Door-to-door signing image quality inspection method, device and equipment and storage medium

Abstract

The invention relates to the technical field of information processing, in particular to a door-to-door signing image quality inspection method, device and equipment and a storage medium, according to a preset flow distribution proportion, a preprocessed image is distributed to a plurality of quality inspection models for parallel reasoning, reasoning results are output respectively, the reasoning results comprise compliance judgment and confidence coefficient, and the accuracy of the quality inspection model is improved. The method comprises the following steps: respectively endowing a plurality of quality inspection models with corresponding model weights based on historical accuracy, carrying out weighted fusion on compliance judgment of the plurality of quality inspection models based on the model weights and confidence, generating final quality inspection result information, and based on multi-model collaborative decision, giving consideration to the maturity of a third-party model and the controllability and optimization space of a self-research model. The performance indexes of all the quality inspection models are monitored, the flow distribution proportion and the model parameters are dynamically adjusted based on the performance indexes, smooth transition and performance optimization between the models are achieved, and the cost is reduced while the quality inspection accuracy is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method, apparatus, equipment, and storage medium for on-site delivery image quality inspection. Background Technology

[0002] In the express delivery industry, couriers are required to take standardized photos of packages upon delivery to ensure service compliance. Currently, a common practice is to use a single AI model for image quality inspection. However, using a single AI model suffers from limited accuracy, insufficient generalization ability, and an inability to flexibly adapt to different scenarios. Although some express delivery companies have attempted to introduce third-party AI services, such as Alibaba Cloud's Visual AI, challenges remain, including model black-box issues, data privacy concerns, and cost control. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, the present invention aims to provide a method, device, equipment, and storage medium for on-site signing image quality inspection based on multi-model collaborative decision-making, which can take into account the maturity of third-party models and the controllability and optimization space of self-developed models, dynamically adjust the traffic allocation ratio based on the real-time performance of models, achieve smooth transition and optimal performance between models, and reduce costs while ensuring the accuracy of quality inspection.

[0004] The first aspect of this invention provides a method for quality inspection of on-site delivery images, comprising: receiving an on-site delivery image sent by a user terminal; preprocessing the on-site delivery image to obtain a preprocessed image; distributing the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio; performing parallel inference on the preprocessed image through the multiple quality inspection models and outputting inference results respectively, the inference results including compliance judgments and confidence levels; assigning corresponding model weights to the multiple quality inspection models based on historical accuracy; performing weighted fusion of the compliance judgments of the multiple quality inspection models based on the model weights and the confidence levels to generate final quality inspection result information; recording the model information used, inference time information, and image ID information; sending the model information used, the inference time information, the image ID information, and the final quality inspection result information to the user terminal; monitoring the performance indicators of each quality inspection model; and dynamically adjusting the traffic allocation ratio and model parameters based on the performance indicators.

[0005] Optionally, in a first implementation of the first aspect of the present invention, receiving the door-to-door delivery image sent by the user terminal and preprocessing the door-to-door delivery image to obtain a preprocessed image includes: receiving the door-to-door delivery image sent by the user terminal; converting the door-to-door delivery image according to a preset standard format; adjusting the size of the door-to-door delivery image according to a preset resolution and maintaining the image ratio of the door-to-door delivery image through a bilinear interpolation algorithm; applying image enhancement technology to enhance the door-to-door delivery image, the image enhancement technology including histogram equalization and Gaussian filtering for noise reduction; and normalizing the pixel values ​​of the door-to-door delivery image to obtain a preprocessed image.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the step of distributing the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio, performing parallel inference on the preprocessed image through the multiple quality inspection models, and outputting inference results respectively, wherein the inference results include compliance judgment and confidence level, includes: distributing the preprocessed image to multiple quality inspection models using a load balancing algorithm according to a preset traffic allocation ratio, wherein the quality inspection models include an image classification model based on a convolutional neural network and an object detection model based on YOLO; performing parallel inference on the preprocessed image through the multiple quality inspection models, and using the image classification model to output the overall compliance judgment and confidence level of the image as the inference result, and using the object detection model to output the compliance judgment and confidence level of a specific region in the image as the inference result; parsing the inference result of each quality inspection model to extract the compliance judgment and the confidence level; and storing the compliance judgment and the confidence level in a temporary result pool.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of assigning corresponding model weights to multiple quality inspection models based on historical accuracy, and performing weighted fusion of the compliance judgments of the multiple quality inspection models based on the model weights and the confidence level to generate final quality inspection result information, includes: querying the historical accuracy of each quality inspection model from the model performance database, calculating model weights based on the historical accuracy; quantifying the compliance judgment of each quality inspection model; fusing the quantified compliance judgments using a weighted average algorithm based on the model weights and the confidence level to generate a fusion score; generating compliant final quality inspection result information when the fusion score is greater than or equal to a preset score threshold; and generating non-compliant final quality inspection result information when the fusion score is less than the score threshold.

