribution data. Assuming a cargo image taken by a drone has a resolution of 1920x1080 pixels, during processing, priority can be given to the grayscale abrupt change areas at the boundary between the cargo and the ground, marking them as the initial cargo edges.

[0035] For example, the calculation and thresholding of cargo edge intensity values ​​can be based on the magnitude of the grayscale gradient to assess the salience of cargo edges. If the intensity value of the current region is lower than a preset threshold, such as 50, it indicates that the cargo edge information in that region may not be clear enough and requires further processing. Suppose that in the cargo image above, some cargo has a small grayscale difference from the background, with an intensity value of only 30, then a deep processing flow will be triggered to focus on analyzing these blurred areas. In the deep feature extraction stage, convolutional neural networks can perform multi-level analysis of the image, uncovering more subtle feature information. For blurred areas at the boundary between cargo and background in the cargo image, the network can extract texture and shape features layer by layer, identifying the subtle features of small cargo that were originally difficult to distinguish. This approach helps improve the integrity of cargo edges, especially in scenes with insufficient lighting or fog interference. For example, the reorganization of fine feature information can integrate the extracted features to construct a clearer cargo edge structure. In the cargo image, the reorganized cargo edge structure can connect discontinuous cargo lines to form a complete outline. This processing method can significantly improve the visualization quality of the image, facilitating subsequent analysis.

[0036] For example, applying smoothing techniques can effectively reduce noise interference when representing optimized cargo edges. Suppose some cargo edges in an image appear jagged due to noise; smoothing can generate more continuous lines, improving the overall visual appeal. This method is particularly suitable for handling subtle noise issues caused by shaking or wind in drone images.

[0037] In the final structured integration phase, the continuous cargo edge lines are integrated into complete cargo edge structure data, which can provide a reliable foundation for subsequent applications.

[0038] From an overall process perspective, the aforementioned steps are interconnected, from initial detection to in-depth optimization, and finally to integration, forming a complete cargo edge enhancement system. This system effectively addresses the problem of blurred cargo edges in complex environments during UAV image processing, providing a higher-quality data foundation for image analysis, especially in scenarios with insufficient lighting or fog interference, where its effectiveness is particularly pronounced.

[0039] S3: Segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture-optimized image.

[0040] Furthermore, in step S3, the process of segmenting the target cargo region based on cargo edge structure information, extracting the texture of the target cargo region and calculating the texture integrity index, and then performing similarity matching and repair with a preset texture template to generate a texture-optimized image includes the following steps: S31: Based on the cargo edge structure information (i.e., the enhanced cargo edge binary image), the closed region where the cargo body is located is determined through contour finding and filling algorithms and used as the cargo target region mask; based on the cargo target region mask, the corresponding cargo region image is cropped from the enhanced image, denoted as... The local binary mode or gray-level co-occurrence matrix algorithm is used to process the cargo area image. Processing is performed to generate the corresponding texture feature map. It achieves cargo region segmentation and texture extraction.

[0041] S32: Texture feature map Adaptive thresholding is performed to obtain a binarized texture saliency map. (Where 1 represents a salient texture pixel); for binarized texture saliency maps Morphological closing operations (using 3x3 rectangular structuring elements) are performed to fill the tiny holes and connect neighboring texture pixels. Subsequently, connected component analysis is performed, and regions with a pixel value of 1 and which are interconnected are marked as a single "texture contiguous region". The texture contiguous region is defined as one with an area greater than a preset minimum area threshold. Connected components (e.g., This implements the steps of texture contiguous region analysis and integrity calculation; and calculates the texture integrity index. : ; In the above formula, The total number of pixels in all contiguous texture regions; Masking the target area of ​​the cargo The total number of pixels within.

[0042] S33: Set integrity threshold (For example, );like If the texture is complete, the cargo area image is directly processed. As a texture-optimized image Output; if If the texture is incomplete, the incomplete region (i.e., the binarized texture saliency map) is recorded. (Pixel areas with a median value of 0) trigger the next step of the texture matching and repair process based on the template library.

[0043] S34: The texture matching and repair process based on the template library includes: establishing a preset texture template library; each template in the preset texture template library contains two parts of data: texture feature vector. (Constructed from LBP histograms or GLCM statistics extracted from standard cargo sample images) and corresponding high-quality texture image patches. For each incomplete region identified, feature vectors are extracted from the surrounding complete texture regions. ; Calculate the eigenvectors Compared with all texture feature vectors in the preset texture template library The cosine similarity is used; the template k with the highest similarity is selected as the matching template; high-quality texture image patches of the matching template are used. Histogram matching and adaptive affine transformation are performed based on the local statistical characteristics of the incomplete region to generate a repair patch. Then, using Poisson image editing or a deep learning-based image inpainting model, the repair patch is seamlessly integrated into the incomplete region to obtain the repaired cargo area image. ; Image of the repaired cargo area (Or the original image of the entire region) is blended with the original image background to output the final texture-optimized image. .

