Image quality inspection method and device, equipment and storage medium

Abstract

The invention relates to the technical field of image detection, and discloses an image quality inspection method and device, equipment and a storage medium. The image quality inspection method comprises the steps of receiving an image to be subjected to quality inspection; performing parallel meta-feature detection on the image to obtain the confidence coefficient of the meta-feature contained in the image; calculating the suspicion degree of the image by adopting a preset weighted fusion strategy based on each confidence coefficient; and based on the suspicion degree, distributing the image to a corresponding quality inspection channel for image quality inspection. According to the method, the image quality inspection safety is ensured, the image quality inspection efficiency is further improved, and the balance of image quality inspection in calculation cost, processing delay and auditing effect quality inspection is ensured.

Description

Technical Field

[0001] This invention relates to the field of image inspection, and more particularly to an image quality inspection method, apparatus, device, and storage medium. Background Technology

[0002] With the rapid development of internet technology, especially the rise of social platforms such as live streaming and short videos, user-generated content (UGC) has experienced explosive growth. While the massive amount of image and video data brings vitality to these platforms, it also presents serious content security challenges. To maintain a healthy online ecosystem and prevent the spread of illegal and harmful information, efficient and accurate content security review systems have become essential infrastructure for major platforms. However, in live streaming scenarios with extremely high real-time requirements, existing content security review systems consistently face an inherent contradiction between computational costs, processing latency, and review effectiveness when handling massive amounts of UGC images or videos.

[0003] To strike a balance between computational cost, processing latency, and review effectiveness, some existing technologies attempt to use a general, small image classifier for initial "normal / abnormal" binary classification. However, this method still suffers from insurmountable drawbacks. First, its target definition is vague. Training a general classifier capable of accurately identifying all anomalous content is extremely difficult because the scope of anomalous content is too broad and its boundaries are blurred, making it difficult to construct a high-quality training dataset. Second, the model's generalization ability is poor. It may be effective for some known violations, but its recognition ability drops significantly when faced with new types of black market images that cleverly embed violating elements into seemingly normal backgrounds (such as landscapes or people). The fundamental reason is that the model attempts to make a black-and-white final judgment rather than a preliminary assessment of the degree of suspicion, resulting in a lack of sensitivity to new and covert violations. Therefore, a new technical solution that can achieve a better balance between computational cost, processing speed, and review accuracy is urgently needed. Summary of the Invention

[0004] The main objective of this invention is to provide an image quality inspection method, apparatus, device, and storage medium, aiming to solve the technical problem that existing image quality inspection technologies have difficulty reconciling computational costs, processing delays, and review effectiveness.

[0005] The first aspect of the present invention provides an image quality inspection method, the image quality inspection method comprising:

[0006] Receive images to be inspected;

[0007] Parallel meta-feature detection is performed on the image to obtain the confidence level that the image contains meta-features;

[0008] Based on the confidence levels, a pre-set weighted fusion strategy is used to calculate the suspicion level of the image;

[0009] Based on the level of suspicion, the image is distributed to the corresponding quality inspection channel for image quality inspection.

[0010] In a first implementation of the first aspect of the present invention, the step of performing parallel meta-feature detection on the image to obtain the confidence that the image contains meta-features includes:

[0011] The image is input into a pre-trained multi-label classification model;

[0012] The image is subjected to parallel detection of multiple meta-features using the multi-label classification model, and the confidence score of the image containing multiple predefined meta-features is output.

[0013] In a second implementation of the first aspect of the present invention, the meta-features include text region density and regularity, QR code / barcode-like patterns, high information entropy local regions, unnatural graphic overlays, and color and saturation anomalies.

[0014] In a third implementation of the first aspect of the present invention, the multi-label classification model includes a backbone network and multiple meta-feature detection heads, each meta-feature detection head corresponding to a meta-feature; the meta-feature detection head is used to detect the probability that the image contains the corresponding meta-feature.

[0015] In a fourth implementation of the first aspect of the present invention, the step of calculating the suspicion level of the image based on each of the confidence levels using a pre-set weighted fusion strategy includes:

[0016] Based on a pre-set weighted fusion strategy, the weight coefficients corresponding to each meta-feature are determined, wherein the magnitude of the weight coefficients is related to the business scenario in which the image is located;

[0017] The suspicion level of the image is obtained by weighted fusion calculation of the weight coefficients and confidence levels corresponding to each meta-feature.

