A cable terminal detection method and system based on bimetallic feature structure counting

Abstract

The application discloses a kind of cable terminal detection method and system based on bimetallic local structure counting, it is suitable for the terminal of two oppositely arranged and allowing position offset and size difference metal local structure in circular port. Obtain end face image, two metal local structures are detected as target in port area respectively, and effective target is obtained by confidence screening and overlap suppression;When effective target number is 2, it is judged to be qualified, otherwise, output exception or pending review. When model is constructed, pre-labeling model is trained with a small amount of artificial labeling image, image is shunted to count exception, low confidence and high confidence random inspection three ways for manual confirmation, and then detection model is trained;Production error is relabeled and fine-tuned according to data set version, and the model is updated after reaching threshold by regression verification. The application improves the explainability of internal structure loss and reduces the labeling and maintenance cost.

Description

Technical Field

[0001] This invention relates to the fields of industrial machine vision and cable terminal quality inspection technology, specifically to a method, model building method, system, and storage medium for target detection, validity judgment, and quantity determination of two metal local structures within a circular port. Background Technology

[0002] After industrial cable terminal assembly, the port structure typically needs to be inspected. For terminals with two opposing metal local structures within a circular port, existing methods treat the entire end face as a single classification object or the entire port as a single inspection object. While these methods can determine the overall appearance, they cannot directly indicate whether the two internal structures are present and identifiable. Furthermore, when one structure is slightly obscured, existing methods struggle to accurately identify its presence, further increasing the risk of missed detections.

[0003] Deep learning-based object detection can provide local regions and confidence levels, but the number of abnormal samples in industrial settings is limited, and manual annotation of each sample is costly. In practice, a limited number of erroneous samples can be collected, and the training samples can be expanded through data augmentation techniques such as geometric transformations and brightness adjustments. While training a pre-labeled model using a small-scale seed data can improve annotation efficiency, if only low-confidence samples are reviewed, erroneous labels generated by the model with high confidence levels may still be included in the training set.

[0004] Therefore, a detection and model building scheme that matches the specific terminal structure is needed. Under the premise that the two metal local structures have positional offsets and size differences, the quantity determination is completed based on whether the structures form independent and valid targets. The model iteration risk is controlled by three-way verification through counting anomalies, low confidence and high confidence random sampling, and production error return flow. Summary of the Invention

[0005] Technical problem: The present invention aims to solve the problems that the whole end face classification or whole port detection cannot directly reflect whether the two internal metal local structures are complete, the cost of manual annotation of samples is high, the automatic pre-annotation error is easy to accumulate, and the production error is difficult to form a traceable model update.

[0006] To address this, the present invention provides a cable terminal detection method based on counting bimetallic feature structures: after acquiring an image of the terminal end face, a target detection model is used to detect two local metal structures respectively. Valid targets are obtained through confidence screening and overlap suppression, with the number of valid targets being two as the criterion. A corresponding model construction method is also provided: a pre-labeled model is trained using a small number of manually labeled images; an expanded training dataset is constructed through three-way random sampling (anomaly counting, low confidence, and high confidence) for manual review; and the model is continuously iterated through production error sample re-processing and regression validation. An industrial cable terminal detection system and a computer-readable storage medium for implementing the above method are also provided.

[0007] Technical Solution: This invention is applicable to industrial cable terminals with two opposing metal partial structures within a circular port. The two metal partial structures may have some positional offset and dimensional differences, and are not required to be strictly symmetrical relative to the port center. In field communication, this metal partial structure can be commonly referred to as a "petal"; this term is only used to help understand its image shape and does not limit the conductive function, mechanical function, material grade, or strict geometric shape of the structure. The method of this invention is applicable to various similar cable terminal end face types, including terminals with different materials, different imaging colors, and different petal positions, such as... Figure 2 As shown. Except as defined in this paragraph, the term "metallic local structure" will be used consistently throughout this article.

[0008] A cable terminal detection method based on counting bimetallic local structures is proposed. The industrial cable terminal has a circular port with two opposing metallic local structures within it. Positional offset and dimensional differences between the two metallic local structures are permitted. The method includes: acquiring an end-face image containing the circular port; detecting the two metallic local structures using a target detection model to obtain candidate target regions and corresponding confidence levels; filtering the candidate target regions to obtain valid targets that can form independent and identifiable structures in the end-face image; counting the number of valid targets; outputting a pass / fail judgment when the number of valid targets is 2, and outputting an anomaly judgment or a judgment pending review when the number of valid targets is not 2.

[0009] The process of acquiring the end face image containing the circular port includes: determining the region of interest corresponding to the circular port, excluding background candidate targets outside the region of interest, and performing size normalization, brightness normalization, contrast adjustment, and geometric correction on the region of interest; the image acquisition area is limited to cover the circular port and two local metal structures inside it.