[0044] S4: Based on texture optimization, image fusion environmental interference parameters are used, and filtering algorithms are employed to suppress interference. Then, the output cargo surface feature descriptor and cargo surface feature set are extracted.

[0045] Furthermore, in step S4, the step of optimizing the image fusion environmental interference parameters based on texture, using a filtering algorithm to suppress interference, and then extracting the output cargo surface feature descriptor and cargo surface feature set includes the following steps: S41: Define the set of environmental interference parameters This set includes the following parameters: light interference factor Fog interference factors Motion fuzziness factor Input texture optimization image, based on interference parameter set Adaptive filtering is performed on the texture-optimized image to obtain a preliminary filtered image. ; S42: In the initial filtered image Within the target area of ​​the goods, calculate the variance of the grayscale gradient magnitude for each local block (e.g., 8x8 pixels); if the variance value of a certain local block is lower than the preset grayscale gradient variance threshold... If the region is identified as a residual blurred region, it is then subjected to secondary processing using soft thresholding denoising based on wavelet transform or bilateral filtering to obtain an optimized surface image. ; S43: Optimize surface image Keypoint detection and feature description are performed, and the coordinates and descriptor vectors of all keypoints are organized together into a set of cargo surface features, as shown in the following formula: ; in, The pixel coordinates of the i-th key point; Let be the feature descriptor vector of the i-th keypoint; n is the total number of keypoints. S44: Analyze the set of cargo surface features The recurrence rate and descriptor consistency of feature points across multiple consecutive frames; if the stability score of the current feature point is lower than the feature stability threshold. If the condition is not met, it will be removed from the set; finally, a clear set of cargo surface features will be obtained and output.

[0046] The above content is detailed below for specific implementation; Step 1: Define the set of environmental interference parameters This set is a combination of the following parameters: Light interference factor ISO sensitivity and exposure time can be read from the image's EXIF ​​data, or estimated by calculating the skewness of the overall image brightness histogram. A high value indicates insufficient light or significant noise.

[0047] Fog interference factor The average value of the global atmospheric light value and the transmittance map estimated using the dark channel prior method. A high value indicates a high fog concentration.

[0048] Motion blur factor The angular velocity and linear velocity at the moment of shooting are extracted from the data of the UAV inertial measurement unit (IMU) and converted into the equivalent fuzzy kernel size and direction through the calibration model. A high value indicates severe blurring. (Note: These parameters can be acquired and calculated synchronously during the image acquisition stage in step S1, and are used as global variables for subsequent steps.)

[0049] Step 2: Perform adaptive filtering: Input image requires texture optimization According to the interference parameter set Dynamically select and configure the filtering algorithm: when Dominant (high ISO noise): Non-local mean filtering is used, and its filter strength h is... Proportional: Where h is the actual filtering intensity, The base filter strength (preset constant); This is a light interference factor. Light noise should be eliminated first.

[0050] when Dominant (Haze Effect): Based on existing dehazing, guided filtering is applied to locally remaining blurred areas to optimize the image for desired texture. The gradient map serves as a guide, enhancing the edges while smoothing out residual fog.

[0051] when Dominant (motion blur): Wiener filtering or Richardson-Lucy deconvolution is used. Image restoration is performed using a defined blur kernel. The above adaptive filtering is then applied to the entire image to obtain a preliminary filtered image. .

[0052] Step 3: Perform residual blur region detection and secondary optimization: In the initial filtered image Within the target area of ​​the goods, calculate the variance of the grayscale gradient magnitude for each local block (e.g., 8x8 pixels). If the variance value of the current local block is lower than a preset grayscale gradient variance threshold... If these areas are found to be residual blurred regions, they are then identified as such. These regions undergo a more aggressive secondary processing using wavelet transform-based soft thresholding or bilateral filtering, emphasizing edge preservation to obtain an optimized surface image. .

[0053] Step 4: Perform clear surface feature extraction and description; optimize the surface image. Perform the following operations to obtain quantifiable surface features: Keypoint detection: Using the Accelerated Robust Feature Detection (SURF) or Oriented Fast and Rotated BRIEF (ORB) detectors to optimize surface images. Extract key points within the cargo area.

[0054] Feature description: Calculate the descriptor vector (such as SURF descriptor, ORB descriptor) for each keypoint.

[0055] Feature set construction: Organize the coordinates of all key points and their descriptor vectors into a cargo surface feature set. .

[0056] Step 5: Analyze the set of surface features of the cargo The recurrence rate of feature points across multiple consecutive frames and descriptor consistency. If the stability score of the current feature point is lower than the feature stability threshold... If a feature is not found in the set, it will be removed from the set. The final output is a stable and clear set of cargo surface features. .