[0018] In a fifth implementation of the first aspect of the present invention, the quality inspection channel includes multiple risk level image quality inspection channels, each risk level corresponds to a quality inspection threshold, and each image quality inspection channel corresponds to several image quality inspection methods and / or several image quality inspection models.

[0019] In a sixth implementation of the first aspect of the present invention, the step of distributing the image to a corresponding quality inspection channel for image quality inspection based on the suspicion level includes:

[0020] Determine the magnitude of the suspicion level and the corresponding quality inspection decision threshold for each risk level;

[0021] Based on the suspicion level and the quality inspection decision threshold corresponding to each risk level, the target image quality inspection channel for the image to be distributed is determined.

[0022] The image is distributed to the target image quality inspection channel, and image quality inspection is performed using the image quality inspection method and / or image quality inspection model corresponding to the target image quality inspection channel.

[0023] A second aspect of the present invention provides an image quality inspection device, the image quality inspection device comprising:

[0024] Image receiving module, used to receive images to be inspected;

[0025] The feature detection module is used to perform parallel meta-feature detection on the image to obtain the confidence level that the image contains meta-features;

[0026] The confidence calculation module is used to calculate the suspicion level of the image based on each confidence level and using a preset weighted fusion strategy.

[0027] The image quality inspection module is used to distribute the image to the corresponding quality inspection channel for image quality inspection based on the suspicion level.

[0028] In a first implementation of the second aspect of the present invention, the feature detection module is specifically used for:

[0029] The image is input into a pre-trained multi-label classification model;

[0030] The image is subjected to parallel detection of multiple meta-features using the multi-label classification model, and the confidence score of the image containing multiple predefined meta-features is output.

[0031] In a second implementation of the first aspect of the present invention, the meta-features include text region density and regularity, QR code / barcode-like patterns, high information entropy local regions, unnatural graphic overlays, and color and saturation anomalies.

[0032] In a third implementation of the second aspect of the present invention, the multi-label classification model includes a backbone network and multiple meta-feature detection heads, each meta-feature detection head corresponding to a meta-feature; the meta-feature detection head is used to detect the probability that the image contains the corresponding meta-feature.

[0033] In a fourth implementation of the second aspect of the present invention, the confidence calculation module is specifically used for:

[0034] Based on a pre-set weighted fusion strategy, the weight coefficients corresponding to each meta-feature are determined, wherein the magnitude of the weight coefficients is related to the business scenario in which the image is located;

[0035] The suspicion level of the image is obtained by weighted fusion calculation of the weight coefficients and confidence levels corresponding to each meta-feature.

[0036] In a fifth implementation of the second aspect of the present invention, the quality inspection channel includes multiple risk level image quality inspection channels, each risk level corresponds to a quality inspection threshold, and each image quality inspection channel corresponds to several image quality inspection methods and / or several image quality inspection models.

[0037] In a sixth implementation of the second aspect of the present invention, the image quality inspection module is specifically used for:

[0038] Determine the magnitude of the suspicion level and the corresponding quality inspection decision threshold for each risk level;

[0039] Based on the suspicion level and the quality inspection decision threshold corresponding to each risk level, the target image quality inspection channel for the image to be distributed is determined.

[0040] The image is distributed to the target image quality inspection channel, and image quality inspection is performed using the image quality inspection method and / or image quality inspection model corresponding to the target image quality inspection channel.

[0041] A third aspect of the present invention provides a computer device, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the computer device to perform the image quality inspection method described above.

[0042] A fourth aspect of the present invention provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the image quality inspection method described above.

