An AFT and ADC joint detection method and device for mass image processing and a storage medium

Abstract

The application discloses an AFT and ADC combined detection method and device for mass image processing and a storage medium. The method comprises the following steps: performing first detection on a to-be-detected image by an AFT device to obtain a first detection result; transmitting the to-be-detected image and the corresponding first detection result stored in a local memory to dispatch software in a memory-to-memory mode by the AFT device; receiving the to-be-detected image from a plurality of AFT devices by the dispatch software, and dispatching each to-be-detected image to at least one ADC server according to a preset binding relationship and a priority scheduling strategy; writing the to-be-detected image into a virtual memory disk by the ADC server, performing second detection on the to-be-detected image by a second detection algorithm, and obtaining a second detection result; fusing the first detection result and the second detection result by the dispatch software to obtain a third detection result, and transmitting the third detection result to an MES system.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and in particular to a method, apparatus and storage medium for joint detection of AFT and ADC for large-capacity image processing. Background Technology

[0002] In the production process of display panels, it is usually necessary to inspect the panels for appearance defects to ensure product performance and quality. Among these, AFT (Array Final Test) equipment is widely used to inspect the performance and defects of display panels, such as checking multiple indicators like color, brightness, and contrast. In practical applications, AFT equipment is often configured with multiple inspection stations. Each station performs automated optical inspection (AOI) and non-uniformity inspection (MURA) through relevant algorithm software, and the inspection results are used for production control and quality evaluation.

[0003] With the development of display technology, high-end display panels such as OLED and LCD are constantly improving in terms of resolution, size, and pixel precision, resulting in a significant increase in the data volume of a single inspection image, and the data volume is changing at an accelerating pace. Faced with the simultaneous requirements of high resolution and high production line speed, existing AFT equipment may experience increased inspection time and intensified data processing pressure when processing large volumes of images in real time. At the same time, due to differences in the software algorithm capabilities of different equipment manufacturers, some equipment has insufficient defect detection capabilities, easily leading to missed detections, over-detection, or inaccurate identification of defect types, thus affecting production quality and yield.

[0004] To address the aforementioned issues, the industry has proposed several measures to improve detection capabilities. For example, negotiating with equipment suppliers to upgrade detection software, improving algorithms or introducing artificial intelligence technology to enhance judgment capabilities. While this approach is convenient to implement, for tasks involving the identification of complex defects or large images, relying solely on the computing power and processing infrastructure of existing equipment often falls short of expectations. It may still result in insufficient detection efficiency or limited reliability of results, failing to fully meet the high-performance detection needs of actual production. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method, apparatus, and storage medium for joint detection of AFT and ADC for large-capacity image processing.

[0006] The technical solution provided in this application is described below:

[0007] A first aspect of this application provides a joint AFT and ADC detection method for large-capacity image processing, the method comprising:

[0008] The first detection is performed on the image to be detected using the AFT device to obtain the first detection result;

[0009] The AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software in a memory-to-memory manner.

[0010] The dispatch software receives images to be detected from multiple AFT devices and dispatches each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy.

[0011] The ADC server writes the image to be detected to a virtual memory disk, and performs a second detection on the image to be detected using a second detection algorithm to obtain a second detection result.

[0012] The first detection result and the second detection result are fused together by the dispatch software to obtain a third detection result, and the third detection result is transmitted to the MES system.

[0013] Optionally, transferring the image to be detected and the corresponding first detection result stored in local memory to the dispatch software via the AFT device in a memory-to-memory manner includes:

[0014] The AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the target shared memory buffer.

[0015] The AFT device negotiates with the dispatch software to adopt a zero-copy memory mapping transmission mechanism;

[0016] Based on the zero-copy memory mapping transfer mechanism, a mapping relationship is constructed between the first memory page and the second memory page in the target shared memory buffer;

[0017] Based on the mapping relationship, the image to be detected and the corresponding first detection result in the memory buffer are transmitted to the dispatch software.

[0018] Optionally, the process of constructing the mapping relationship between the first memory page and the second memory page in the target shared memory buffer based on the zero-copy memory mapping transfer mechanism includes:

[0019] The dispatch software allocates a shared memory identifier and sends the shared memory identifier and the corresponding target page table address to the AFT device.

[0020] The kernel interface is invoked through the AFT device according to the shared memory identifier, and a lock is requested for the first memory page used for transmission;

[0021] In the AFT device, the physical address of the first memory page is registered in the page table of the transmission channel;

[0022] The dispatch software creates a second memory page in the local process address space and maps the virtual address of the second memory page to the corresponding physical address through the page table.

[0023] Optionally, the step of assigning each of the images to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy includes:

[0024] The assignment software establishes a binding table between the AFT device and the ADC server;

[0025] A priority is determined for each arriving image to be detected by the dispatch software, and the image is then added to a priority queue according to the priority.

[0026] Based on the real-time load of the ADC server, the queue length of the priority queue, and the response time, a scheduling score is calculated, and the target ADC server is determined based on the scheduling score.

[0027] Optionally, writing the image to be detected to the virtual memory disk via the ADC server includes:

[0028] Construct a specified capacity region in the physical disk group of the ADC server;

[0029] A temporary virtual disk is created based on the specified capacity area, and the temporary virtual disk is mapped to an independent logical drive letter;

[0030] The image to be detected, passed in by the dispatch software, is written to the temporary virtual disk, and the memory pages occupied are locked during the writing process.

[0031] The space occupied by the temporary virtual disk is monitored. When the space occupied reaches the preset write-back threshold, a write-back operation is performed. The target disk group is determined in turn among multiple physical disk groups through a roulette-style selection strategy.

[0032] The images to be detected in the temporary virtual disk are scattered and written to the target disk group.

[0033] Optionally, the step of fusing the first detection result and the second detection result through the dispatch software to obtain the third detection result includes:

[0034] Defect candidate sets are extracted from the first detection result and the second detection result respectively. Each candidate in the defect candidate set includes defect location, defect type, confidence level, piece number and timestamp.

[0035] In the set of defect candidates, alignment is performed based on the piece number, timestamp, and defect location to obtain paired candidate groups and unilateral candidate groups;

[0036] For the paired candidate groups, the confidence levels of the first detection result and the second detection result are mapped to a unified confidence scale.

[0037] For each pair of candidate pairs, a fusion weight is determined, which is determined by at least one of the following:

[0038] The model's historical accuracy calibration coefficients are obtained based on statistical analysis of the processing records of the dispatch software.

[0039] The credibility verification value of the current model version;

[0040] The defect type weight of the paired candidate group;

[0041] The importance weight of the region where the paired candidate group is located;

[0042] The fusion parameters are calculated based on the fusion weights and the unified confidence scale.

[0043] The results of the paired candidate group and the unilateral candidate are classified according to the fusion parameters and the preset fusion threshold set to obtain a third detection result.

[0044] Optionally, the third detection result is obtained by classifying the paired candidate group and the unilateral candidate group according to the fusion parameters and the preset fusion threshold set.

[0045] If the fusion parameter is greater than a preset first threshold, the corresponding pairing candidate group is marked as "determined defect";

[0046] If the fusion parameter is between a preset first threshold and a second threshold, the corresponding pairing candidate group is marked as "suspected defect";

[0047] If the fusion parameter is lower than the second threshold, the corresponding pairing candidate group is ignored;

[0048] Optionally, the method further includes:

[0049] If the difference in confidence between two candidates in the paired candidate group is greater than a preset difference threshold, the defect type with the higher confidence level is determined as the classification result.