[0057] S5: Extract cargo surface features from the cargo surface feature set, match them with preset multi-scale cargo judgment indicators to obtain cargo categories, and output preliminary identification labels.

[0058] Furthermore, in step S5, the step of extracting cargo surface features from the cargo surface feature set and matching them with preset multi-scale cargo judgment indicators to obtain cargo categories and outputting preliminary identification labels specifically includes the following steps: S51: Extract a clear cargo surface image from the cargo surface feature set, and perform segmentation and enhancement processing on the cargo surface image using image processing technology to obtain clear surface information data.

[0059] S52: Use multi-scale analysis to extract features from surface information data at different levels and determine the multi-scale feature set.

[0060] S53: Construct a corresponding judgment index system based on the multi-scale feature set, perform index calculation and normalization on the multi-scale feature set, and obtain standardized index data.

[0061] That is, based on the multi-scale feature set, a corresponding judgment index system is constructed, and standardized index data is obtained through index calculation and normalization.

[0062] S54: Compare the standardized indicator data with the category parameters in the preset model. Output the matching data as preliminary category labels; record the non-matching data as pending and proceed to subsequent processing. That is, if the standardized indicator data matches the category parameters in the preset model, output the preliminary category labels; if they do not match, record them as pending and proceed to subsequent processing.

[0063] Furthermore, the preset model is a category determination model based on multi-scale cargo features. Its core function is to compare the standardized indicator data with the predefined cargo category parameters to achieve preliminary matching and label output of cargo categories.

[0064] The preset model is a "feature-category mapping model" pre-trained and stored in the system. It is a "multi-scale feature-to-cargo category" correspondence library constructed from historical cargo sample data, containing two core components: Goods Category Parameter Library: A set of "standard feature parameters" stored according to the type of goods (such as food boxes, medicine boxes, tool bags, disaster relief supply packages, etc.). Each category corresponds to one or more sets of feature parameter thresholds (such as texture density, edge contour ratio, surface marking features, etc.).

[0065] Matching rule engine: Defines the comparison logic of "standardized indicator data and category parameters" (such as Euclidean distance matching, cosine similarity matching, threshold interval determination, etc.) to determine whether the input features meet the standard features of a certain type of goods.

[0066] S55: Perform a secondary comparison between the surface parameters associated with the initial category label and the surface information data to determine the final recognition result. That is, for the initial category label, perform a secondary comparison with the surface information data, and determine the final recognition result through preset logical rules.

[0067] S56: Generate preliminary identification labels based on the final identification results.

[0068] S6: Obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy.

[0069] Furthermore, in step S6, the step of obtaining associated preset cargo status data based on the preliminary identification tag, verifying the consistency between the cargo status data and the cargo surface features, and determining the final identification accuracy includes the following steps: S61: Extract cargo status information related to the initial identification tag from the database, and sort it according to preset rules to obtain the initial sorted status dataset; Furthermore, the preset rules include: 1. Data filtering rules: The original cargo status information extracted from the database may contain redundant data, and irrelevant / invalid information needs to be removed; 2. Field standardization rules: Standardize the format of the filtered fields to achieve "association mapping between actual feature data and status data"; 3. Hierarchical division rules: Based on the importance of the cargo status to "label consistency verification", the status information is divided into a core layer and an auxiliary layer to facilitate subsequent priority comparison of key features; 4. Association Mapping Rules: Establish the correspondence between status data and identification tags to ensure that each piece of status data can be accurately associated with the corresponding preliminary identification tag.

[0070] S62: Obtain the actual feature data of the cargo surface corresponding to each preliminary identification label and perform association mapping with the data in the preliminary sorting status dataset to form a feature mapping set; the association mapping refers to establishing a one-to-one or many-to-many association relationship between the actual feature data of the cargo surface corresponding to the preliminary identification label and the preset standard data in the preliminary sorting cargo status dataset according to preset rules, forming a structured feature mapping set, which provides a matching basis for subsequent verification of the "consistency between label and actual features" through convolutional neural networks.

[0071] S63: Use a convolutional neural network to process the feature mapping set, analyze the degree of matching between the preliminary identification labels and the actual feature data, and determine the feature consistency result; S64: If the feature consistency result is lower than the preset threshold, the feature mapping set is extracted a second time to obtain supplementary actual feature data and determine whether the consistency has been improved. S65: By classifying and organizing the results after consistency improvement, the final identified label data is labeled to obtain the basis for accuracy evaluation; S66: Based on the accuracy assessment criteria, a logical comparison tool is used to verify the final identification result and determine the final matching conclusion between the cargo surface features and the preliminary identification label. Furthermore, the logical comparison tool is an automated algorithm or module used to verify the "final matching conclusion between cargo surface features and the preliminary identification label," ensuring the accuracy, consistency, and reasonableness of the identification results. Specifically, it includes two parts: a "preset logical rule base" and a "data verification engine." 1. The preset logical rule base includes: Core matching rules: Verify the consistency of core fields between "actual surface feature data of goods" and "preliminary standard features corresponding to the identification label" (such as whether the texture integrity index is within the preset threshold range, whether the key surface feature points (such as packaging markings and material textures) are completely matched, and whether the size / contour ratio conforms to the standard parameters of this type of goods).