[0043] The image quality inspection method provided by this invention focuses on achieving intelligent image pre-screening. By constructing a funnel-shaped multi-level processing architecture, it enables efficient and accurate segmentation of massive images. First, high throughput and low latency are achieved through parallelized, lightweight meta-feature extraction. When the system receives an image to be inspected, it does not directly send it to a computationally intensive complex model, but instead first enters a lightweight pre-screening module. The core of this module is the parallel detection of multiple meta-features in the image. Meta-feature detection has extremely low computational overhead and can be executed in parallel, allowing the system to complete a multi-dimensional preliminary evaluation of an image within milliseconds. Each detector does not output a hard "yes / no" judgment, but rather outputs a confidence level of a meta-feature. This parallel processing architecture is the foundation for the system's high throughput and low latency response, ensuring that even when faced with a massive influx of image data, the system can perform rapid and unblocked preliminary screening at the entry point. Second, high robustness is achieved through comprehensive risk assessment based on a weighted fusion strategy. After obtaining multiple independent meta-feature confidence scores, the system enters the core stage of intelligent decision-making, specifically employing a pre-set weighted fusion strategy to integrate these dispersed confidence scores into a unified, quantified suspicion score. This weighted fusion strategy is not a simple weighting; the weight coefficients used reflect the differences in importance of different meta-features in different application scenarios. This multi-dimensional cross-validation mechanism can more accurately identify novel variant attacks that cleverly embed illegal elements into normal backgrounds, greatly improving the system's robustness. Finally, funnel-shaped intelligent distribution based on suspicion score optimizes computational resources. The calculated suspicion score becomes the sole basis for subsequent resource scheduling. The system presets multiple quality inspection decision thresholds, thereby intelligently distributing images to different quality inspection channels, forming an efficient funnel structure.

[0044] This invention can quickly and accurately identify and filter out the vast majority of benign, low-risk images with extremely low computational overhead, while precisely directing limited and expensive computing resources to at least a few highly suspicious images for in-depth and complex analysis. Through this "funnel-shaped" resource allocation strategy, the cost, efficiency, and security of the entire content review system are optimized at the macro level, effectively coping with the onslaught of massive amounts of data and meeting the business requirement of second-level response to high-risk content. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of one embodiment of the image quality inspection method in this invention;

[0046] Figure 2 This is a schematic diagram of one embodiment of the image quality inspection device in this invention;

[0047] Figure 3 This is a schematic diagram of one embodiment of the computer device in this invention. Detailed Implementation

[0048] 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.

[0049] 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 image quality inspection method in this invention includes:

[0050] 101. Receive the image to be inspected;

[0051] In this embodiment, the image to be inspected is uploaded by the user. To prevent users from maliciously uploading non-compliant images, quality inspection is required after the user uploads the image. The upstream business system (such as the user upload platform or content management system) sends the image data via an HTTP POST request. The request body may contain the image itself (Base64 encoded or binary stream) and the image's metadata (such as ID, source, upload time, etc.). To ensure high availability and high performance, a reverse proxy such as Nginx is typically used for load balancing, and the backend service is deployed using a microservice architecture.

[0052] For high-concurrency, high-data-volume applications (such as live streaming), asynchronous message queues (such as Kafka, RabbitMQ, and RocketMQ) are preferred. The upstream system sends image information as messages to a designated Topic. The image quality inspection service, acting as a consumer, pulls messages from the queue for processing. Furthermore, after receiving the data, preprocessing such as format validation, image decoding, color space conversion, and size normalization can be performed to obtain a standardized image suitable for subsequent analysis.

[0053] 102. Perform parallel meta-feature detection on the image to obtain the confidence level that the image contains meta-features;

[0054] In this embodiment, to address the problems of ambiguous classification target definitions and poor generalization ability in existing quality inspection models, meta-features are introduced. Meta-features refer to general visual patterns that are independent of the semantic content of the image and highly correlated with the information implantation behavior.

[0055] In an optional embodiment, the meta-features preferably include:

[0056] (1) Text region density and regularity: refers to the density of text arrangement in an image and its degree of harmony with the background. Malicious images often contain text blocks that are not naturally arranged, have high density, or are inconsistent with the background. This meta-feature aims to detect whether there are regions in the image that resemble text textures or stroke structures.

[0057] (2) QR code / barcode-like pattern: refers to the presence of content resembling a QR code or barcode in an image. This meta-feature aims to detect the presence of high-frequency, square visual patterns with location point features in an image.

[0058] (3) High information entropy local regions: These refer to local regions in an image that contain high information content. Attackers often paste small pieces of high-information-content illegal content (such as non-compliant images, various signs, etc.) onto a smooth background in an image. This meta-feature aims to detect whether there are isolated regions in an image that have abnormally high information entropy (or texture complexity) compared to the surrounding areas.