[0050] If the difference in confidence between two candidates in the paired candidate group is lower than a preset difference threshold, the corresponding paired candidate group will be transferred to the backup ADC model for re-execution of the detection.

[0051] A second aspect of this application provides a joint AFT and ADC detection apparatus for large-capacity image processing, the apparatus comprising:

[0052] The first detection unit is used to perform a first detection on the image to be detected using an AFT device to obtain a first detection result;

[0053] The transmission unit is used to transmit the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software via the AFT device in a memory-to-memory manner.

[0054] The dispatch unit is used to receive images to be detected from multiple AFT devices through the dispatch software, and dispatch each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy.

[0055] The second detection unit is used to write the image to be detected to a virtual memory disk through the ADC server, and to perform a second detection on the image to be detected through a second detection algorithm to obtain a second detection result.

[0056] The fusion unit is used to fuse the first detection result and the second detection result through the dispatch software to obtain a third detection result, and then transmit the third detection result to the MES system.

[0057] A third aspect of this application provides a computer-readable storage medium on which a program is stored, which, when executed on a computer, performs the methods of the first aspect and any one of the first aspects.

[0058] As can be seen from the above technical solutions, this application has the following beneficial effects:

[0059] 1. By using the AFT device to transfer the image to be detected and the first detection result to the dispatch software in a memory-to-memory manner, zero-copy data transfer between devices is achieved, avoiding the intermediate buffer and multiple memory copies introduced by traditional file reading and writing or network copying. This significantly reduces CPU utilization and IO waiting time, resulting in a significant reduction in the transmission latency of large-capacity images in multi-device collaborative detection.

[0060] 2. The AFT device and the ADC server perform detection independently. The first detection result and the second detection result are fused by the dispatch software, which can greatly improve the accuracy of the detection result.

[0061] 3. After receiving detection tasks from multiple AFT devices, the dispatch software can perform real-time load distribution on the ADC server based on preset binding relationships and priority scheduling strategies, dynamically adjusting the flow of detection tasks. This strategy not only avoids overload or idleness of a single server, but also improves the resource utilization and parallel processing capabilities of the entire detection system.

[0062] 4. By writing the image to be detected to a virtual memory disk through the ADC server, a high-speed memory-based caching mechanism is implemented, which can temporarily store large amounts of image data without relying on a high-performance physical disk. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a schematic flowchart of an embodiment of the AFT and ADC joint detection method for large-capacity image processing provided in this application;

[0065] Figure 2 This is a schematic flowchart of an embodiment of step S102 in the AFT and ADC joint detection method for large-capacity image processing provided in this application;

[0066] Figure 3 This is a schematic flowchart of an embodiment of step S104 in the AFT and ADC joint detection method for large-capacity image processing provided in this application;

[0067] Figure 4 This is a schematic flowchart of an embodiment of step S105 in the AFT and ADC joint detection method for large-capacity image processing provided in this application;

[0068] Figure 5 A schematic diagram of an embodiment of the AFT and ADC joint detection device for large-capacity image processing provided in this application;

[0069] Figure 6 This is a schematic diagram of an embodiment of another AFT and ADC joint detection device for large-capacity image processing provided in this application. Detailed Implementation

[0070] It should be understood that the method described in this invention can be implemented by software, hardware, or a combination of software and hardware, and the subject performing the method is not limited.

[0071] In practical implementation, the various functional steps described in this invention can be completed collaboratively by multiple system nodes distributed in the same local area network or industrial Ethernet. For example, the AFT detection equipment, the dispatch software server, and the ADC server can be deployed in different industrial control units or computing servers on the production line, or they can be integrated into the same data processing platform.

[0072] Among them, the AFT detection equipment can be a detection terminal with image acquisition and preliminary detection functions, and its built-in detection algorithm module can be executed by a local CPU or GPU; the dispatch software can run on an independent middleware server or main control computing node to realize task scheduling, memory management and result fusion; the ADC server can be a high-performance computing node with deep learning model inference capabilities, supporting deep detection and feature analysis of image data.

[0073] This invention does not limit the specific hardware form of the above-mentioned execution units. Each module can be implemented by physical devices or by software programs, virtual machines, containerized services, etc. running on general computing platforms.

[0074] In some implementations, the functional steps of the above method can be executed sequentially by the same processor, or they can be executed in parallel in a distributed manner in a multiprocessor, multi-node system.

[0075] Furthermore, the program instructions of the method described in this invention can be stored in a computer-readable storage medium, which may include a hard disk, a solid-state drive, flash memory, EPROM, EEPROM, or other non-volatile memory.

[0076] When the program instructions are loaded and executed by the processor, all or part of the functional flow of large-capacity image detection and result fusion described in this invention can be realized.

[0077] Therefore, the embodiments of the present invention are not limited to specific hardware deployments, network topologies, or computing architectures. As long as the corresponding functions of AFT detection, ADC detection, result fusion, and data transmission can be achieved, they all fall within the protection scope of the present invention.

[0078] To facilitate a more accurate understanding of the technical solutions of this invention by those skilled in the art, some terms used in this specification are explained below. Unless otherwise specified, the terms used in this specification should be interpreted according to the common understanding of those skilled in the art upon reading this specification.

[0079] 1. AFT equipment (Automatic Functional Test equipment)

[0080] AFT equipment is an automated functional testing device, mainly used for functional and performance testing during the display panel production process.

[0081] Its main functions include pixel detection, image quality analysis, brightness and color uniformity detection, etc., to quickly identify defects such as dead pixels, color difference, uneven brightness, and MURA in the display panel.

[0082] AFT devices typically achieve automated detection by connecting to the display panel under test and executing a preset detection program, and can output detection data including detection images and initial detection results.

[0083] In this invention, the AFT device plays the role of primary detection and data generation, and is the front-end execution unit of the joint detection system.

[0084] 2. ADC Server (Automatic Defect Classification)

[0085] ADC server is an automated defect detection and classification system used for in-depth defect identification and classification in display panel production lines.

[0086] This system uses high-resolution image acquisition devices and image analysis algorithms to automatically detect and classify defects on the panel surface.

[0087] In this invention, the ADC server undertakes the tasks of high-precision detection and defect classification, and is the core back-end computing unit of the joint detection system.

[0088] 3. Dispatch Middleware

[0089] The dispatch software is a scheduling and control module deployed in network nodes or servers, responsible for task scheduling and data distribution between AFT devices and ADC servers.

[0090] Its main functions include: receiving images to be inspected uploaded by multiple AFT devices, maintaining the binding relationship between AFT and ADC, dynamically allocating inspection tasks according to priority and load, performing memory sharing and zero-copy transmission, and performing result fusion and output after inspection.

[0091] In this invention, the dispatch software is the core control and data scheduling layer of the entire joint detection system.

[0092] 4. MES system (Manufacturing Execution System)

[0093] MES (Manufacturing Execution System) is a real-time information system for manufacturing process management, used to establish an information loop between the production planning layer and the equipment control layer.

[0094] The MES system can receive inspection results in real time, record production status, track the source of defects, and generate quality statistical reports.

[0095] In this invention, the MES system serves as the final receiving and management platform for test results, used to receive the third test results output by the dispatch software, thereby enabling the linkage between test and production information.

[0096] 5. Memory-to-Memory Transfer

[0097] Memory-to-memory transfer is a data transfer mechanism that directly transfers data between devices or processes based on memory mapping, without going through disk files or traditional I / O paths.