[0008] Optionally, in the fourth implementation of the first aspect of the present invention, the step of recording the model information used, inference time information, and image ID information, and sending the model information used, the inference time information, the image ID information, and the final quality inspection result information to the user terminal includes: recording the model information used, which includes the model name, version number, and parameter configuration; recording the inference time information, which includes the inference start time and end time; recording the image ID information, which includes a unique image identifier and a user terminal identifier; integrating the model information used, the inference time information, the image ID information, and the final quality inspection result information and encapsulating them into JSON format data to obtain integrated information; and sending the integrated information to the user terminal via the HTTP protocol and receiving a confirmation response from the user terminal.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, monitoring the performance indicators of each of the quality inspection models and dynamically adjusting the traffic allocation ratio and model parameters based on the performance indicators includes: real-time monitoring of the performance indicators of each of the quality inspection models, the performance indicators including accuracy, response time, and resource utilization; calculating a comprehensive score for each of the quality inspection models based on the performance indicators within a time window, and dynamically adjusting the traffic allocation ratio based on the comprehensive score; updating the adjusted traffic allocation ratio to the system configuration center; and adjusting the model parameters of the corresponding quality inspection model when the accuracy of the performance indicator of the quality inspection model is less than a preset threshold.

[0010] Optionally, in the sixth implementation of the first aspect of the present invention, after monitoring the performance indicators of each of the quality inspection models and dynamically adjusting the traffic allocation ratio and model parameters based on the performance indicators, the method further includes: summarizing the quality inspection records of all the on-site delivery images, generating a quality inspection statistical report based on the quality inspection records; encrypting the quality inspection statistical report to obtain encrypted report information; and uploading the encrypted report information to the blockchain.

[0011] A second aspect of the present invention provides a door-to-door delivery receipt image quality inspection device, comprising: a receiving and preprocessing module, configured to receive a door-to-door delivery receipt image sent by a user terminal, preprocess the door-to-door delivery receipt image to obtain a preprocessed image; a distribution and inference output module, configured to distribute the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio, perform parallel inference on the preprocessed image through the multiple quality inspection models, and output inference results respectively, the inference results including compliance judgments and confidence levels; an assignment and fusion generation module, configured to assign corresponding model weights to the multiple quality inspection models based on historical accuracy, and perform weighted fusion of the compliance judgments of the multiple quality inspection models based on the model weights and the confidence levels to generate final quality inspection result information; a recording and sending module, configured to record the model information used, inference time information, and image ID information, and send the model information used, the inference time information, the image ID information, and the final quality inspection result information to the user terminal; and a monitoring and adjustment module, configured to monitor the performance indicators of each quality inspection model, and dynamically adjust the traffic allocation ratio and model parameters based on the performance indicators.

[0012] Optionally, in a first implementation of the second aspect of the present invention, the receiving preprocessing module includes: a receiving unit for receiving a door-to-door delivery receipt image sent by a user terminal; a conversion unit for converting the door-to-door delivery receipt image into a format based on a preset standard format; an adjustment and holding unit for adjusting the size of the door-to-door delivery receipt image based on a preset resolution and maintaining the image ratio of the door-to-door delivery receipt image through a bilinear interpolation algorithm; an enhancement processing unit for applying image enhancement techniques to enhance the door-to-door delivery receipt image, the image enhancement techniques including histogram equalization and Gaussian filtering for noise reduction; and a normalization processing unit for normalizing the pixel values ​​of the door-to-door delivery receipt image to obtain a preprocessed image.

[0013] Optionally, in a second implementation of the second aspect of the present invention, the distribution inference output module includes: a distribution unit, configured to distribute the preprocessed image to multiple quality inspection models using a load balancing algorithm according to a preset traffic allocation ratio, wherein the quality inspection models include an image classification model based on a convolutional neural network and an object detection model based on YOLO; an inference output unit, configured to perform parallel inference on the preprocessed image through multiple quality inspection models, and use the image classification model to output the overall compliance judgment and confidence level of the image as the inference result, and use the object detection model to output the compliance judgment and confidence level of a specific region in the image as the inference result; a parsing and extraction unit, configured to parse the inference result of each quality inspection model and extract the compliance judgment and the confidence level; and a storage unit, configured to store the compliance judgment and the confidence level in a temporary result pool.

[0014] Optionally, in a third implementation of the second aspect of the present invention, the fusion generation module includes: a query calculation unit, configured to query the historical accuracy of each of the quality inspection models from the model performance database and calculate model weights based on the historical accuracy; a numerical processing unit, configured to perform numerical processing on the compliance judgment of each of the quality inspection models; a fusion unit, configured to fuse the numerically quantified compliance judgments based on the model weights and the confidence level using a weighted average algorithm to generate a fusion score; a first generation unit, configured to generate compliant final quality inspection result information when the fusion score is greater than or equal to a preset score threshold; and a second generation unit, configured to generate non-compliant final quality inspection result information when the fusion score is less than the score threshold.

[0015] Optionally, in a fourth implementation of the second aspect of the present invention, the recording and sending module includes: a recording unit, used to record the model information used, the model information including model name, version number and parameter configuration, record inference time information, the inference time information including inference start time and end time, and record image ID information, the image ID information including image unique identifier and user terminal identifier; an integration and encapsulation unit, used to integrate the model information used, the inference time information, the image ID information and the final quality inspection result information and encapsulate them into JSON format data to obtain integrated information; and a sending unit, used to send the integrated information to the user terminal via the HTTP protocol and receive a confirmation response from the user terminal.

[0016] Optionally, in a fifth implementation of the second aspect of the present invention, the monitoring and adjustment module includes: a monitoring unit, configured to monitor the performance indicators of each quality inspection model in real time, the performance indicators including accuracy, response time, and resource utilization; a calculation and adjustment unit, configured to calculate the comprehensive score of each quality inspection model based on the performance indicators within a time window, and dynamically adjust the traffic allocation ratio according to the comprehensive score; an update unit, configured to update the adjusted traffic allocation ratio to the system configuration center; and an adjustment unit, configured to adjust the model parameters of the corresponding quality inspection model when the accuracy of the performance indicator of the quality inspection model is less than a preset threshold.