[0059] (4) Unnatural graphic overlay: This refers to the presence of unnatural graphic overlays in an image. Malicious images often use post-processing techniques such as arrows, circles, color blocks, blurring, and mosaics to emphasize or obscure information in the image. This meta-feature aims to detect whether there are graphic elements with obvious signs of manual editing in an image.

[0060] (5) Color and Saturation Anomalies: This refers to anomalies in the specific color distribution and saturation of an image. Some types of malicious images have specific preferences in color distribution or attract user attention by adjusting saturation. This meta-feature aims to detect whether there are anomalies in the global or local color distribution of an image.

[0061] In an optional embodiment, step 102 above further includes:

[0062] 1021. Input the image into a pre-trained multi-label classification model;

[0063] 1022. The image is subjected to parallel detection of multiple meta-features using the multi-label classification model, and the confidence score of the image containing multiple predefined meta-features is output.

[0064] In this optional embodiment, the multi-label classification model is a lightweight multi-head meta-feature extraction network, preferably a depth-optimized ultra-lightweight convolutional neural network. This convolutional neural network is not a standard classification network, but rather a multi-task output network.

[0065] In one embodiment, the multi-label classification model includes a backbone network and multiple meta-feature detection heads, each corresponding to a meta-feature. The meta-feature detection heads are used to detect the probability that an image contains the corresponding meta-feature. For example, the backbone network of the convolutional neural network adopts the MobileNetV3 architecture, but instead of a single classification head, multiple parallel meta-feature detection heads are connected to the end of the convolutional neural network. Inputting the image to be inspected into the multi-label classification model allows for the real-time parallel output of confidence scores for multiple predefined meta-features.

[0066] For example, suppose the meta-feature detection heads used in a multi-label classification model include: text region density and regularity detection head, QR code / barcode pattern detection head, high information entropy local region detection head, unnatural graphic overlay detection head, and color and saturation anomaly detection head. When an image is input into this multi-label classification model for meta-feature detection, the model will simultaneously output the confidence scores of each meta-feature, i.e., the probability that the image contains each meta-feature. For example, the confidence score for text region density and regularity might be 0.3, the confidence score for QR code / barcode pattern might be 0.7, the confidence score for high information entropy local regions might be 0.9, the confidence score for unnatural graphic overlay might be 0.1, and the confidence score for color and saturation anomaly might be 0.6.

[0067] This embodiment introduces meta-feature detection to more quickly and in parallel assess potential non-compliance defects in images from multiple dimensions, quantifying the severity of each non-compliance defect into a confidence score. In the image pre-screening stage, this embodiment does not focus on the specific content of the violation (e.g., whether it's gambling or pornography), but rather on identifying the more fundamental, structural "information-carrying patterns" that are often associated with malicious content—that is, meta-feature detection. Furthermore, the model in this embodiment focuses on the more fundamental and harder-to-forge "information-carrying patterns" (i.e., meta-features) rather than the easily changeable specific content, thus ensuring high recall and robustness for highly suspicious images.

[0068] 103. Based on the aforementioned confidence levels, a pre-set weighted fusion strategy is used to calculate the suspicion level of the image;

[0069] In this embodiment, multi-dimensional evaluation of the images to be inspected is achieved through meta-feature detection. To facilitate accurate image screening, a weighted fusion strategy is further introduced to integrate the multi-dimensional evaluation results into a single, decision-oriented comprehensive indicator: suspicion level. This suspicion level specifically measures the degree of risk of an image violating regulations. This embodiment calculates the final suspicion level of an image through a configurable weighted fusion strategy.

[0070] In an optional embodiment, step 103 above further includes:

[0071] 1031. Based on a pre-set weighted fusion strategy, determine the weight coefficients corresponding to each of the meta-features, wherein the magnitude of the weight coefficients is related to the business scenario in which the image is located;

[0072] 1032. Perform weighted fusion calculation on the weight coefficients and confidence scores corresponding to each of the meta-features to obtain the suspicion level of the image.