[0098] This mechanism, through shared memory or DMA (Direct Memory Access) technology, allows data to be directly copied between the sender's and receiver's memory spaces, reducing CPU intervention and lowering latency.

[0099] In this invention, the mechanism is used for AFT devices to transmit large amounts of detection image data to dispatch software, thereby improving data throughput efficiency.

[0100] 6. Zero-Copy Memory Mapping

[0101] Zero-copy memory mapping is a data transfer optimization technique that uses the operating system kernel to directly map source memory pages to the address space of the target process, achieving "zero-copy" data access.

[0102] In this invention, zero-copy mapping is used for data exchange between the AFT device and the dispatch software, enabling high-speed transmission of the detected image between different processes and avoiding the performance loss caused by traditional file I / O.

[0103] 7. Virtual Memory Disk

[0104] A virtual memory disk is a temporary logical disk that is dynamically built based on system memory or cache, used to cache large amounts of image data.

[0105] In this invention, the ADC server achieves high-speed image writing and reading by creating a virtual memory disk, and supports writing cached content to the physical disk in batches according to the write-back strategy, thus achieving both high-speed temporary storage and persistence.

[0106] 8. TT (Throughput Time)

[0107] TT refers to the total time spent on a single product on a machine from the start of testing to the end of testing. It is an important parameter for measuring the efficiency and capacity of equipment operation.

[0108] The lower the TT value, the higher the processing efficiency of the testing equipment and the stronger the throughput capacity of the production line.

[0109] In the joint detection system of the present invention, the overall TT value can be effectively reduced by optimizing memory-to-memory transfer and task assignment, thereby improving the detection cycle time of the production line.

[0110] 9.RAID 0 (Redundant Array of Independent Disks Level 0)

[0111] RAID 0, also known as striping disk array, is a storage structure that improves read and write speeds by writing data in blocks to multiple hard drives in parallel.

[0112] In this invention, the ADC server can use a physical disk group with a RAID 0 structure when performing virtual disk write-back to further improve the write-back and read performance of batch image data.

[0113] 10. Fusion Processing

[0114] Fusion processing refers to the process of jointly analyzing and weighting the confidence levels of AFT and ADC detection results to generate a unified third detection result.

[0115] This process may include steps such as confidence analysis, weighted averaging, conflict resolution, and backtracking to improve detection accuracy and result consistency.

[0116] Please see Figure 1 This application first provides an embodiment of an AFT (Automatic Detection Technique) and ADC (Automatic Defect Classification) joint detection method, apparatus, and storage medium for large-capacity image processing. This embodiment addresses the problems of insufficient detection capability, low image processing efficiency, and limited multi-device collaboration efficiency in existing display panel inspection processes. The method establishes an efficient data collaboration mechanism between the AFT inspection device and the ADC automatic defect classification system, achieving hierarchical processing and fusion judgment of the detection task, thereby improving detection accuracy and system throughput efficiency. This embodiment includes:

[0117] S101. Perform a first detection on the image to be detected using the AFT device to obtain a first detection result;

[0118] In step S101, the AFT device performs a first detection on the image to be tested to obtain a first detection result. Specifically, during the display of each test pattern, the image acquisition module acquires images displayed on the panel using an industrial camera or a line scan camera at a fixed sampling frequency, and inputs the acquired raw pixel matrix data to the internal image processing unit. This image processing unit may include preprocessing modules for image equalization, illumination correction, gamma adjustment, and noise filtering to eliminate deviations caused by ambient lighting or the angle of capture.

[0119] The detection algorithm module of the AFT device performs rapid feature extraction and regional defect detection on the preprocessed image data.

[0120] In one optional implementation, the detection algorithm can employ a template-matching-based differential detection method: by calculating the difference between the currently acquired image and the standard template image in the pixel grayscale domain, pixel regions with brightness deviations exceeding a set threshold are identified, thereby determining whether a defect exists. For brightness differences in continuous regions, clustering is performed using a connected component algorithm to distinguish between point defects, line defects, and area defects.

[0121] AFT devices can also smooth complex texture areas by combining local mean filtering with dynamic thresholding algorithms, avoiding false detections. During the detection process, corresponding detection index parameters can be output based on different detection modes (such as spot inspection mode, brightness uniformity mode, and chroma mode).

[0122] After completing all inspection items, the AFT device structures and encapsulates the analysis results corresponding to each inspection area to generate the first inspection result data. This first inspection result may include the following information:

[0123] Image identification information and pixel coordinate index of each detection unit;

[0124] Defect categories (such as bright spots, dark spots, line defects, streaks, etc.);

[0125] Defect severity level or confidence score;

[0126] Regional grayscale statistical parameters (average brightness, standard deviation, brightness offset rate, etc.).

[0127] In some embodiments, to ensure detection efficiency, the detection tasks of the AFT device are executed in a multi-threaded manner, with different detection areas processed in parallel by independent task threads. After a detection cycle is completed at the image acquisition end, the detection results are cached in the device's local high-speed memory in the form of key-value pairs. The key-value pairs include the detection image identifier and the corresponding result data structure, so that they can be quickly transferred to the dispatch software in the future via zero-copy memory mapping.

[0128] S102. The image to be detected and the corresponding first detection result stored in the local memory are transmitted to the dispatch software via the AFT device in a memory-to-memory manner.

[0129] In step S102, after completing the first detection, the AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software in a memory-to-memory manner through the dispatch software.

[0130] Specifically, the system's network architecture adopts a unified network segment design, with each AFT device, distribution software host, ADC server, and central storage node all grouped into the same logical subnet to ensure low-latency communication within the local area network. Each device is assigned an independent and contiguous IP address and communicates via a 10 Gigabit Ethernet interface, ensuring sufficient bandwidth and controllable latency for image-level data transmission. The core switch utilizes a 10 Gigabit Ethernet switch with low-latency forwarding capabilities, enabling image data output from the AFT devices to be forwarded to the distribution software host within sub-millisecond latency, achieving real-time processing of large-capacity image data.

[0131] After the detection task is completed, the file transfer module of the AFT device first packages the image to be detected and the structured first detection result into a data package. This packaging process is not a traditional disk transfer method, but rather the data mapping and buffer locking are directly completed through the internal high-speed cache pool. At this time, the image data and detection results are stored in the shared memory area of ​​the AFT device, and the system establishes a zero-copy data channel with the dispatch software through the memory page mapping mechanism.

[0132] In one alternative implementation, data transfer is achieved through Direct Memory Access (DMA) technology. The DMA controller is responsible for directly copying the memory data blocks of the AFT device to the corresponding receive buffer in the host memory of the dispatch software at the hardware level, without the CPU participating in intermediate data transfer. This approach effectively reduces CPU utilization and system context switching overhead, thereby improving data transfer rate and parallel processing capabilities.

[0133] To further ensure transmission stability and smoothness, this technical solution employs a flow control mechanism based on a memory queue. When the receiving thread of the dispatch software is under high load, the AFT device adjusts the data transmission rate based on the returned TCP signals and temporarily stores unfinished transmission tasks in the memory queue. The memory queue is automatically scheduled according to a first-in, first-out (FIFO) strategy, automatically resuming data transmission when the network load recovers. This mechanism not only avoids data loss caused by momentary congestion but also achieves dynamic flow control and optimal allocation of memory resources.