[0017] Optionally, in the sixth implementation of the second aspect of the present invention, it further includes: a summary generation module, used to summarize the quality inspection records of all the on-site delivery images and generate a quality inspection statistical report based on the quality inspection records; an encryption module, used to encrypt the quality inspection statistical report to obtain encrypted report information; and an upload module, used to upload the encrypted report information to the blockchain.

[0018] A third aspect of the present invention provides a door-to-door delivery image quality inspection device, the door-to-door delivery image quality inspection device comprising: a memory and at least one processor, the memory storing instructions; at least one processor calling the instructions in the memory to cause the door-to-door delivery image quality inspection device to perform each step of the door-to-door delivery image quality inspection method described above.

[0019] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed by a processor, implement the steps of the on-site delivery image quality inspection method described in any of the preceding claims.

[0020] In the technical solution of this invention, preprocessed images are distributed to multiple quality inspection models for parallel inference according to a preset traffic allocation ratio, and inference results are output respectively. The inference results include compliance judgment and confidence level. Based on historical accuracy, corresponding model weights are assigned to multiple quality inspection models. Based on model weights and confidence levels, the compliance judgments of multiple quality inspection models are weighted and fused to generate the final quality inspection result information. Based on multi-model collaborative decision-making, the maturity of third-party models and the controllability and optimization space of self-developed models can be taken into account. The performance indicators of each quality inspection model are monitored, and the traffic allocation ratio and model parameters are dynamically adjusted based on the performance indicators to achieve smooth transition and optimal performance between models, ensuring the accuracy of quality inspection while reducing costs. Attached Figure Description

[0021] Figure 1 This is a first flowchart of the on-site delivery image quality inspection method provided in an embodiment of the present invention; Figure 2 This is a second flowchart of the on-site delivery image quality inspection method provided in this embodiment of the invention; Figure 3 This is a third flowchart of the on-site signing image quality inspection method provided in this embodiment of the invention; Figure 4 This is a fourth flowchart of the on-site delivery image quality inspection method provided in this embodiment of the invention; Figure 5 This is a schematic diagram of a door-to-door delivery image quality inspection device provided in an embodiment of the present invention; Figure 6 This is another structural schematic diagram of the on-site delivery image quality inspection device provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the structure of the on-site signing image quality inspection equipment provided in an embodiment of the present invention. Detailed Implementation

[0022] This invention provides a method, device, equipment, and storage medium for on-site signing image quality inspection. Based on multi-model collaborative decision-making, it can balance the maturity of third-party models with the controllability and optimization space of self-developed models. It can dynamically adjust the traffic allocation ratio based on the real-time performance of the models, achieve smooth transition and optimal performance between models, and reduce costs while ensuring the accuracy of quality inspection.

[0023] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] For ease of understanding, the specific process of the embodiments of the present invention is described below. Please refer to [link / reference]. Figure 1 One embodiment of the on-site delivery image quality inspection method of the present invention includes: 101. Receive the door-to-door delivery confirmation image sent by the user terminal, preprocess the door-to-door delivery confirmation image to obtain a preprocessed image; In this embodiment, the system receives images of home delivery uploaded by the user through a network interface (such as a RESTful API). The image data may be transmitted in Base64 encoding or file format. The system first verifies the image format and size to ensure data integrity and security. Preprocessing includes format conversion (such as converting the image to a uniform JPEG format), size adjustment (using bilinear interpolation to maintain the image ratio according to a preset resolution, such as 224x224 pixels), image enhancement (applying histogram equalization to improve contrast and using Gaussian filtering to remove noise), and pixel value normalization (scaling pixel values ​​to the [0,1] range or standardizing them to the mean and standard deviation of the ImageNet dataset).

[0025] 102. Based on the preset traffic allocation ratio, the preprocessed image is distributed to multiple quality inspection models. The preprocessed image is then subjected to parallel inference by the multiple quality inspection models, and the inference results are output respectively. The inference results include compliance judgment and confidence level. In this embodiment, the system distributes preprocessed images to multiple quality inspection models using a load balancing algorithm (such as round-robin or weight-based allocation) according to a preset traffic allocation ratio (e.g., 30% of the traffic is allocated to the self-developed model and 70% to the third-party model). These models include image classification models based on convolutional neural networks (CNN) (such as ResNet) and object detection models based on YOLO. Each model performs parallel inference on the images. The image classification model outputs the overall compliance judgment (such as "compliant" or "non-compliant") and confidence score (a score between 0 and 1) of the image. The object detection model outputs the compliance judgment and confidence score of a specific region in the image. The inference results are temporarily stored for subsequent fusion.

[0026] 103. Based on historical accuracy, assign corresponding model weights to multiple quality inspection models. Based on model weights and confidence levels, perform weighted fusion of compliance judgments from multiple quality inspection models to generate final quality inspection results. In this embodiment, the system queries the historical accuracy of each quality inspection model from the model performance database (based on validation data over a period of time). The model weights are calculated based on the historical accuracy, for example, models with higher accuracy are given higher weights. Then, the compliance judgment of each model is numerically processed (e.g., "compliant" is 1, "non-compliant" is 0). Based on the model weights and confidence levels, a weighted average algorithm is used to calculate the fusion score. If the fusion score is greater than or equal to a preset threshold (e.g., 0.5), a compliant final quality inspection result is generated; otherwise, a non-compliant result is generated.