[0072] Anomaly exclusion rules: Define specific scenarios for "non-matching" (such as ≥2 inconsistent core feature fields, missing or mismatched key features (such as cargo-specific identification), and supplementary feature data deviating from the label standard deviation by more than a preset threshold (such as ±15%)).

[0073] Priority verification rules: Core layer features (such as texture, outline, key identifiers) are verified first, followed by auxiliary layer features (such as surface reflectivity, local details). If the core layer features do not match, they are directly judged as "invalid label". If the auxiliary layer features do not match, scene tolerance can be combined (such as reflectivity deviation caused by slight stains).

[0074] 2. The data validation engine includes: Data format verification: Ensure that the format of "cargo surface feature data" and "preset standard data" is consistent (such as standardized feature vectors, coded classification fields, and quantified index values) to avoid comparison errors caused by format differences.

[0075] Logical consistency verification: Verify the synergistic matching of multi-dimensional features through Boolean logic (AND / OR / NOT) (e.g., "texture feature matching", "contour feature matching" and "key identifier feature matching" must be matched to be considered a valid label; if any core feature does not match, it is considered an invalid label).

[0076] Threshold interval verification: Perform interval judgment on quantified feature indicators (such as texture integrity and feature similarity score) (e.g., if "feature consistency score ≥ 0.8 (preset threshold)" is determined to be a match, otherwise it is determined to be a mismatch).

[0077] S67: By storing the final matching results, an index structure is built for subsequent query needs to obtain traceable identification records.

[0078] Specifically, cargo status information related to identification tags is extracted from a database (data source), and this information is processed using automated tools to obtain a preliminary status dataset. Based on this dataset, corresponding actual feature data is obtained for each cargo status associated with an identification tag, forming a feature mapping set. A convolutional neural network is used to process the feature mapping set, analyzing the matching degree between the identification tag and the actual features to determine the feature consistency result. If the feature consistency result is lower than a preset threshold, the feature mapping set is further extracted to obtain supplementary actual feature data, and the consistency is assessed to determine if it has improved. The improved consistency results are categorized and organized, and the final identified tag data is labeled to obtain an accuracy evaluation criterion. Based on the accuracy evaluation criterion, a logical comparison tool is used to verify the final identification result, determining the final matching conclusion between the cargo status and the identification tag. The final matching conclusion is stored, and an index structure is built for subsequent query needs to obtain traceable identification records.

[0079] S7: Output cargo category and status report from the final identification accuracy to obtain complete cargo identification results for subsequent cargo transportation management.

[0080] From the category results after classification confirmation, key regional features in the cargo image data are obtained, and state analysis is performed on the key regional features to determine whether the cargo's state information meets the preset standards.

[0081] If the status information meets the preset criteria, the status information will be associated with and stored with the category result. If it does not meet the requirements, it is marked as an abnormal status and the abnormality details are recorded to obtain a complete description of the cargo status.

[0082] Based on the complete cargo status description, a structured identification record is generated. Data integration tools are used to uniformly format the category results and status descriptions to determine the final cargo identification profile.

[0083] Through the data transmission interface, the final cargo identification file is pushed to the cargo transportation management module, and the file content is allocated to the workflow to complete the systematic update of cargo information.

[0084] Beneficial effects: I. Deep Enhancement Effect of Environmental Robustness Multi-source interference collaborative suppression capability: The solution not only targets the two core interferences of "insufficient lighting and fog," but also integrates motion blur interference during drone flight through an environmental interference parameter set (lighting interference factor, fog interference factor, and motion blur factor). Furthermore, it dynamically matches different interference-dominant scenarios (such as high ISO noise, residual fog, and motion blur) through adaptive filtering, achieving robustness of "one solution for multiple complex environments," rather than targeted processing of a single interference. This collaborative suppression capability avoids recognition failure caused by changes in environmental interference type (such as insufficient lighting and slight motion blur due to cloudy days and light winds), adapting to the material delivery needs of various scenarios such as mountainous areas, rainy seasons, and winter.

[0085] Detail Preservation Capability in Extreme Environments: Through a combined strategy of "texture integrity calculation + template library repair," even under severe interference (such as dense fog causing texture loss exceeding 25%), key textures on the cargo surface (such as packaging labels and material textures) can still be restored through neighboring texture feature matching and a deep learning repair model, rather than simply discarding or blurring them. This detail preservation capability is crucial in emergency material delivery (such as disaster relief material packaging label identification), avoiding "class misjudgment" (such as confusing food boxes with medicine boxes) caused by texture loss.