[0134] At the network layer, the transport channel can enable the TLS encryption protocol to ensure the security and integrity of data during transmission. Before data packets are transmitted, a checksum is calculated and a timestamp is appended. Once the dispatch software completes the verification at the receiving end, it immediately releases an acknowledgment signal (ACK) to notify the AFT device that the data packet has been successfully received and written to the buffer. The entire transmission process does not rely on any disk I / O operations and is completed entirely based on the memory path, thereby significantly reducing latency and improving data throughput.

[0135] In a preferred embodiment, the receiving module of the distribution software runs on a server host supporting a 10 Gigabit network, which is configured with a RAID 0 array for auxiliary cache management. Although image data is directly stored in memory without passing through a disk, the system periodically writes some transmission logs, index files, and verification information to the RAID array for subsequent data tracking and anomaly recovery. The read / write speed of the RAID 0 array can be several times that of a single disk, thereby further improving overall I / O capabilities and system redundancy.

[0136] After transmission is complete, the dispatch software caches the received image data and corresponding detection results in a shared memory pool, forming a task queue to await subsequent dispatch and scheduling. The data in the task queue is indexed according to product identifiers (such as film numbers) and AFT device numbers to ensure accurate matching and binding of multi-source image data.

[0137] Through the above technical solution, a high-speed memory-to-memory transfer mechanism based on a 10 Gigabit local area network is realized between the AFT device and the dispatch software. Compared with traditional disk transfer or file copying methods, this solution achieves significant optimizations in transmission latency, CPU utilization, and data consistency. It can support the large-capacity image transmission requirements under multi-station parallel operation conditions and provide low-latency, high-quality input data for subsequent ADC detection.

[0138] See Figure 2 In a specific implementation, step S102 may include:

[0139] S1021. The image to be detected and the corresponding first detection result stored in the local memory are transmitted to the target shared memory buffer through the AFT device.

[0140] After the AFT device completes the first detection and caches the first detection result and the image to be detected in local memory, the AFT device first allocates a target shared memory buffer (hereinafter referred to as the "shared buffer") in local memory for this transmission task, and writes the image data to be transmitted and the corresponding first detection result into the shared buffer according to the agreed data format. The shared buffer can consist of several contiguous or non-contiguous memory pages. When writing, the AFT device simultaneously generates data digest information (e.g., a checksum based on CRC or SHA-256) and metadata (including image number, workstation number, timestamp, data length, page list, etc.), and atomically writes this metadata into the header of the shared buffer or a dedicated metadata area so that the receiver can perform verification and indexing later. To improve concurrency performance, the AFT device can use multi-threading or asynchronous I / O to write the image into blocks into the shared buffer, and record the page table entry and status bit (e.g., "written / pending transmission") of each block after writing, to support subsequent zero-copy mapping and breakpoint resumption.

[0141] S1022. Negotiate with the dispatch software through the AFT device to adopt a zero-copy memory mapping transmission mechanism;

[0142] After the shared buffer is ready, the AFT device and the dispatch software conduct a handshake negotiation through a pre-established control channel (e.g., a TCP-based control session) to determine whether to use zero-copy memory mapping as the data transfer mechanism. During the handshake phase, the AFT device sends a metadata packet for this transmission to the dispatch software (containing a shared buffer identifier, page table information, data digest, priority identifier, and required transmission mode marker). Upon receiving this metadata packet, the dispatch software allocates the corresponding receive buffer for the task in its address space locally and returns the shared buffer identifier and agreement information to the receiving end. Both parties can also negotiate several optional parameters in the handshake message, such as whether to use DMA or RDMA, the maximum number of pages per mapping, backpressure threshold, transmission timeout, and maximum number of retries. If the network or device supports RDMA / hardware direct connection, the handshake can agree to use RDMA to register page tables to reduce CPU paths; if not, it will switch to kernel-level memory mapping (mmap) or other zero-copy implementations. After the handshake is complete and both parties agree on the transmission parameters, the dispatch software returns a "mapping ready" response to AFT, proceeding to the next mapping establishment phase.

[0143] S1023. Based on the zero-copy memory mapping transmission mechanism, construct a mapping relationship between the first memory page and the second memory page in the target shared memory buffer;

[0144] Based on the zero-copy memory mapping mechanism negotiated by both parties during the handshake phase, the system constructs a mapping relationship between the first memory page in the shared buffer and the second memory page at the receiving end. The specific implementation process may include:

[0145] The dispatch software allocates a shared memory identifier and sends the shared memory identifier and the corresponding target page table address to the AFT device. The AFT device calls the kernel interface based on the shared memory identifier and requests to lock a first memory page for transmission. In the AFT device, the physical address of the first memory page is registered in the page table of the transmission channel. The dispatch software creates a second memory page in the local process address space and maps the virtual address of the second memory page to the corresponding physical address using the page table.

[0146] Specifically, the AFT device calls its operating system or driver interface to register the memory pages to be transferred with the kernel / DMA engine (or RDMA registry) and obtains the physical addresses or DMA-visible addresses of these pages. Subsequently, the AFT sends this list of physical addresses to the dispatch software (this transmission can be done via a control channel and signed or encrypted to prevent tampering). Upon receiving the data, the dispatch software locally calls the kernel mapping interface (or RDMA mapping API) to create corresponding virtual address mappings, mapping several memory pages from the receiving end to the physical page addresses provided by the remote end, thus forming a one-to-one page mapping relationship between the two ends. During the mapping process, both sides perform page locking operations (e.g., mlock or driver-level pinning) on ​​the involved memory pages to prevent the operating system from swapping these pages out of the swap space or reclaiming them. After successful mapping, the dispatch software performs cache coherence operations on the mapped pages (e.g., executing cache refresh / invalidation instructions) to ensure that the CPU cache is consistent with the memory pages. To ensure traceability, the system writes the mapping relationship (including page table entries, mapping time, and mapping length) into the task mapping table and returns a "mapping confirmation" message to AFT. If any page registration or mapping fails during the mapping process, the system will trigger a rollback process—releasing the registered page, notifying the peer to reapply or switch to a backup transmission channel (e.g., writing to a sharded disk and completing the process via file transfer), and recording the cause of the exception and error code in the log for subsequent manual intervention or automatic retry.

[0147] S1024. Based on the mapping relationship, the image to be detected and the corresponding first detection result in the memory buffer are transmitted to the dispatch software.

[0148] After the mapping relationship is established and confirmed to be ready, the AFT device can directly set the image to be detected and its first detection result in the shared buffer to a "readable" state based on the mapping relationship. This allows the dispatch software to directly read the data to be transmitted through the mapped virtual address in its local address space without performing a kernel copy. The specific process of data transmission can adopt one of two cooperative modes: push mode or pull mode. In push mode, after the mapping is confirmed, the AFT device sends a data ready signal (including the page range) to the dispatch software to prompt the dispatch software to read the corresponding page from the mapped virtual address. In pull mode, after receiving the mapping confirmation, the dispatch software actively reads the mapped pages in sequence according to the mapping table and completes the write operation locally. Regardless of the mode used, after reading each mapped page, the dispatch software can perform an integrity check on the page data (based on the digest provided with the data in S1021). If the check passes, an ACK is sent to the AFT and the page is marked as "received" in the task queue. If the check fails, the error of the page is reported to the AFT and the page is retransmitted or remapped according to the negotiated retransmission policy.