[0073] For an image to be inspected, after pre-screening, a set of confidence scores for meta-features is obtained, such as {Score_1, Score_2, Score_3, Score_4,...}. Each score represents the confidence score of a meta-feature. In this optional embodiment, the weighted fusion strategy preferably adopts linear weighted summation, the mathematical formula of which is:

[0074] S = Σ(wi*Score_i)

[0075] Where S represents the suspicion level of the image (with a value between 0 and 1), Score_i represents the confidence level of the i-th meta-feature, and wi represents the weight coefficient of the i-th meta-feature.

[0076] In this optional embodiment, the weighting coefficient is the core of the strategy, and its setting directly affects the final calculation result of the suspicion level. The magnitude of the weighting coefficient is preferably related to the business scenario in which the image exists; that is, the weighting coefficient can be dynamically adjusted. For example, in a comment section scenario, promotional text is more common, so the weighting coefficient of text meta-features can be increased. In a profile picture upload scenario, QR code advertising carries a higher risk, so the weighting coefficient of QR code meta-features can be increased. Furthermore, the weighting coefficient can also be periodically and automatically optimized using feedback results from the deep analysis channel (i.e., which meta-feature combinations of images are ultimately identified as malicious).

[0077] 104. Based on the level of suspicion, the image is distributed to the corresponding quality inspection channel for image quality inspection.

[0078] In this embodiment, the magnitude of the suspicion value corresponds to different risk levels. Therefore, this embodiment pre-sets multiple different image quality inspection channels, each of which corresponds to different processing logic and resources.

[0079] In one embodiment, the quality inspection channel preferably includes image quality inspection channels of multiple risk levels. Each risk level corresponds to a quality inspection threshold, and each image quality inspection channel corresponds to several image quality inspection methods and / or several image quality inspection models. Among them, the image quality inspection methods can be filtering based on simple rules, duplicate filtering based on image hash values, etc. The image quality inspection models can be image recognition models for specific categories, such as pornographic recognition models, violent and terrorist recognition models, text recognition and analysis models, face recognition models, etc. To further improve the review accuracy of image quality inspection, it is preferred to deploy multiple image quality inspection methods and / or multiple image quality inspection models in the same image quality inspection channel, and only when all the image quality inspection methods or all the image quality inspection models in the same image quality inspection channel pass, it is determined that the image passes the quality inspection.

[0080] In an alternative embodiment, step 104 above further includes:

[0081] 1041. Respectively judge the magnitudes of the suspicion degree and the quality inspection decision thresholds corresponding to each risk level;

[0082] 1042. Based on the magnitudes of the suspicion degree and the quality inspection decision thresholds corresponding to each risk level, determine the target image quality inspection channel for the image to be distributed;

[0083] 1043. Distribute the image to the target image quality inspection channel, and perform image quality inspection through the image quality inspection methods and / or image quality inspection models corresponding to the target image quality inspection channel.

[0084] In this alternative embodiment, to achieve the balance of computing cost, processing speed and review accuracy, a dynamic threshold method is adopted for decision-making. Specifically: compare the calculated suspicion degree score S with the quality inspection decision threshold T, and according to the comparison result, determine the quality inspection channel for image distribution.

[0085] For example, if S < T, it is determined as low risk and goes through the fast quality inspection channel, which can improve the quality inspection processing speed while ensuring the quality inspection safety and review accuracy. If S >= T, it is determined as high suspicion and goes through the in-depth analysis quality inspection channel.

[0086] There is no limit to the dynamic setting method of the quality inspection decision threshold in this alternative embodiment. For example: when the system load is too high (such as the system load exceeds 60%), T can be appropriately increased (such as increasing by 20% based on the base value T0), sacrificing a certain recall rate to ensure the stability of the core service. When performing quality inspection for high-risk services, T can be reduced (such as reducing by 20% based on the base value T0) to intercept all suspicious images with the strictest standard to ensure absolute safety.

[0087] In addition, by setting multiple thresholds (T1 < T2 < T3), more refined hierarchical processing can be achieved, and the image quality inspection channels can be divided into multiple levels such as "safe", "low suspicion", and "high suspicion". Different levels correspond to different image quality inspection strategies. For example, S < 0.2: Distribute the image to the safe channel for quality inspection; 0.2 <= S < 0.7: Distribute it to the low suspicion channel for quality inspection; S >= 0.7: Distribute the image to the high suspicion channel for quality inspection.