[0027] 104. Record the model information, inference time information, and image ID information used, and send the model information, inference time information, image ID information, and final quality inspection result information to the user terminal; In this embodiment, the system records detailed inference process information, including model information (model name, version number, parameter configuration), inference time information (start time and end time, used to calculate processing delay), and image ID information (unique image identifier and user terminal identifier). This information is integrated and encapsulated into JSON format data and sent back to the user terminal via HTTP protocol. After receiving the data, the user terminal returns an acknowledgment response to ensure reliable data transmission.

[0028] 105. Monitor the performance indicators of each quality inspection model, and dynamically adjust the flow allocation ratio and model parameters based on the performance indicators; In this embodiment, the system monitors the performance metrics of each quality inspection model in real time, including accuracy (based on manual review or test data), response time (inference latency), and resource utilization (CPU / GPU utilization). Based on the performance metrics within the time window, the system calculates the comprehensive score of each model (e.g., weighted average accuracy and response time). The system dynamically adjusts the traffic allocation ratio according to the comprehensive score, for example, by increasing the traffic ratio of well-performing models. The adjusted ratio is updated to the system configuration center. If the accuracy of a model is lower than a preset threshold, the system triggers model parameter adjustments, such as retraining or fine-tuning.

[0029] In this embodiment of the invention, images of on-site delivery are received and preprocessed, and then distributed to multiple quality inspection models for parallel inference based on a preset traffic allocation ratio. Inference results containing compliance judgments and confidence levels are obtained. Subsequently, weights are assigned based on the historical accuracy of each model, and the judgment results are weighted and fused in combination with the confidence level to generate the final quality inspection result. Finally, complete information is recorded and sent to the user terminal. At the same time, model performance is monitored and traffic and parameters are dynamically adjusted. Based on multi-model collaborative decision-making, the problem of limited accuracy and insufficient generalization ability of a single model is effectively solved. It takes into account the maturity of third-party models and the controllability of self-developed models. Through dynamic adjustment, a smooth transition between models and optimal system performance are achieved, reducing operating costs while ensuring the accuracy of quality inspection.

[0030] Please see Figure 2 The second embodiment of the on-site delivery image quality inspection method in this invention includes: 201. Receive the delivery confirmation image sent by the user; In this embodiment, the system receives image data uploaded by the user through a RESTful API. The image data may be transmitted in Base64 encoding or as a file. The system verifies the image format and size to ensure data integrity and security.

[0031] 202. Based on the preset standard format, convert the format of the on-site delivery signing image; In this embodiment, the system converts the uploaded image into a preset standard format, such as JPEG or PNG, to eliminate format differences. The system uses an image processing library (such as OpenCV) to perform the format conversion, ensuring that the image is compatible with subsequent processing steps.

[0032] 203. Based on the preset resolution, adjust the size of the door-to-door delivery image and maintain the image ratio of the door-to-door delivery image through bilinear interpolation algorithm; In this embodiment, the system adjusts the image size to a fixed resolution, such as 224x224 pixels, and uses a bilinear interpolation algorithm for scaling. Bilinear interpolation smoothly processes pixel values, reduces image distortion, and maintains visual consistency.

[0033] 204. Apply image enhancement techniques to enhance the images of the door-to-door delivery. Image enhancement techniques include histogram equalization and Gaussian filtering for noise reduction. In this embodiment, the system applies histogram equalization to enhance image contrast and make details clearer. Then, Gaussian filtering (kernel size such as 3x3) is used to remove noise and improve image quality, making it easier for the model to analyze accurately.

[0034] 205. Normalize the pixel values ​​of the on-site delivery receipt image to obtain a preprocessed image; In this embodiment, the system normalizes the image pixel values, scaling them from 0-255 to the range of 0-1, or standardizes them according to the requirements of the pre-trained model (e.g., subtracting the mean and dividing by the standard deviation). This makes the input data distribution consistent and improves the stability of the model.

[0035] 206. Based on the preset traffic allocation ratio, the preprocessed image is distributed to multiple quality inspection models using a load balancing algorithm. The quality inspection models include an image classification model based on a convolutional neural network and an object detection model based on YOLO. In this embodiment, the system distributes images to multiple model instances using a load balancing algorithm (such as minimum number of connections or hash allocation) based on the traffic allocation ratio (e.g., 30% for self-developed models and 70% for third-party models). The image classification model (e.g., ResNet-50) is responsible for overall compliance judgment, while the object detection model (e.g., YOLOv5) detects whether a specific object meets the requirements.

[0036] 207. Parallel reasoning is performed on the preprocessed image using multiple quality inspection models. The image classification model outputs the overall compliance judgment and confidence level of the image as the reasoning result, and the object detection model outputs the compliance judgment and confidence level of a specific region in the image as the reasoning result. In this embodiment, multiple quality inspection models process images simultaneously. The image classification model outputs a global label and confidence score, while the object detection model outputs bounding box coordinates, category, and confidence score. All inference results are collected for fusion.

[0037] 208. Analyze the inference results of each quality inspection model and extract compliance judgments and confidence levels; In this embodiment, the inference results of each quality inspection model are analyzed, and the compliance judgment and confidence level in the analysis information are extracted.