[0149] Furthermore, to avoid overloading the receive buffer due to instantaneous concurrent reads, the dispatch software can limit the number of pages mapped in parallel based on memory queues and flow control mechanisms (for example, it can be configured to map a maximum of N page blocks concurrently per task, where N is an adjustable parameter), and send a backpressure command to AFT when the queue reaches its limit, requesting it to postpone subsequent data ready notifications.

[0150] The dispatch software locally moves the complete image read through mapping and the first detection result from the receiving buffer to the task processing area (or directly hands the mapping data over to the subsequent processing module), and sends a final transmission completion confirmation message to the AFT device, while simultaneously unlocking and mapping the relevant memory pages.

[0151] S103. Receive images to be detected from multiple AFT devices through the dispatch software, and dispatch each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy.

[0152] In one specific implementation, step S103 involves receiving images to be detected from multiple AFT devices via dispatch software, and dispatching each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy. This process includes: establishing a binding table between the AFT devices and the ADC server using the dispatch software; determining the priority for each image arriving at the dispatch software and merging it into a priority queue based on the priority; calculating a scheduling score based on the real-time load of the ADC server, the queue length of the priority queue, and the response time; and determining the target ADC server based on the scheduling score.

[0153] Specifically, after the dispatch software starts, it first loads the binding relationship table and server status table during the system initialization phase. The binding relationship table describes the logical mapping relationship between each AFT device and one or more ADC servers. Its establishment method can be determined based on factors such as production line layout, network topology, bandwidth allocation, or physical proximity of devices. For example, in a multi-station production line, every three adjacent AFT devices can be bound to the same group of ADC servers to reduce cross-switch traffic and data transmission latency. The binding table may include fields such as AFT device identifier, corresponding ADC server list, default priority, historical average task latency, and binding update timestamp. The system can periodically adjust these fields based on actual load during operation.

[0154] During the task reception phase, the dispatch software receives images to be detected and initial detection results from multiple AFT devices from shared memory or a queue interface. For each received task, the system generates a task description object (TaskDescriptor) with fields including: task ID, AFT source identifier, image size, task priority, arrival timestamp, summary of the initial detection result, and network path identifier. The system then inserts this task object into the scheduling queue.

[0155] To achieve dynamic task scheduling, the dispatch software maintains a priority scheduling queue. This queue can adopt a hierarchical multi-queue structure, with each queue level corresponding to a priority range. For example, task priorities can be divided into five levels: urgent, high, medium, low, and delayed. Priority determination can be based on calculations from multiple dimensions:

[0156] Image types from AFT (e.g., high-resolution image priority processing);

[0157] The detection confidence level of the AFT report (images with low confidence levels need to be re-examined first);

[0158] MES system production line cycle time requirements (adjust task weights based on current production cycle time or batch priority).

[0159] Historical average wait time (tasks that have not been assigned for a long time are automatically given higher priority).

[0160] After priority calculation, the dispatch software places tasks into the corresponding priority queue. The queue can be sorted using timestamps or the earliest due date first (EDF) algorithm based on response time limits to ensure that high-priority tasks are scheduled first.

[0161] Before tasks are dispatched, the dispatch software needs to calculate a scheduling score based on the real-time load status of the ADC servers. Each ADC server can periodically report its current CPU utilization, memory usage, task queue length, virtual disk utilization, and recent average response time, among other status parameters. The dispatch software calculates the load coefficient L based on these status parameters and combines it with the task priority weight W to obtain the scheduling score S. The scoring model can be expressed as:

[0162] S = α·(1-L) + β·W + γ·H; where α, β, and γ are adjustable parameters, and H represents the historical task completion rate or delay penalty factor. The scheduling engine selects the ADC server with the highest score as the task assignment target based on the S value. When multiple server scores are similar, the system can make a decision based on the load balancing strategy, using round-robin, shortest queue first, or least recently used (LRU) strategy.

[0163] The dispatch software writes the image data of the task and the index address of the first detection result into a dispatch instruction packet, which is then sent to the target ADC server via efficient RPC (Remote Procedure Call) or message queue middleware (such as ZeroMQ or gRPC). Before sending, the system can add a task sequence number and integrity check code to the dispatch instruction packet and record the dispatch event in the task log. After receiving the dispatch instruction, the ADC server returns an acknowledgment response, including the reception status, the assigned task ID, and the estimated processing completion time. Upon receiving the acknowledgment response, the dispatch software updates the task status table, marks the task as "dispatched," and synchronously writes the acknowledgment information to the database or cache system for querying by the MES system.

[0164] To improve system stability under high concurrency, the dispatch software can also run a scheduling monitoring thread in the background to dynamically adjust priority thresholds and binding relationships. When it is detected that an ADC server is under high load for an extended period (e.g., CPU utilization exceeds 80% for 30 seconds), the system will automatically remove the server from the active list of the binding table and transfer its bound tasks to a less loaded backup server.

[0165] S104. Write the image to be detected to the virtual memory disk through the ADC server, and perform a second detection on the image to be detected through the second detection algorithm to obtain a second detection result;

[0166] In one specific implementation, step S104 can write the image to be detected to a virtual memory disk through the ADC server, and perform a second detection on the image to be detected through a second detection algorithm to obtain a second detection result.

[0167] After the dispatch software sends task instructions to the ADC server, the ADC server first creates a dedicated high-speed data caching environment on its local server to support the parallel loading and processing of large-capacity images. (See also...) Figure 3 In one embodiment, this step may include the following steps:

[0168] S1041. Construct a specified capacity region in the physical disk group of the ADC server;

[0169] Upon startup, the ADC server constructs a designated capacity region within its physical disk group. This region can consist of multiple solid-state drives (SSDs) or a hybrid storage array, providing physical support space for the virtual memory disk. The size of this designated capacity region can be dynamically allocated based on system load and task image size; for example, it can be expanded to 2–4 times the physical memory capacity in large-scale detection batches to prevent memory write bottlenecks.

[0170] S1042. Create a temporary virtual disk based on the specified capacity area, and map the temporary virtual disk as an independent logical drive letter;

[0171] In this step, the ADC server creates a temporary virtual disk based on the specified capacity area and registers this virtual disk as an independent logical drive letter recognizable by the system. For example, a virtual disk node can be registered at the system kernel level using memory file system technology (such as tmpfs or ramdisk) to achieve high-speed read and write access. At this time, the system will allocate an independent file descriptor and path identifier to the virtual disk, making it directly accessible to the ADC detection process.

[0172] S1043. Write the image to be detected transmitted by the dispatch software into the temporary virtual disk, and lock the memory pages occupied during writing;

[0173] After the virtual disk is initialized, the server receives the task data stream from the dispatch software and writes the image to be detected, passed in by the dispatch software, to the temporary virtual disk. During the writing process, the system uses a memory mapping (mmap) mechanism to directly map the image data to contiguous memory pages of the virtual disk, omitting intermediate buffer copying steps and achieving "zero-copy" writing.

[0174] Meanwhile, to avoid data access delays caused by the operating system's page swapping, the server locks the memory pages it occupies during writing, ensuring that these memory pages remain resident in memory and are not swapped out or overwritten before the detection task is completed.

[0175] This mechanism effectively ensures the continuity of access to high-resolution images during processing, reduces disk I / O jitter, and improves the overall detection speed.

[0176] S1044. Monitor the occupied space of the temporary virtual disk, and perform a write-back operation when the occupied space reaches the preset write-back threshold. Then, determine the target disk group in turn among multiple physical disk groups through a roulette-style selection strategy.