[0088] In this embodiment, through meta-feature detection, a multi-dimensional fusion-based suspicion degree evaluation is realized, so as to quickly filter a large number of benign and compliant images, greatly improving the processing efficiency of image quality inspection and effectively solving the real-time problem under the impact of a large amount of image data. At the same time, through meta-feature detection to complete the pre-screening of images, the precision of computing resource utilization is realized, ensuring that valuable computing power is concentrated on analyzing a small number of pictures that are truly suspected of having risks. With the same hardware investment, the detection ability and response speed of the system for high-risk content in images are greatly enhanced. In addition, since the image pre-screening focuses on the more underlying and more difficult-to-forge "information-bearing mode" meta-features rather than the specific content that is easy to change, the recall rate of high-suspicion pictures is guaranteed. At the same time, the strategy design of weighted fusion and dynamic threshold makes it more convenient for the system to be adjusted according to different business scenarios and operation requirements, achieving a dynamic balance of cost, efficiency, and security.

[0089] The image quality inspection method in the embodiment of the present invention has been described above. Next, the image quality inspection device in the embodiment of the present invention will be described. Please refer to Figure 2 , an embodiment of the image quality inspection device in the embodiment of the present invention includes:

[0090] An image receiving module 201, configured to receive an image to be quality inspected;

[0091] A feature detection module 202, configured to perform parallel meta-feature detection on the image to obtain the confidence level of the meta-features included in the image;

[0092] A confidence level calculation module 203, configured to calculate the suspicion degree of the image based on each of the confidence levels by using a preset weighted fusion strategy;

[0093] An image quality inspection module 204, configured to distribute the image to the corresponding quality inspection channel for image quality inspection based on the suspicion degree. <00,00197>In an optional embodiment, the feature detection module 202 is specifically configured to:

[0095] Input the image into a pre-trained multi-label classification model;

[0096] The image is subjected to parallel detection of multiple meta-features using the multi-label classification model, and the confidence score of the image containing multiple predefined meta-features is output.

[0097] In one alternative embodiment, the meta-features include text region density and regularity, QR code / barcode-like patterns, high information entropy local regions, unnatural graphic overlays, and color and saturation anomalies.

[0098] In one optional embodiment, the multi-label classification model includes a backbone network and multiple meta-feature detection heads, each meta-feature detection head corresponding to a meta-feature; the meta-feature detection head is used to detect the probability that the image contains the corresponding meta-feature.

[0099] In an optional embodiment, the confidence calculation module 203 is specifically used for:

[0100] Based on a pre-set weighted fusion strategy, the weight coefficients corresponding to each meta-feature are determined, wherein the magnitude of the weight coefficients is related to the business scenario in which the image is located;

[0101] The suspicion level of the image is obtained by weighted fusion calculation of the weight coefficients and confidence levels corresponding to each meta-feature.

[0102] In one optional embodiment, the quality inspection channel includes multiple risk level image quality inspection channels, each risk level corresponds to a quality inspection threshold, and each image quality inspection channel corresponds to several image quality inspection methods and / or several image quality inspection models.

[0103] In an optional embodiment, the image quality inspection module 204 is specifically used for:

[0104] Determine the magnitude of the suspicion level and the corresponding quality inspection decision threshold for each risk level;

[0105] Based on the suspicion level and the quality inspection decision threshold corresponding to each risk level, the target image quality inspection channel for the image to be distributed is determined.

[0106] The image is distributed to the target image quality inspection channel, and image quality inspection is performed using the image quality inspection method and / or image quality inspection model corresponding to the target image quality inspection channel.

[0107] Since the embodiments of the device part correspond to the embodiments of the above method, the description of the image quality inspection device provided by the present invention should refer to the above method embodiments. The present invention will not be described again here, but it has the same beneficial effects as the above image quality inspection method.

[0108] above Figure 2The image quality inspection device in this embodiment of the invention will be described in detail from the perspective of modular functional entities. The computer equipment in this embodiment of the invention will be described in detail from the perspective of hardware processing.