[0038] 209. Store compliance judgments and confidence levels in a temporary results pool; In this embodiment, the inference results are stored in a temporary result pool (such as Redis or a memory cache) in key-value pair format for quick access and subsequent fusion processing. The temporary pool is set with an expiration time to avoid data accumulation.

[0039] In this embodiment of the invention, the complete image preprocessing process is described in detail, including format conversion, size adjustment and aspect ratio preservation based on bilinear interpolation algorithm, image enhancement techniques using histogram equalization and Gaussian filtering, and pixel value normalization. This ensures that the image data input to the quality inspection model has a uniform format, stable quality, and significant features, laying a solid foundation for accurate inference of subsequent models. Furthermore, the preprocessed images are distributed to multiple models, including a CNN image classification model and a YOLO object detection model, for parallel inference using a load balancing algorithm according to the traffic allocation ratio. The overall and regional compliance judgments and confidence scores of the outputs are parsed and temporarily stored, providing a clear and operable implementation path for achieving efficient and reliable multi-model collaborative inference and result fusion.

[0040] Please see Figure 3 The third embodiment of the on-site signing image quality inspection method in this invention includes: 301. Query the historical accuracy of each quality inspection model from the model performance database, and calculate the model weight based on the historical accuracy; In this embodiment, the system queries the model performance database (such as MySQL or MongoDB) to obtain the accuracy of each model in a recent period (such as 7 days), calculates the model weight based on the historical accuracy, and the model weight is proportional to the accuracy. For example, weight = accuracy / total (accuracy of all models).

[0041] 302. Numericalize the compliance judgment of each quality inspection model; In this embodiment, the compliance judgment is converted into a numerical value, for example, compliance is 1 and non-compliance is 0.

[0042] 303. Based on model weights and confidence levels, a weighted average algorithm is used to fuse the numerical compliance judgments and generate a fusion score. In this embodiment, a weighted average algorithm is used to fuse the numerical compliance judgments based on model weights and confidence levels to generate a fusion score. The fusion score = Σ(model weights * confidence levels * numerical judgments). For example, if the two models output (compliant, 0.9) and (non-compliant, 0.8) respectively, with weights of 0.6 and 0.4 respectively, then the fusion score = 0.6 * 0.9 * 1 + 0.4 * 0.8 * 0 = 0.54.

[0043] 304. When the fusion score is greater than or equal to the preset score threshold, the final quality inspection result information that complies with regulations will be generated. In this embodiment, when the fusion score is greater than or equal to a preset score threshold, a compliant final quality inspection result is generated. For example, the preset score threshold is 0.5. If the fusion score is greater than or equal to 0.5, the final quality inspection result is compliant; otherwise, it is non-compliant. The score threshold can be adjusted according to business needs.

[0044] 305. When the fusion score is less than the score threshold, non-compliant final quality inspection result information will be generated; In this embodiment, when the fusion score is less than the score threshold, non-compliant final quality inspection result information is generated. For example, the preset score threshold is 0.5. If the fusion score is less than 0.5, a non-compliant result is generated, and an alarm or manual review process can be triggered.

[0045] 306. Record the model information used, including the model name, version number and parameter configuration; record the inference time information, including the inference start time and end time; and record the image ID information, including the image unique identifier and the user terminal identifier. In this embodiment, the system records detailed model information (such as ResNet-v1, YOLOv5s), inference timestamps, image IDs (such as UUIDs), and user IDs (such as courier IDs). This data is used for auditing and performance analysis.

[0046] 307. Integrate the model information, inference time information, image ID information, and final quality inspection result information used, and encapsulate them into JSON format data to obtain integrated information; In this embodiment, the model information, inference time information, image ID information, and final quality inspection result information are integrated and encapsulated into JSON format data. All information is integrated into a single JSON object to obtain integrated information.

[0047] 308. Send the integration information to the user terminal via the HTTP protocol and receive the user terminal's confirmation response; In this embodiment, the system sends JSON data to the callback URL specified by the user via an HTTP POST request. The user returns an acknowledgment response to ensure that the data is delivered.

[0048] In this embodiment of the invention, historical accuracy is queried from the model performance database to calculate model weights, and the compliance judgment of each model is numerically processed. Then, a weighted average algorithm is used to perform fusion calculation based on weights and confidence levels to generate a fusion score. Finally, the final quality inspection result is generated based on a preset threshold, forming a scientific and objective multi-model decision fusion mechanism, which effectively improves the accuracy and reliability of the final quality inspection result. At the same time, by recording the entire process data such as model information, inference time information, and image ID information used by the system, and integrating and encapsulating it into JSON format, it is sent to the user terminal via HTTP protocol and received and confirmed, realizing full-link traceability of the quality inspection process and reliable delivery of results, enhancing the transparency and reliability of the system.

[0049] Please see Figure 4 The fourth embodiment of the on-site delivery image quality inspection method in this invention includes: 401. Monitor the performance indicators of each quality inspection model in real time, including accuracy, response time and resource utilization. In this embodiment, the system uses monitoring tools (such as Prometheus) to collect model accuracy (through manual sampling verification), response time (time from request to response), and resource utilization (CPU / GPU memory usage), and the data is aggregated in real time.