[0177] To achieve write balancing and disk lifespan management, a round-robin strategy can be used to select the target disk group from multiple physical disk groups in turn for batch data write-back. For example, if the current write target is disk group A, the next write-back will automatically switch to disk group B, the next to disk group C, and so on, until the cycle ends and returns to disk group A. This strategy can avoid I / O hotspots caused by prolonged high write loads on a single disk group, thus achieving dynamic balancing of write pressure.

[0178] S1045. The images to be detected in the temporary virtual disk are scattered and written into the target disk group.

[0179] Upon triggering a write-back, the images to be detected in the virtual disk are distributed and written to the target disk group. During the write process, block-level compression and index reconstruction operations are performed for subsequent traceability. If the detection task is completed and the images no longer need to be retained, the system can also start an automatic cleanup process to release the corresponding memory pages and destroy the temporary virtual disk to improve system resource utilization.

[0180] After completing the above data preparation process, the ADC server enters the second detection stage. In this embodiment, the ADC server calls a preset second detection algorithm module to perform depth detection on the image to be detected that has been written to the virtual disk. The second detection algorithm can be based on a deep convolutional neural network, feature extraction and pattern recognition model, or a combination of statistical thresholding and spatial filtering algorithms. Specifically, the ADC server maps the virtual disk path to the input interface of the detection engine, and the detection engine directly loads image data from the virtual disk into the video memory buffer, avoiding repeated disk I / O.

[0181] S105. The first detection result and the second detection result are fused together by the dispatch software to obtain a third detection result, and the third detection result is transmitted to the MES system.

[0182] In one specific implementation, step S105 is used to fuse the first detection result and the second detection result through dispatch software to obtain a third detection result, and then transmit the third detection result to the MES system.

[0183] Specifically, after the ADC server completes the second detection, it packages the detection results, including metadata such as the detected image identifier, defect coordinates, defect type, confidence value, and processing time, and sends them to the result fusion module of the dispatch software. After receiving the detection results from multiple ADC servers, the dispatch software initiates the fusion processing flow to achieve result consistency and information enhancement across the detection system.

[0184] See Figure 4 In one feasible implementation, the fusion process can be achieved as follows:

[0185] S1051. Extract defect candidate sets from the first detection result and the second detection result respectively. Each candidate in the defect candidate set includes defect location, defect type, confidence level, piece number and timestamp.

[0186] In this step, defect candidate sets are extracted from the first and second detection results, respectively. Each candidate in the defect candidate set includes the defect location, defect type, confidence level, slice number, and timestamp. In actual implementation, the dispatch software can call a standardized interface to read the AFT detection result file and the ADC detection result file, and parse the defect candidate information based on a unified data structure (e.g., JSON or binary table structure). Each candidate is indexed using image spatial coordinates and time identifiers to facilitate subsequent matching and fusion. To improve retrieval efficiency, the system can construct a hash index structure based on slice number and timestamp in memory to achieve rapid location of candidate items.

[0187] S1052. In the defect candidate set, alignment is performed based on the piece number, timestamp, and defect location to obtain paired candidate groups and unilateral candidate groups.

[0188] Specifically, the unique correspondence between images is first determined based on the image number and timestamp. Then, spatial coordinate matching algorithms (such as Euclidean distance or pixel neighborhood window matching) are used to spatially align the defect locations. When the defect coordinates of two detection results are within a preset tolerance range (e.g., within 3 pixels), they are considered to be the same defect target, thus forming a pairing candidate group. If a result appears only on one side, it is marked as a unilateral candidate for subsequent confidence assessment or backtracking.

[0189] S1053. For the paired candidate groups, map the confidence levels of the first detection result and the second detection result to a unified confidence scale.

[0190] Because the AFT and ADC detection algorithms use different confidence definition models (e.g., AFT uses a probability score based on brightness difference measurement, while ADC uses a confidence probability based on the output of a convolutional feature classifier), the system needs to standardize the confidence. In this embodiment, a confidence mapping function can be constructed based on the statistical distribution of the two types of results in historical detection records to unify them to the [0,1] interval; for example, linear interpolation or piecewise normalization methods can be used to align the AFT output interval (0.3–0.9) and the ADC output interval (0.6–0.99) to achieve comparability.

[0191] S1054. Determine the fusion weight for each pair of candidate pairs;

[0192] For each pair of candidate pairs, a fusion weight is determined; the fusion weight is determined by at least one of the following:

[0193] The historical accuracy calibration coefficient of the model is obtained based on the processing records of the dispatch software.

[0194] The credibility verification value of the current model version;

[0195] The defect type weight of the paired candidate group;

[0196] The importance weight of the region where the paired candidate group is located.

[0197] In practice, the assignment software maintains a dynamic database of model accuracy, recording the accuracy and false positive rates of different detection model versions in historical batches. The fusion weights can be automatically adjusted based on the historical performance of each model in this database.

[0198] For example, for ADC models that have passed multiple rounds of verification and have high stability, their historical accuracy calibration coefficient can be set to 0.9; while if the current version is in the testing phase, the corresponding confidence verification value is reduced to prevent the results of the new model from having an excessive impact on the overall fusion. In addition, the system can also set different weighting ratios based on the importance of defect types (such as point defects, line defects, and uneven brightness) and image regions (such as the central area and the edge area), thereby achieving targeted optimization during fusion.

[0199] S1055. Calculate the fusion parameters based on the fusion weights and the unified confidence scale;

[0200] In this embodiment, the fusion parameters can be obtained through a weighted average or a nonlinear combination function, for example:

[0201] ;

[0202] in, and Let w1 and w2 represent the detection confidence levels of the two classes, respectively, and w1 and w2 be the fusion weights. In some implementations, to enhance sensitivity to low-confidence samples, a sigmoid function can be used for fusion, making the high-confidence results have a greater impact on the fusion output, thereby improving the robustness of the overall judgment.

[0203] S1056. Classify the paired candidate group and the unilateral candidate according to the fusion parameters and the preset fusion threshold set to obtain a third detection result.

[0204] In one specific embodiment, if the fusion parameter is greater than a preset first threshold, the corresponding pairing candidate group is marked as "determined defect";

[0205] If the fusion parameter is between a preset first threshold and a second threshold, the corresponding pairing candidate group is marked as "suspected defect";

[0206] If the fusion parameter is lower than the second threshold, the corresponding pairing candidate group is ignored;

[0207] The fusion threshold set may include multiple levels of thresholds for distinguishing between "definite defects", "potential defects", and "non-defects". For example, when the fusion confidence level is higher than 0.85, it is determined to be a definite defect; when the confidence level is between 0.6 and 0.85, it is marked as a defect to be confirmed and added to the re-evaluation list; if it is lower than 0.6, it is considered a non-defect.

[0208] For unilateral candidates, the system can independently apply a unilateral threshold strategy based on their source model and historical accuracy. For example, when a candidate comes from a high-precision ADC model and has a confidence level exceeding 0.9, it can be directly included in the final result. The final third detection result includes defect location, type, fusion confidence level, fusion status identifier, and corresponding task identifier information, which the system packages and transmits to the MES system.

[0209] In another optional embodiment, if the difference in confidence between two candidates in the paired candidate group is greater than a preset difference threshold, the defect type with the higher confidence level is determined as the classification result.

[0210] If the difference in confidence between two candidates in the paired candidate group is lower than a preset difference threshold, the corresponding paired candidate group will be transferred to the backup ADC model for re-execution of the detection.