[0086] II. Optimization Effects on Computational Resources and Operational Efficiency Computational power adaptability of layered processing: The solution adopts a layered logic of "preliminary processing - deep optimization" (e.g., in edge extraction, the initial edge is first obtained through traditional operators, and deep processing of the convolutional neural network is only triggered for low-intensity edge regions), avoiding the waste of computing power caused by indiscriminate deep learning of the entire image. For the embedded processor of the drone (with limited computing power), this mode of "lightweight preprocessing + heavyweight local optimization" can reduce power consumption and latency during real-time operation, ensuring that the drone can still output recognition results stably in scenarios with limited battery life.

[0087] The redundancy reduction effect of feature screening: In step S4, low-stability feature points are removed by "continuous multi-frame feature stability verification", which reduces the amount of computation in the subsequent matching and verification stages. The feature set that has not been screened may contain a large number of noisy features (such as temporary reflections and shadows of interfering objects). Screening can improve the feature matching efficiency by more than 30%, while reducing the memory usage of the convolutional neural network verification stage and adapting to more low-cost hardware platforms.

[0088] III. Improved System Scalability and Compatibility Flexible iteration capability of the template library: The preset texture template library adopts a dual data structure of "feature vector + high-quality image blocks", which supports the rapid access of new goods types (only the corresponding template data needs to be added, without modifying the core algorithm process). For example, when adding new types of goods such as "cold chain material boxes" and "fragile packaging", there is no need to reconstruct the recognition model. Only the LBP histogram / GLCM statistics and standard texture blocks of this type of goods need to be added to expand the recognition range and reduce the system upgrade cost.

[0089] Multi-device and multi-data-format compatibility: Step S11, “Preliminary format unification and noise filtering”, does not limit specific image formats (such as JPG, PNG, RAW), and can adapt to the image output formats of different drone models; at the same time, the environmental interference parameter set can be obtained from multiple sources such as drone IMU sensors and image EXIF ​​data, without relying on specific sensor configurations, thus improving the portability of the solution on drones of different brands and configurations.

[0090] IV. Supporting Effect on Business Implementation Traceability and troubleshooting capabilities: Step S67, “building an index structure and storing identification records”, not only meets compliance requirements but also provides data support for subsequent business optimization. For example, by analyzing “low consistency” cases in historical identification records, the template library can be optimized in reverse (by supplementing texture templates for frequently misjudged goods) and threshold parameters can be adjusted (such as texture integrity threshold and feature stability threshold), forming a closed loop of “identification-feedback-optimization” to continuously improve the system’s adaptability in actual business.

[0091] Potential for refined cargo status assessment: The solution not only outputs "cargo category" but also implicitly assesses the integrity of cargo packaging through indicators such as texture completeness and surface feature consistency (e.g., texture breaks may correspond to damaged packaging, and abnormal surface features may correspond to moisture damage). This refined assessment can assist in subsequent transportation management (e.g., prioritizing the delivery of intact goods and conducting secondary verification of damaged goods), exceeding the basic requirement of "simple category identification" and providing data support for the full-process control of the logistics chain.

[0092] V. Enhanced resistance to misjudgment and fault tolerance The false positive rate is reduced through dual verification: By using "multi-scale indicator matching (preliminary label) + convolutional neural network consistency verification (final label)," the risk of false positives from a single matching logic is avoided. For example, when the surface features of goods are similar to a certain type of goods template due to interference (false positive in the preliminary label), a secondary verification can be performed by associating with preset goods status data (such as size and material features), which can reduce the false positive rate to below 1%, making it particularly suitable for the accurate identification of high-value, high-priority materials.

[0093] Error tolerance for incomplete areas: For areas with incomplete textures (such as partially occluded goods), the solution does not abandon recognition outright. Instead, it achieves effective recognition through "local complete texture matching + repair," overcoming the limitation that "a complete image of the goods is required for recognition." In actual deployment scenarios, goods may be partially invisible due to stacking or occlusion. This error tolerance can improve the recognition success rate by more than 20%, reducing "unrecognizable" situations caused by occlusion.

[0094] Example 2 like Figure 2 As shown, a cargo intelligent identification system for a material delivery drone is used to implement the above-described cargo intelligent identification method for a material delivery drone, including: The enhancement processing module is used to process the original images acquired by the UAV using image enhancement algorithms to obtain enhanced images; The edge enhancement module is used to extract the edge structure information of the cargo based on the enhanced image; The optimization processing module is used to segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture optimization image; based on the texture optimization image, environmental interference parameters are fused, and a filtering algorithm is used to suppress interference, and then the cargo surface feature descriptor and cargo surface feature set are extracted and output. The preliminary label recognition module is used to extract cargo surface features from the cargo surface feature set, match them with preset multi-scale cargo judgment indicators to obtain cargo categories, and output preliminary recognition labels. The verification module is used to obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy. The report output module is used to output cargo category and status reports from the final identification accuracy, obtaining complete cargo identification results for subsequent cargo transportation management.