[0010] This invention inspects the metal structure within a circular port. If a local metal structure has minor defects but still forms an independently identifiable target, it is counted as a valid target. If a local metal structure is missing, obviously broken, severely folded, or severely obscured and cannot form an independently identifiable target, it is not counted as a valid target. When an anomaly judgment or a judgment pending review is output, the initial image, candidate target region, confidence level, and model version are saved as detection data. Compared to existing solutions that only classify and inspect the entire end face or the entire circular port, this invention directly detects two local metal structures within the circular port, avoiding interference from the overall contour features of the circular port in the determination of internal structural defects. This reduces the likelihood of terminals with abnormal internal structures being mistakenly identified as qualified products due to a normal overall appearance of the port.

[0011] A model building method for constructing an object detection model includes: acquiring a seed end-face image set; manually annotating each effective local metal structure in each seed end-face image; training a pre-annotated model using the manually annotated seed end-face image set; using the pre-annotated model to infer unannotated end-face images to obtain candidate annotations, the number of candidate annotations, and the confidence level of candidate annotations; manually verifying the candidate annotations to form an extended training dataset; and training an object detection model for industrial cable terminal detection using the extended training dataset.

[0012] Specifically, end-face images with no more than 2 candidate annotations are assigned to the count anomaly review queue; end-face images with 2 candidate annotations but whose confidence level is lower than the automatic acceptance threshold are assigned to the low confidence review queue; and a portion of end-face images are randomly selected from those with 2 candidate annotations and whose confidence level is not lower than the automatic acceptance threshold according to a preset ratio and assigned to the high confidence random sampling queue; annotations are corrected or accepted based on the manual confirmation results of the three queues.

[0013] The system acquires manually verified false detection images, missed detection images, or images awaiting review generated by the target detection model during production testing. It re-annotates these images and records their source, anomaly type, and corresponding model version. The re-annotated images are then added to a training dataset with version identifiers for model fine-tuning. Regression validation is performed on a fixed regression dataset and a newly added error example dataset. The fine-tuned model is updated to the production model only when the product judgment index, anomaly detection index, and inference performance meet the preset update threshold.

[0014] An industrial cable terminal inspection system includes an image acquisition unit, a processor, and a memory. The memory stores a target detection model and a computer program. When the computer program is executed by the processor, the processor performs an industrial cable terminal inspection method. The system also includes a sample distribution interface, a manual review interface, a dataset version management module, and a model regression verification module to execute the model construction method.

[0015] Beneficial effects: Compared with the prior art, the present invention has the following significant beneficial effects:

[0016] First, the overall port-level judgment is transformed into the independent detection and counting of two internal metal local structures. This allows structures that are missing or have lost their independent identifiable state to be directly expressed by the number of effective targets, avoiding the situation where terminals with abnormal internal structures are mistakenly judged as qualified products due to the normal overall appearance of the port. At the same time, since the two metal local structures are detected and located separately, it can adapt to the positional offsets and dimensional differences that occur during actual assembly and imaging.

[0017] Second, by employing a combination of manual annotation of a small number of seeds, automatic pre-annotation of batches of unannotated images, and a three-way split-checking process, the need for entirely manual annotation is significantly reduced, thereby improving overall annotation efficiency. The three-way checking process—including outlier counting, low-confidence, and high-confidence random sampling—takes into account both difficult samples and high-confidence errors, reducing the risk of continuous accumulation of erroneous pre-annotations.

[0018] Third, save detection metadata, dataset version, and model version to make production error re-flow, model fine-tuning, and regression validation traceable.

[0019] Fourth, the initial detection network is not limited to a specific version, which makes it easy to select the matching model according to industrial computing power and speed requirements, and has high compatibility. Attached Figure Description

[0020] Figure 1 The flowchart of the overall method for industrial cable terminal testing shows the closed loop of model construction, target counting and judgment, and error return model update.

[0021] Figure 2 : Schematic diagrams of various cable terminal faces, showing typical terminal types with different materials, colors, and petal positions.

[0022] Figure 3 : Schematic diagram of local features of bimetals to be detected on various end faces, indicating the typical shape and position of two local metal structures (petals).

[0023] Figure 4 The initial model building flowchart includes seed data annotation, pre-annotated model training, and three-way split manual review process.