[0050] 402. Calculate the comprehensive score of each quality inspection model based on the performance indicators within the time window, and dynamically adjust the flow allocation ratio according to the comprehensive score; In this embodiment, the overall score = w1 * accuracy + w2 * (1 / response time) + w3 * (1 / resource utilization), where w1, w2, and w3 are weights. Traffic is redistributed based on the overall score; for example, models with higher scores receive a higher traffic allocation ratio.

[0051] 403. Update the adjusted traffic allocation ratio to the system configuration center; In this embodiment, the adjusted ratio is pushed to the configuration center (such as ZooKeeper or Apollo), and all processing nodes are updated synchronously to ensure consistency.

[0052] 404. When the accuracy rate in the performance indicators of the quality inspection model is less than the preset threshold, the model parameters of the corresponding quality inspection model shall be adjusted. In this embodiment, if the accuracy of a certain model is lower than a threshold (such as 90%), the system triggers a model update process, such as retraining with new data or adjusting hyperparameters. After the model version is updated, it is redeployed.

[0053] 405. Summarize all quality inspection records for on-site delivery and sign-off images, and generate a quality inspection statistical report based on the quality inspection records; In this embodiment, the system periodically (e.g., daily) summarizes all quality inspection records and generates a quality inspection statistical report based on the records. The generated quality inspection statistical report includes compliance rate, model performance, common problems, etc., and is displayed in the form of charts.

[0054] 406. Encrypt the quality inspection statistical report to obtain encrypted report information; In this embodiment, an encryption algorithm (such as AES-256) is used to encrypt the quality inspection statistical report to protect sensitive data.

[0055] 407. Upload the encrypted report information to the blockchain; In this embodiment, encrypted report information is uploaded to a blockchain network, and the immutability of the blockchain is used to ensure the authenticity and auditability of the report.

[0056] In this embodiment of the invention, by real-time monitoring of key performance indicators such as accuracy, response time, and resource utilization of each quality inspection model, a comprehensive score is calculated based on the indicator data within the time window, and the traffic allocation ratio is dynamically adjusted. The system configuration is updated in a timely manner, and its parameters are adjusted when the model accuracy does not meet the standard. A complete model performance monitoring and adaptive optimization mechanism has been established to ensure that the system can continuously operate in the best state. In addition, by aggregating all quality inspection records to generate statistical reports, and encrypting the reports before uploading them to the blockchain for evidence storage, not only is data support provided for operational decisions, but the immutability of blockchain technology is also used to ensure the security and auditability of quality inspection data, providing a solid foundation of trust for the entire quality inspection process.

[0057] The above describes the on-site delivery image quality inspection method in the embodiments of the present invention. The following describes the on-site delivery image quality inspection device in the embodiments of the present invention. Please refer to [link / reference]. Figure 5 One embodiment of the on-site delivery image quality inspection device of the present invention includes: The receiving preprocessing module 501 is used to receive the door-to-door delivery signing image sent by the user terminal, and preprocess the door-to-door delivery signing image to obtain a preprocessed image. The distribution inference output module 502 is used to distribute the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio. The preprocessed image is then subjected to parallel inference by the multiple quality inspection models, and the inference results are output respectively. The inference results include compliance judgment and confidence level. The fusion generation module 503 is used to assign corresponding model weights to multiple quality inspection models based on historical accuracy, and to perform weighted fusion of the compliance judgments of multiple quality inspection models based on model weights and confidence levels to generate the final quality inspection result information. The recording and sending module 504 is used to record the model information used, inference time information, and image ID information, and send the model information used, inference time information, image ID information, and final quality inspection result information to the user terminal. The monitoring and adjustment module 505 is used to monitor the performance indicators of each quality inspection model and dynamically adjust the traffic allocation ratio and model parameters based on the performance indicators.

[0058] In this embodiment, according to a preset traffic allocation ratio, the preprocessed image is distributed to multiple quality inspection models for parallel inference, and the inference results are output respectively. The inference results include compliance judgment and confidence level. Based on the historical accuracy, corresponding model weights are assigned to the multiple quality inspection models. Based on the model weights and confidence levels, the compliance judgments of the multiple quality inspection models are weighted and fused to generate the final quality inspection result information. Based on multi-model collaborative decision-making, the maturity of third-party models and the controllability and optimization space of self-developed models can be taken into account. The performance indicators of each quality inspection model are monitored, and the traffic allocation ratio and model parameters are dynamically adjusted based on the performance indicators to achieve smooth transition and optimal performance between models, ensuring the accuracy of quality inspection while reducing costs.

[0059] Please see Figure 6 Another embodiment of the on-site delivery image quality inspection device of the present invention includes: The receiving preprocessing module 501 is used to receive the door-to-door delivery signing image sent by the user terminal, and preprocess the door-to-door delivery signing image to obtain a preprocessed image. The distribution inference output module 502 is used to distribute the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio. The preprocessed image is then subjected to parallel inference by the multiple quality inspection models, and the inference results are output respectively. The inference results include compliance judgment and confidence level. The fusion generation module 503 is used to assign corresponding model weights to multiple quality inspection models based on historical accuracy, and to perform weighted fusion of the compliance judgments of multiple quality inspection models based on model weights and confidence levels to generate the final quality inspection result information. The recording and sending module 504 is used to record the model information used, inference time information, and image ID information, and send the model information used, inference time information, image ID information, and final quality inspection result information to the user terminal. The monitoring and adjustment module 505 is used to monitor the performance indicators of each quality inspection model and dynamically adjust the traffic allocation ratio and model parameters based on the performance indicators. In this embodiment, the receiving preprocessing module 501 includes: a receiving unit 5011 for receiving a door-to-door delivery receipt image sent by a user terminal; a conversion unit 5012 for converting the door-to-door delivery receipt image into a format based on a preset standard format; an adjustment and holding unit 5013 for adjusting the size of the door-to-door delivery receipt image based on a preset resolution and maintaining the image ratio of the door-to-door delivery receipt image through a bilinear interpolation algorithm; an enhancement processing unit 5014 for applying image enhancement techniques to enhance the door-to-door delivery receipt image, including histogram equalization and Gaussian filtering for noise reduction; and a normalization processing unit 5015 for normalizing the pixel values ​​of the door-to-door delivery receipt image to obtain a preprocessed image.