[0095] Example 3 A computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device where the computer-readable storage medium is located to perform the above-described intelligent cargo identification method for unmanned aerial vehicles (UAVs) used for material delivery.

[0096] Example 4 A processor for running a program, wherein the program executes the above-described intelligent cargo identification method for unmanned aerial vehicles (UAVs) used for material delivery.

[0097] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This application discloses a cargo recognition method based on UAV image processing. Addressing the business scenario problem of image blurring and loss of cargo edge and texture details caused by insufficient lighting and fog interference when UAVs acquire images, it achieves accurate cargo recognition through a series of technical means. Specifically, firstly, the original image acquired by the UAV is processed using an image enhancement algorithm to remove interference caused by insufficient lighting and fog, resulting in an enhanced image used to restore edge and texture details of the cargo features. Based on the enhanced image, preliminary cargo edge information is extracted. If the cargo edge intensity is lower than a preset threshold, a convolutional neural network is applied for deep feature learning to determine the enhanced cargo edge structure and output an enhanced cargo edge binary image. Based on the enhanced cargo edge structure, the cargo target region is segmented, and its texture detail distribution is extracted and analyzed. By calculating a texture integrity index and performing similarity matching and repair with a preset texture template library, a texture-optimized image is generated. Based on texture-optimized images, this invention integrates environmental interference parameters obtained from image metadata and sensors, employs a parameter-adaptive filtering algorithm to suppress specific interference, and then extracts clear cargo surface feature descriptors and outputs a cargo surface feature set. From this set, clear cargo surface features are extracted, and multi-scale cargo judgment indicators are derived. If the indicators match a preset cargo category model, a preliminary identification label is output. Based on the preliminary identification label, associated cargo status data is obtained, and the consistency between the label and actual features is verified using a convolutional neural network to determine the final identification accuracy. From the final identification accuracy, a cargo category and status report are output, yielding a complete cargo identification result. Analysis of the above execution scheme shows that this invention first processes the original image using an image enhancement algorithm to restore cargo edges and texture details. Then, it utilizes deep learning with a convolutional neural network to strengthen the cargo edge structure and optimizes texture distribution by combining texture template similarity comparison. Simultaneously, for the remaining blurred areas, this invention integrates environmental interference parameters, employs adaptive filtering technology to extract clear cargo surface features, and ensures consistency between the identification label and actual features through multi-scale indicator matching and status data verification, ultimately outputting a cargo category and status report. The core innovation of this invention lies in combining image enhancement, deep learning, and adaptive filtering, which solves the problems of image quality in complex environments and the difficulty of cargo feature extraction, and significantly improves the accuracy and efficiency of cargo identification in cargo transportation management.

[0098] This application discloses a method, system, medium, and processor for intelligent cargo identification using drones for material delivery, relating to the field of image cargo recognition technology. Addressing the problems of image blurring and loss of cargo edge and texture details caused by insufficient lighting and fog interference during drone image acquisition, this invention uses image enhancement algorithms to process the original image to obtain an enhanced image, extracting and strengthening cargo edge structure information, segmenting the target region, calculating texture integrity indices, and combining a preset template library to repair textures and generate an optimized image. Environmental interference parameters are fused, and adaptive filtering is used to extract a set of cargo surface features. Multi-scale judgment indicators are matched to output preliminary identification labels, and a convolutional neural network is used to verify the consistency between the labels and actual features. Finally, a cargo category and status report are output. This invention combines image enhancement, deep learning, and adaptive filtering to improve the accuracy and efficiency of cargo identification in complex environments, providing reliable support for cargo transportation management.

[0099] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0100] In the embodiments provided by the present invention, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0101] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0102] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention 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 or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of this application.

Claims

1. A method for intelligent cargo identification for unmanned aerial vehicles (UAVs) used for material delivery, characterized in that, Includes the following steps: S1: The original image acquired by the drone is processed using an image enhancement algorithm to obtain the enhanced image; S2: Extract cargo edge structure information based on the enhanced image; S3: Segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture-optimized image; S4: Based on texture optimization, image fusion environmental interference parameters are used, and filtering algorithms are employed to suppress interference. Then, the output cargo surface feature descriptor and cargo surface feature set are extracted. S5: Extract cargo surface features from the cargo surface feature set and match them with preset multi-scale cargo judgment indicators to obtain cargo categories and output preliminary identification labels; S6: Obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy; S7: Output cargo category and status report from the final identification accuracy to obtain complete cargo identification results for subsequent cargo transportation management.

2. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S1, the original image acquired by the drone is processed using an image enhancement algorithm to obtain an enhanced image, specifically including the following steps: S11: The raw images acquired by the drone are subjected to preliminary format unification and noise filtering to obtain the preliminarily processed images; S12: Detect the illumination of the pre-processed image, and perform global illumination compensation on the insufficient image using brightness adjustment technology to obtain the illumination-corrected image; S13: For the image after illumination correction, the histogram equalization method in the image enhancement algorithm is used to adjust the contrast in order to highlight the edge information of the goods in the image, and obtain the image with improved contrast. S14: Detect fog interference areas in the image after contrast enhancement. If fog interference areas are detected, apply defogging technology to clarify the corresponding areas and obtain the defogging image. S15: Extract cargo edge information from the dehazed image to enhance cargo edge details, resulting in a cargo edge-enhanced image; S16: Obtain texture feature information from the image, perform detail reconstruction processing on the enhanced image of the cargo edge to restore texture details, and obtain the final enhanced image.

3. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S2, extracting cargo edge structure information based on the enhanced image includes the following steps: S21: Obtain initial cargo edge information of the enhanced image through a pre-established cargo edge detection operator to obtain preliminary cargo edge distribution data; S22: Calculate the edge strength value of the cargo based on the preliminary cargo edge distribution data. If the calculated strength value is lower than the preset threshold, trigger the deep processing flow to determine the area that needs further processing. S23: For areas requiring further processing, use a convolutional neural network to extract deep features from the image data and obtain fine feature information; S24: The extracted fine feature information is reorganized to construct an enhanced cargo edge structure, resulting in an optimized cargo edge representation; S25: Apply smoothing technology to the optimized cargo edge representation to eliminate noise interference and obtain continuous cargo edge lines; S26: Structure and integrate continuous cargo edge lines to generate the final cargo edge structure information.

4. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S3, the process of segmenting the target cargo region based on cargo edge structure information, extracting the texture of the target cargo region and calculating the texture integrity index, and then performing similarity matching and repair with a preset texture template to generate a texture-optimized image includes the following steps: S31: Based on the cargo edge structure information, a contour lookup and filling algorithm is used to determine the closed region where the cargo body is located, which serves as the cargo target region mask; based on the cargo target region mask, the corresponding cargo region image is cropped from the enhanced image, denoted as... The local binary mode or gray-level co-occurrence matrix algorithm is used to process the cargo area image. Processing is performed to generate the corresponding texture feature map. ; S32: Texture feature map Adaptive thresholding is performed to obtain a binarized texture saliency map. ; for binarized texture saliency maps Morphological closing operations are performed to fill tiny holes and connect neighboring texture pixels. Subsequently, connected component analysis is performed, and regions with a pixel value of 1 and which are interconnected are marked as a single "texture contiguous region". The texture contiguous region is defined as one with an area greater than a preset minimum area threshold. Connected components; calculate texture integrity index : ； In the above formula, The total number of pixels in all contiguous texture regions; Masking the target area of ​​the cargo Total number of pixels within; S33: Set integrity threshold ;like If the texture is complete, the cargo area image is directly processed. As a texture-optimized image Output; if If the texture is incomplete, the incomplete area is recorded, triggering the next step of the texture matching and repair process based on the template library; S34: The texture matching and repair process based on the template library includes: establishing a preset texture template library; each template in the preset texture template library contains two parts of data: texture feature vector. and corresponding high-quality texture image blocks For each incomplete region identified, feature vectors are extracted from the surrounding complete texture regions. ; Calculate the eigenvectors Compared with all texture feature vectors in the preset texture template library The cosine similarity is used; the template k with the highest similarity is selected as the matching template; high-quality texture image patches of the matching template are used. Histogram matching and adaptive affine transformation are performed based on the local statistical characteristics of the incomplete region to generate a repair patch. Then, using Poisson image editing or a deep learning-based image inpainting model, the repair patch is seamlessly integrated into the incomplete region to obtain the repaired cargo area image. ; Image of the repaired cargo area The image is blended with the original image background to output the final texture-optimized image. .

5. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S4, the process of optimizing image fusion environmental interference parameters based on texture, suppressing interference using a filtering algorithm, and then extracting the output cargo surface feature descriptor and cargo surface feature set includes the following steps: S41: Define the set of environmental interference parameters This set includes the following parameters: light interference factor Fog interference factors Motion fuzziness factor Input texture optimization image, based on interference parameter set Adaptive filtering is performed on the texture-optimized image to obtain a preliminary filtered image. ; S42: In the initial filtered image Within the target area of ​​the goods, calculate the variance of the grayscale gradient magnitude for each local block; if the variance value of a certain local block is lower than the preset grayscale gradient variance threshold... If the area is identified as a residual blurred region, soft thresholding based on wavelet transform is applied to denoise these regions to obtain an optimized surface image. ; Furthermore, the formula for calculating the variance of the grayscale gradient magnitude for each local block is as follows: In the above formula, Let V be the variance of the grayscale gradient magnitude of the k-th local block; (This corresponds to an 8×8 pixel block size for a local block). The grayscale gradient magnitude at coordinates (x, y) within the local block; The average grayscale gradient magnitude of the k-th local block; S43: Optimize surface image Keypoint detection and feature description are performed, and the coordinates and descriptor vectors of all keypoints are organized together into a set of cargo surface features, as shown in the following formula: ； in, The pixel coordinates of the i-th key point; Let be the feature descriptor vector of the i-th keypoint; n is the total number of keypoints. S44: Analyze the set of cargo surface features The recurrence rate and descriptor consistency of feature points across multiple consecutive frames; if the stability score of the current feature point is lower than the feature stability threshold. If the condition is not met, it will be removed from the set; finally, a clear set of cargo surface features will be obtained and output.

6. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S5, the step of extracting cargo surface features from the cargo surface feature set and matching them with preset multi-scale cargo judgment indicators to obtain cargo categories and outputting preliminary identification labels specifically includes the following steps: S51: Extract a clear cargo surface image from the cargo surface feature set, and perform segmentation and enhancement processing on the cargo surface image using image processing technology to obtain clear surface information data; S52: Use multi-scale analysis to extract features from surface information data at different levels and determine the multi-scale feature set; S53: Construct a corresponding judgment index system based on the multi-scale feature set, perform index calculation and normalization on the multi-scale feature set, and obtain standardized index data; S54: Compare the standardized indicator data with the category parameters in the preset model, output the matching data as the initial category label, and record the non-matching data as pending and transfer it to subsequent processing; S55: Perform a secondary comparison between the surface parameters associated with the initial category label and the surface information data to determine the final identification result; S56: Generate preliminary identification labels based on the final identification results.

7. The intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery according to claim 1, characterized in that, In step S6, the step of obtaining the associated preset cargo status data based on the preliminary identification tag, verifying the consistency between the cargo status data and the cargo surface features, and determining the final identification accuracy includes the following steps: S61: Extract cargo status information related to the initial identification tag from the database, and sort it according to preset rules to obtain the initial sorted status dataset; S62: Obtain the actual feature data of the cargo surface features corresponding to each preliminary identification tag and associate and map it with the data in the preliminary sorting status dataset to form a feature mapping set; S63: Use a convolutional neural network to process the feature mapping set, analyze the degree of matching between the preliminary identification labels and the actual feature data, and determine the feature consistency result; S64: If the feature consistency result is lower than the preset threshold, the feature mapping set is extracted a second time to obtain supplementary actual feature data and determine whether the consistency has been improved. S65: By classifying and organizing the results after consistency improvement, the final identified label data is labeled to obtain the basis for accuracy evaluation; S66: Based on the accuracy assessment criteria, a logical comparison tool is used to verify the final identification results and determine the final matching conclusion between the surface features of the goods and the preliminary identification label; S67: By storing the final matching results, an index structure is built for subsequent query needs to obtain traceable identification records.

8. A cargo intelligent identification system for unmanned aerial vehicles (UAVs) used for material delivery, characterized in that, The intelligent cargo identification method for implementing the unmanned aerial vehicle (UAV) for material delivery as described in any one of claims 1 to 7 includes: The enhancement processing module is used to process the original images acquired by the UAV using image enhancement algorithms to obtain enhanced images; The edge enhancement module is used to extract the edge structure information of the cargo based on the enhanced image; The optimization processing module is used to segment the target region of the cargo based on the edge structure information of the cargo, extract the texture of the target region of the cargo and calculate the texture integrity index, and then perform similarity matching and repair with the preset texture template to generate a texture optimization image; based on the texture optimization image, environmental interference parameters are fused, and a filtering algorithm is used to suppress interference, and then the cargo surface feature descriptor and cargo surface feature set are extracted and output. The preliminary label recognition module is used to extract cargo surface features from the cargo surface feature set, match them with preset multi-scale cargo judgment indicators to obtain cargo categories, and output preliminary recognition labels. The verification module is used to obtain the associated preset cargo status data based on the initial identification tag, verify the consistency between the cargo status data and the cargo surface features, and determine the final identification accuracy. The report output module is used to output cargo category and status reports from the final identification accuracy, obtaining complete cargo identification results for subsequent cargo transportation management.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein, when the program is executed, it controls the device containing the computer-readable storage medium to perform the intelligent cargo identification method for unmanned aerial vehicles used for material delivery as described in any one of claims 1 to 7.

10. A processor, characterized in that, The processor is used to run a program, wherein the program executes the intelligent cargo identification method for unmanned aerial vehicles (UAVs) for material delivery as described in any one of claims 1 to 7.

Images