[0211] In one feasible implementation, for cases where the confidence scores from AFT detection and ADC detection differ in dimensions, distribution, or interpretability, a confidence mapping process can be further performed to calibrate the two types of confidence scores to a unified scale. This confidence mapping can be implemented in one or more of the following ways.

[0212] In one implementation, a linear normalization (min–max scaling) strategy can be used for the confidence scores of the two model outputs. Specifically, the historical distribution intervals [Cmin, Cmax] of the confidence scores of the AFT model and the ADC model can be statistically analyzed separately, and the current confidence score C can be mapped to:

[0213] ;

[0214] This method compresses both types of confidence levels into the [0,1] interval. When the confidence level distribution is relatively stable, this method can quickly achieve dimensional alignment.

[0215] Secondly, in another implementation, a probabilistic calibration method can be used to uniformly calibrate the confidence level in a probabilistic sense. For example, the system can use Plattscaling to input the original output s into the following logic function:

[0216] ;

[0217] Here, A and B are calibration parameters obtained from the validation set statistics. Alternatively, isotonic regression can be used to construct a monotonic regression model based on the true labels, mapping the original confidence scores to a scale that better reflects the true probabilities. The above calibration methods can effectively address situations where the model output is biased by excessively high or low confidence, resulting in greater consistency after aggregating results from different sources.

[0218] Furthermore, in some application scenarios, the confidence distributions output by the two types of models may exhibit overall offset or shape differences. In such cases, a distribution-driven scale alignment method can be employed. For example, by minimizing the KL divergence between the two confidence distributions, the mapped confidence distributions can be made statistically more consistent. Specifically, a mapping function f(·) can be constructed to minimize the following function:

[0219] ;

[0220] This allows for the acquisition of a mapping relationship that best aligns the distribution features. This method enables deep alignment across model distribution levels and exhibits better stability in large-scale production line data.

[0221] In another feasible implementation, in order to improve the accuracy of determining the correspondence between defect candidates from AFT detection and ADC detection, overlap matching of the two types of candidates can be performed based on the defect box intersection over union (IoU).

[0222] Specifically, the system obtains the set of defect box coordinates from the AFT detection results and the set of defect box coordinates from the ADC detection results. For any defect box of an AFT candidate, the Intersection over Union (IoU) is calculated with the defect boxes of all ADC candidate. A set of preset overlap thresholds can be pre-set. When the calculated IoU value is greater than the overlap threshold, the corresponding defect candidates are considered to point to the same potential defect and are added to the paired candidate group. If the IoU of an AFT box with all ADC boxes is lower than the overlap threshold, the AFT box is classified as a unilateral candidate. The same applies to ADC boxes.

[0223] Furthermore, in another alternative approach, to avoid matching errors caused by a single IoU threshold, the system can adopt a maximum IoU matching strategy, that is, for each AFT defect box, select the ADC defect box with the largest IoU as its matching object, while ensuring that the maximum IoU value exceeds the minimum recognition threshold.

[0224] Furthermore, for cases where multiple AFT boxes or ADC boxes point to the same defect area, the Hungarian Algorithm can be used to find the matching group with the lowest cost, thereby obtaining a one-to-one candidate combination.

[0225] The foregoing embodiments have provided a detailed description of the AFT and ADC joint detection method for large-capacity image processing provided in this application. The embodiments of the AFT and ADC joint detection related apparatus for large-capacity image processing provided in this application are described below.

[0226] See Figure 5 This embodiment provides a joint AFT and ADC detection device for large-capacity image processing. The device includes:

[0227] The first detection unit 501 is used to perform a first detection on the image to be detected through the AFT device and obtain a first detection result;

[0228] Transmission unit 502 is used to transmit the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software via the AFT device in a memory-to-memory manner;

[0229] The dispatch unit 503 is used to receive images to be detected from multiple AFT devices through the dispatch software, and dispatch each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy.

[0230] The second detection unit 504 is used to write the image to be detected to a virtual memory disk through the ADC server, and to perform a second detection on the image to be detected through a second detection algorithm to obtain a second detection result.

[0231] The fusion unit 505 is used to fuse the first detection result and the second detection result through the dispatch software to obtain a third detection result, and transmit the third detection result to the MES system.

[0232] Optionally, the transmission unit 502 is specifically used for:

[0233] The AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the target shared memory buffer.

[0234] The AFT device negotiates with the dispatch software to adopt a zero-copy memory mapping transmission mechanism;

[0235] Based on the zero-copy memory mapping transfer mechanism, a mapping relationship is constructed between the first memory page and the second memory page in the target shared memory buffer;

[0236] Based on the mapping relationship, the image to be detected and the corresponding first detection result in the memory buffer are transmitted to the dispatch software.

[0237] Optionally, the transmission unit 502 is specifically used for:

[0238] The dispatch software allocates a shared memory identifier and sends the shared memory identifier and the corresponding target page table address to the AFT device.

[0239] The kernel interface is invoked through the AFT device according to the shared memory identifier, and a lock is requested for the first memory page used for transmission;

[0240] In the AFT device, the physical address of the first memory page is registered in the page table of the transmission channel;

[0241] The dispatch software creates a second memory page in the local process address space and maps the virtual address of the second memory page to the corresponding physical address through the page table.

[0242] Optionally, dispatch unit 503 is specifically used for:

[0243] The assignment software establishes a binding table between the AFT device and the ADC server;

[0244] A priority is determined for each arriving image to be detected by the dispatch software, and the image is then added to a priority queue according to the priority.

[0245] Based on the real-time load of the ADC server, the queue length of the priority queue, and the response time, a scheduling score is calculated, and the target ADC server is determined based on the scheduling score.

[0246] Optionally, the second detection unit 504 is specifically used for:

[0247] Construct a specified capacity region in the physical disk group of the ADC server;

[0248] A temporary virtual disk is created based on the specified capacity area, and the temporary virtual disk is mapped to an independent logical drive letter;

[0249] The image to be detected, passed in by the dispatch software, is written to the temporary virtual disk, and the memory pages occupied are locked during the writing process.

[0250] The space occupied by the temporary virtual disk is monitored. When the space occupied reaches the preset write-back threshold, a write-back operation is performed. The target disk group is determined in turn among multiple physical disk groups through a roulette-style selection strategy.

[0251] The images to be detected in the temporary virtual disk are scattered and written to the target disk group.

[0252] Optionally, the fusion unit 505 is specifically used for:

[0253] Defect candidate sets are extracted from the first detection result and the second detection result respectively. Each candidate in the defect candidate set includes defect location, defect type, confidence level, piece number and timestamp.

[0254] In the set of defect candidates, alignment is performed based on the piece number, timestamp, and defect location to obtain paired candidate groups and unilateral candidate groups;

[0255] For the paired candidate groups, the confidence levels of the first detection result and the second detection result are mapped to a unified confidence scale.

[0256] For each pair of candidate pairs, a fusion weight is determined, which is determined by at least one of the following:

[0257] The model's historical accuracy calibration coefficients are obtained based on statistical analysis of the processing records of the dispatch software.

[0258] The credibility verification value of the current model version;

[0259] The defect type weight of the paired candidate group;

[0260] The importance weight of the region where the paired candidate group is located;

[0261] The fusion parameters are calculated based on the fusion weights and the unified confidence scale.

[0262] The results of the paired candidate group and the unilateral candidate are classified according to the fusion parameters and the preset fusion threshold set to obtain a third detection result.