[0109] Figure 3 This is a schematic diagram of a computer device 500 provided in an embodiment of the present invention. The computer device 500 can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 510 (e.g., one or more processors) and a memory 520, and one or more storage media 530 (e.g., one or more mass storage devices) for storing application programs 533 or data 532. The memory 520 and storage media 530 can be temporary or persistent storage. The program stored in the storage media 530 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the computer device 500. Furthermore, the processor 510 may be configured to communicate with the storage media 530 and execute the series of instruction operations in the storage media 530 on the computer device 500.

[0110] Computer device 500 may also include one or more power supplies 540, one or more wired or wireless network interfaces 550, one or more input / output interfaces 560, and / or one or more operating systems 531, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art will understand that... Figure 3 The computer device structure shown does not constitute a limitation on the computer device and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0111] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing computer-readable instructions, which, when executed by the processor, cause the processor to perform the steps of the image quality inspection method described in the above embodiments.

[0112] The present invention also provides a computer-readable storage medium, which may 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 image quality inspection method.

[0113] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0114] 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.

[0115] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image quality inspection method, characterized in that, The image quality inspection method includes: Receive images to be inspected; Parallel meta-feature detection is performed on the image to obtain the confidence level that the image contains meta-features; Based on the confidence levels, a pre-set weighted fusion strategy is used to calculate the suspicion level of the image; Based on the level of suspicion, the image is distributed to the corresponding quality inspection channel for image quality inspection.

2. The image quality inspection method according to claim 1, characterized in that, The parallel meta-feature detection of the image to obtain the confidence level that the image contains meta-features includes: The image is input into a pre-trained multi-label classification model; The image is subjected to parallel detection of multiple meta-features using the multi-label classification model, and the confidence score of the image containing multiple predefined meta-features is output.

3. The image quality inspection method according to claim 2, characterized in that, The meta-features include text region density and regularity, QR code / barcode-like patterns, high information entropy local regions, unnatural graphic overlays, and color and saturation anomalies.

4. The image quality inspection method according to claim 2, characterized in that, The multi-label classification model includes a backbone network and multiple meta-feature detection heads, each corresponding to a meta-feature; the meta-feature detection head is used to detect the probability that the image contains the corresponding meta-feature.

5. The image quality inspection method according to any one of claims 1-4, characterized in that, The calculation of the suspicion level of the image based on each confidence level and using a pre-set weighted fusion strategy includes: Based on a pre-set weighted fusion strategy, the weight coefficients corresponding to each meta-feature are determined, wherein the magnitude of the weight coefficients is related to the business scenario in which the image is located; The suspicion level of the image is obtained by weighted fusion calculation of the weight coefficients and confidence levels corresponding to each meta-feature.

6. The image quality inspection method according to any one of claims 1-4, characterized in that, The quality inspection channel includes multiple risk levels of image quality inspection channels, each risk level corresponds to a quality inspection threshold, and each image quality inspection channel corresponds to several image quality inspection methods and / or several image quality inspection models.

7. The image quality inspection method according to claim 6, characterized in that, The step of distributing the image to the corresponding quality inspection channel for image quality inspection based on the suspicion level includes: Determine the magnitude of the suspicion level and the corresponding quality inspection decision threshold for each risk level; Based on the suspicion level and the quality inspection decision threshold corresponding to each risk level, the target image quality inspection channel for the image to be distributed is determined. The image is distributed to the target image quality inspection channel, and image quality inspection is performed using the image quality inspection method and / or image quality inspection model corresponding to the target image quality inspection channel.

8. An image quality inspection device, characterized in that, The image quality inspection device includes: Image receiving module, used to receive images to be inspected; The feature detection module is used to perform parallel meta-feature detection on the image to obtain the confidence level that the image contains meta-features; The confidence calculation module is used to calculate the suspicion level of the image based on each confidence level and using a preset weighted fusion strategy. The image quality inspection module is used to distribute the image to the corresponding quality inspection channel for image quality inspection based on the suspicion level.

9. A computer device, characterized in that, The computer device includes: a memory and at least one processor, wherein the memory stores instructions; The at least one processor invokes the instructions in the memory to cause the computer device to perform the image quality inspection method as described in any one of claims 1-7.

10. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the image quality inspection method as described in any one of claims 1-7.

Images