[0060] In this embodiment, the distribution inference output module 502 includes: a distribution unit 5021, used to distribute the preprocessed image to multiple quality inspection models according to a preset traffic allocation ratio using a load balancing algorithm; the quality inspection models include an image classification model based on a convolutional neural network and an object detection model based on YOLO; an inference output unit 5022, used to perform parallel inference on the preprocessed image through multiple quality inspection models, and use the image classification model to output the overall compliance judgment and confidence level of the image as the inference result, and use the object detection model to output the compliance judgment and confidence level of a specific region in the image as the inference result; a parsing and extraction unit 5023, used to parse the inference result of each quality inspection model and extract the compliance judgment and confidence level; and a storage unit 5024, used to store the compliance judgment and confidence level in a temporary result pool.

[0061] In this embodiment, the fusion generation module 503 includes: a query calculation unit 5031, used to query the historical accuracy of each quality inspection model from the model performance database and calculate the model weight based on the historical accuracy; a numerical processing unit 5032, used to perform numerical processing on the compliance judgment of each quality inspection model; a fusion unit 5033, used to fuse the numerical compliance judgment based on the model weight and confidence level using a weighted average algorithm to generate a fusion score; a first generation unit 5034, used to generate compliant final quality inspection result information when the fusion score is greater than or equal to a preset score threshold; and a second generation unit 5035, used to generate non-compliant final quality inspection result information when the fusion score is less than the score threshold.

[0062] In this embodiment, the recording and sending module 504 includes: a recording unit 5041, used to record the model information used, including the model name, version number, and parameter configuration; record inference time information, including the inference start time and end time; and record image ID information, including a unique image identifier and a user terminal identifier; an integration and encapsulation unit 5042, used to integrate the model information used, inference time information, image ID information, and final quality inspection result information and encapsulate them into JSON format data to obtain integrated information; and a sending unit 5043, used to send the integrated information to the user terminal via the HTTP protocol and receive a confirmation response from the user terminal.

[0063] In this embodiment, the monitoring and adjustment module 505 includes: a monitoring unit 5051, used to monitor the performance indicators of each quality inspection model in real time, including accuracy, response time, and resource utilization; a calculation and adjustment unit 5052, used to calculate the comprehensive score of each quality inspection model based on the performance indicators within a time window, and dynamically adjust the traffic allocation ratio according to the comprehensive score; an update unit 5053, used to update the adjusted traffic allocation ratio to the system configuration center; and an adjustment unit 5054, used to adjust the model parameters of the corresponding quality inspection model when the accuracy of the performance indicator of the quality inspection model is less than a preset threshold.

[0064] In this embodiment, it also includes: a summary generation module 506, used to summarize the quality inspection records of all on-site signed images and generate a quality inspection statistical report based on the quality inspection records; an encryption module 507, used to encrypt the quality inspection statistical report to obtain encrypted report information; and an upload module 508, used to upload the encrypted report information to the blockchain.

[0065] above Figure 5 and Figure 6 The on-site delivery image quality inspection device in this embodiment of the invention is described in detail from the perspective of modular functional entities. The on-site delivery image quality inspection device in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0066] Figure 7This is a schematic diagram of the structure of an on-site delivery image quality inspection device 600 provided in an embodiment of the present invention. The on-site delivery image quality inspection device 600 can vary significantly due to different configurations or performance. It may include one or more central processing units (CPUs) 610 (e.g., one or more processors) and a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage media 630 can be temporary or persistent storage. The program stored in the storage media 630 may include one or more modules (not shown in the diagram), each module may include a series of instruction operations on the on-site delivery image quality inspection device 600. Furthermore, the processor 610 may be configured to communicate with the storage media 630 and execute the series of instruction operations in the storage media 630 on the on-site delivery image quality inspection device 600 to implement the steps of the on-site delivery image quality inspection method provided in the above-described method embodiments.

[0067] The on-site image quality inspection equipment 600 may also include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input / output interfaces 660, and / or one or more operating systems 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 7 The illustrated structure of the on-site delivery image quality inspection equipment does not constitute a limitation on the on-site delivery image quality inspection equipment. It may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0068] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the steps of the on-site signing image quality inspection method.