[0263] Optionally, the fusion unit 505 is specifically used for:

[0264] If the fusion parameter is greater than a preset first threshold, the corresponding pairing candidate group is marked as "determined defect";

[0265] If the fusion parameter is between a preset first threshold and a second threshold, the corresponding pairing candidate group is marked as "suspected defect";

[0266] If the fusion parameter is lower than the second threshold, the corresponding pairing candidate group is ignored.

[0267] Optionally, the fusion unit 505 is specifically used for:

[0268] If the difference in confidence between two candidates in the paired candidate group is greater than a preset difference threshold, the defect type with the higher confidence level is determined as the classification result.

[0269] If the difference in confidence between two candidates in the paired candidate group is lower than a preset difference threshold, the corresponding paired candidate group will be transferred to the backup ADC model for re-execution of the detection.

[0270] Please see Figure 6 This application also provides a joint AFT and ADC detection device for large-capacity image processing, comprising:

[0271] Processor 601, memory 602, input / output unit 603, bus 604;

[0272] The processor 601 is connected to the memory 602, the input / output unit 603, and the bus 604;

[0273] The memory 602 stores a program, and the processor 601 calls the program to execute any of the methods described above.

[0274] This application also relates to a computer-readable storage medium on which a program is stored, which, when run on a computer, causes the computer to perform any of the methods described above.

[0275] 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.

[0276] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between apparatuses or units through some interfaces, and may be electrical, mechanical, or other forms.

[0277] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0278] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0279] 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 this application, 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 this application. 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.

Claims

1. A method for joint detection of AFT and ADC for large-capacity image processing, characterized in that, The method includes: The first detection is performed on the image to be detected using the AFT device to obtain the first detection result; The AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software in a memory-to-memory manner. The dispatch software receives images to be detected from multiple AFT devices and dispatches each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy. The ADC server writes the image to be detected to a virtual memory disk, and performs a second detection on the image to be detected using a second detection algorithm to obtain a second detection result. The first detection result and the second detection result are fused together by the dispatch software to obtain a third detection result, and the third detection result is transmitted to the MES system.

2. The AFT and ADC joint detection method for large-capacity image processing according to claim 1, characterized in that, The AFT device transmits the image to be detected and the corresponding first detection result stored in local memory to the dispatch software in a memory-to-memory manner, including: The AFT device transmits the image to be detected and the corresponding first detection result stored in the local memory to the target shared memory buffer. The AFT device negotiates with the dispatch software to adopt a zero-copy memory mapping transmission mechanism; Based on the zero-copy memory mapping transfer mechanism, a mapping relationship is constructed between the first memory page and the second memory page in the target shared memory buffer; Based on the mapping relationship, the image to be detected and the corresponding first detection result in the memory buffer are transmitted to the dispatch software.

3. The AFT and ADC joint detection method for large-capacity image processing according to claim 2, characterized in that, The transfer mechanism based on the zero-copy memory mapping, which constructs the mapping relationship between the first memory page and the second memory page in the target shared memory buffer, includes: The dispatch software allocates a shared memory identifier and sends the shared memory identifier and the corresponding target page table address to the AFT device. The kernel interface is invoked through the AFT device according to the shared memory identifier, and a lock is requested for the first memory page used for transmission; In the AFT device, the physical address of the first memory page is registered in the page table of the transmission channel; The dispatch software creates a second memory page in the local process address space and maps the virtual address of the second memory page to the corresponding physical address through the page table.

4. The AFT and ADC joint detection method for large-capacity image processing according to claim 1, characterized in that, The step of assigning each of the images to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy includes: The assignment software establishes a binding table between the AFT device and the ADC server; A priority is determined for each arriving image to be detected by the dispatch software, and the image is then added to a priority queue according to the priority. Based on the real-time load of the ADC server, the queue length of the priority queue, and the response time, a scheduling score is calculated, and the target ADC server is determined based on the scheduling score.

5. The AFT and ADC joint detection method for large-capacity image processing according to claim 1, characterized in that, The step of writing the image to be detected to the virtual memory disk through the ADC server includes: Construct a specified capacity region in the physical disk group of the ADC server; A temporary virtual disk is created based on the specified capacity area, and the temporary virtual disk is mapped to an independent logical drive letter; The image to be detected, passed in by the dispatch software, is written to the temporary virtual disk, and the memory pages occupied are locked during the writing process. The space occupied by the temporary virtual disk is monitored. When the space occupied reaches the preset write-back threshold, a write-back operation is performed. The target disk group is determined in turn among multiple physical disk groups through a roulette-style selection strategy. The images to be detected in the temporary virtual disk are scattered and written to the target disk group.

6. The AFT and ADC joint detection method for large-capacity image processing according to claim 1, characterized in that, The step of fusing the first detection result and the second detection result through the dispatch software to obtain the third detection result includes: Defect candidate sets are extracted from the first detection result and the second detection result respectively. Each candidate in the defect candidate set includes defect location, defect type, confidence level, piece number and timestamp. In the set of defect candidates, alignment is performed based on the piece number, timestamp, and defect location to obtain paired candidate groups and unilateral candidate groups; For the paired candidate groups, the confidence levels of the first detection result and the second detection result are mapped to a unified confidence scale. For each pair of candidate pairs, a fusion weight is determined, which is determined by at least one of the following: The historical accuracy calibration coefficient of the model is obtained based on the processing records of the dispatch software. The credibility verification value of the current model version; The defect type weight of the paired candidate group; The importance weight of the region where the paired candidate group is located; The fusion parameters are calculated based on the fusion weights and the unified confidence scale. The results of the paired candidate group and the unilateral candidate are classified according to the fusion parameters and the preset fusion threshold set to obtain a third detection result.

7. The AFT and ADC joint detection method for large-capacity image processing according to claim 6, characterized in that, The third detection result is obtained by classifying the paired candidate groups and the unilateral candidate groups according to the fusion parameters and the preset fusion threshold set. If the fusion parameter is greater than a preset first threshold, the corresponding pairing candidate group is marked as "determined defect"; If the fusion parameter is between a preset first threshold and a second threshold, the corresponding pairing candidate group is marked as "suspected defect"; If the fusion parameter is lower than the second threshold, the corresponding pairing candidate group is ignored.

8. The AFT and ADC joint detection method for large-capacity image processing according to claim 7, characterized in that, The method further includes: If the difference in confidence between two candidates in the paired candidate group is greater than a preset difference threshold, the defect type with the higher confidence level is determined as the classification result. If the difference in confidence between two candidates in the paired candidate group is lower than a preset difference threshold, the corresponding paired candidate group will be transferred to the backup ADC model for re-execution of the detection.

9. A combined AFT and ADC detection device for large-capacity image processing, characterized in that, The device includes: The first detection unit is used to perform a first detection on the image to be detected using an AFT device to obtain a first detection result; The transmission unit is used to transmit the image to be detected and the corresponding first detection result stored in the local memory to the dispatch software via the AFT device in a memory-to-memory manner. The dispatch unit is used to receive images to be detected from multiple AFT devices through the dispatch software, and dispatch each image to be detected to at least one ADC server according to a preset binding relationship and priority scheduling strategy. The second detection unit is used to write the image to be detected to a virtual memory disk through the ADC server, and to perform a second detection on the image to be detected through a second detection algorithm to obtain a second detection result. The fusion unit is used to fuse the first detection result and the second detection result through the dispatch software to obtain a third detection result, and then transmit the third detection result to the MES system.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains a program that, when executed on a computer, performs the method as described in any one of claims 1 to 8.

Images