[0069] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system, device, or unit described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0070] 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, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0071] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for door-to-door image quality inspection, characterized in that, The method comprises the following steps: receiving an on-site signing image sent by a user terminal, and preprocessing the on-site signing image to obtain a preprocessed image; distributing the preprocessed image to multiple quality inspection models according to a preset traffic distribution ratio, performing parallel inference on the preprocessed image by the multiple quality inspection models, and respectively outputting inference results, wherein the inference results comprise compliance judgment and confidence; assigning corresponding model weights to the multiple quality inspection models based on historical accuracy, and performing weighted fusion on the compliance judgment of the multiple quality inspection models based on the model weights and the confidence to generate final quality inspection result information; recording model information, inference time information and image ID information, and sending the model information, the inference time information, the image ID information and the final quality inspection result information to the user terminal; monitoring performance indicators of each quality inspection model, and dynamically adjusting the traffic distribution ratio and model parameters based on the performance indicators.

2. The method of claim 1, wherein, The method comprises the following steps: receiving an on-site signing image sent by a user terminal; performing format conversion on the on-site signing image based on a preset standard format; adjusting the size of the on-site signing image based on a preset resolution, and maintaining the image proportion of the on-site signing image by using a bilinear interpolation algorithm; performing enhancement processing on the on-site signing image by using an image enhancement technique, wherein the image enhancement technique comprises histogram equalization and Gaussian filter denoising; performing normalization processing on the pixel value of the on-site signing image to obtain a preprocessed image.

3. The method of claim 1, wherein, The method comprises the following steps: distributing the preprocessed image to multiple quality inspection models according to a preset traffic distribution ratio, performing parallel inference on the preprocessed image by the multiple quality inspection models, and respectively outputting inference results, wherein the inference results comprise compliance judgment and confidence; distributing the preprocessed image to multiple quality inspection models according to a preset traffic distribution ratio by using a load balancing algorithm, wherein the quality inspection models comprise an image classification model based on a convolutional neural network and a target detection model based on YOLO; performing parallel inference on the preprocessed image by the multiple quality inspection models, and outputting the compliance judgment and the confidence of the entire image by using the image classification model as the inference result, and outputting the compliance judgment and the confidence of a specific region in the image by using the target detection model as the inference result; analyzing the inference result of each quality inspection model, and extracting the compliance judgment and the confidence; 4. The method of claim 1, wherein, storing the compliance judgment and the confidence in a temporary result pool. The method comprises the following steps: querying the historical accuracy of each quality inspection model from a model performance database, and calculating the model weight based on the historical accuracy; performing numerical value processing on the compliance judgment of each quality inspection model; Fusing the compliance judgments after the numericalization using a weighted average algorithm based on the model weights and the confidence levels to generate a fused score; When the fused score is greater than or equal to a preset score threshold, a final quality inspection result information of compliance is generated; When the fused score is less than the score threshold, a final quality inspection result information of non-compliance is generated.

5. The method of claim 1, wherein, The used model information, inference time information and image ID information are recorded, and the used model information, inference time information, image ID information and final quality inspection result information are sent to the user end, including: The used model information, the inference time information, the image ID information and the final quality inspection result information are integrated and packaged as JSON format data to obtain integrated information; The integrated information is sent to the user end through the HTTP protocol, and the confirmation response of the user end is received. The performance indicators of each quality inspection model are monitored, and the traffic distribution ratio and model parameters are dynamically adjusted based on the performance indicators, including:

6. The method of door-to-door image quality inspection according to claim 1, characterized in that, The performance indicators of each quality inspection model are monitored in real time, and the performance indicators include accuracy, response time and resource utilization rate; The comprehensive score of each quality inspection model is calculated based on the performance indicators within a time window, and the traffic distribution ratio is dynamically adjusted according to the comprehensive score; The adjusted traffic distribution ratio is updated to the system configuration center; When the accuracy in the performance indicators of the quality inspection model is less than a preset threshold, the model parameters corresponding to the quality inspection model are adjusted. After the performance indicators of each quality inspection model are monitored and the traffic distribution ratio and model parameters are dynamically adjusted based on the performance indicators, the method further includes:

7. The method of claim 1, wherein, All quality inspection records of the on-site signing images are summarized, and a quality inspection statistical report is generated according to the quality inspection records; The quality inspection statistical report is encrypted to obtain encrypted report information; The encrypted report information is uploaded to the blockchain. Including:

8. An image quality inspection device for door-to-door signature, characterized in that, A receiving pre-processing module is configured to receive an on-site signing image sent by a user end, and to pre-process the on-site signing image to obtain a pre-processed image; A distribution inference output module is configured to distribute the pre-processed image to multiple quality inspection models according to a preset traffic distribution ratio, and to perform parallel inference on the pre-processed image through the multiple quality inspection models to respectively output inference results, the inference results including compliance judgments and confidence levels; A fusion generation module is configured to respectively assign corresponding model weights to the multiple quality inspection models based on historical accuracy rates, and to perform weighted fusion on the compliance judgments of the multiple quality inspection models based on the model weights and the confidence levels to generate final quality inspection result information; ​ The record sending module is configured to record used model information, inference time information and image ID information, and send the used model information, the inference time information, the image ID information and the final quality inspection result information to the user end. The monitoring and adjusting module is configured to monitor performance indexes of each quality inspection model, and dynamically adjust the traffic distribution ratio and model parameters based on the performance indexes.

9. An image quality inspection device for door-to-door signature, characterized by, The door-to-door signature image quality inspection device includes a memory and at least one processor, and the memory stores instructions. The at least one processor invokes the instructions in the memory, so that the door-to-door signature image quality inspection device performs each step of the door-to-door signature image quality inspection method in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions are executed by the processor to implement each step of the door-to-door signature image quality inspection method in any one of claims 1-7.

Images