[0024] Figure 5 The flowchart of the bimetallic structure detection, counting and judgment module includes ROI extraction, target detection, screening and deduplication, validity judgment and quantity judgment. Figure 6 The diagram shows the architecture of the industrial cable terminal testing system of this invention. It is a layered architecture diagram of the industrial cable terminal testing system, which is divided into a hardware access layer and a software functional module layer from top to bottom. The software layer is divided into two independent working sub-processes: real-time quality inspection of the production line and model iterative optimization. The diagram shows the two-way closed-loop interaction relationship between the synchronization of images to be reviewed and the iterative update of the model. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0026] Terminology and Judgment Principles

[0027] "Relative arrangement" refers to two metal local structures located in two expected areas within a circular port, but it is not required that the line connecting their centers pass through the center of the port, nor that their areas, angles, or distances to the port center be exactly the same. The two preset detection areas can be translated or scaled as a whole based on the port positioning results.

[0028] "Independently identifiable targets" refer to local metallic structures that, under given imaging and annotation rules, still possess visible areas distinguishable from the background and another structure, and can be represented by an independent target bounding box. Minor wear, minor deformation, or localized reflections that do not prevent the formation of an independent target are still counted; missing, obvious breaks, severe folding, or severe occlusion that prevent the formation of an independent target are not counted.

[0029] Example:

[0030] Example 1: Image Acquisition and Data Preparation

[0031] This embodiment is used to complete the acquisition, preprocessing, and initial seed dataset construction of terminal end face images. Figure 2 The sample shows various types of terminals.

[0032] An industrial CCD camera and a ring-shaped supplementary light source are deployed at a fixed imaging station on the production line to acquire images of the cable terminal end faces. The image resolution is set to no less than 640×640 pixels to ensure that the circular port and its internal bimetallic structure are completely within the effective field of view. After acquisition, port positioning and ROI cropping are performed on the original image, and standardized preprocessing operations such as size normalization, brightness equalization, and contrast enhancement are performed as needed to eliminate interference caused by imaging lighting and terminal assembly misalignment.

[0033] Figure 2Images (a), (b), and (c) respectively show terminal facets made of different materials, metals, and from different production lines, representing typical test samples compatible with this invention. This embodiment uses... Figure 2 (a) The dataset is built based on the type of terminal, and the solution can be directly reused in (b)(c) production lines of the same type of terminal.

[0034] In this embodiment, a total of 152 manually labeled seed images were prepared, including 129 qualified samples with complete bimetallic structures, 12 single-metal incomplete samples, and 11 samples with poor metal structures. For each image, the OK / NG primary classification label was simultaneously saved, and target annotation boxes were drawn for each independently identifiable local metal structure within the image. The annotation rules were uniform: metal petals with only slight wear, reflection, or local deformation, but still possessing independently distinguishable contours, were marked as valid targets; metal petals that were completely missing, extensively broken, or severely folded and obscured, unable to form an independently identifiable area, were not annotated. Figure 3 This is a schematic diagram of the annotation effect of a single sample, which intuitively demonstrates how the manual annotation box selects local structures of metal.

[0035] Example 2: Pre-annotated model and three-way verification

[0036] This embodiment fully implements the closed-loop process of expanding a large-scale labeled dataset with a small amount of seed data, and the overall logic corresponds to... Figure 4 The model construction flowchart.

[0037] Step 1: Lightweight pre-labeled model training.

[0038] This embodiment trains a lightweight pre-labeled model based on the aforementioned dataset of 152 manually labeled seed images, using the YOLOv8n network architecture. However, this invention is not limited to this network, and those skilled in the art can replace it with other detection networks according to device computing power and inference latency requirements. The training parameters are configured as follows: input image size 640×640, total number of iterations 100, batch size automatically adapted to the training GPU memory; the optimizer is AdamW, and the initial learning rate is set to 0.001.

[0039] Samples with a confidence level below 0.3 are directly removed; samples with a confidence level greater than or equal to 0.5 are considered high-confidence samples; and samples with a confidence level between 0.3 and 0.5 are considered low-confidence samples.

[0040] Step 2: Automatic inference and three-way splitting of unlabeled images.

[0041] After the pre-annotated model is trained, automatic inference is performed on a large number of unannotated terminal face images, outputting candidate bounding boxes and corresponding confidence scores for local metal structures within the images. Based on the number of candidate targets and the confidence scores, the images are divided into three categories of manual review queues, with the following division rules:

[0042] 1. Counting anomaly verification queue: The number of candidate targets automatically detected in the image is not equal to 2;

[0043] 2. Low-confidence review queue: The total number of candidate targets is 2, but there are candidate boxes whose confidence is between the automatic acceptance threshold (0.3) and the high-confidence threshold (0.5);

[0044] 3. High-confidence random sampling queue: The total number of candidate targets is 2, and the confidence level of all candidates is ≥0.5. This type of image is automatically entered into the database by default. At the same time, samples are randomly selected according to the 10% stratified sampling ratio and sent for manual review to screen for hidden labeling errors under high confidence.

[0045] Step 3: Manual review, correction, and expansion of the dataset generation

[0046] Human reviewers uniformly verify the images in the three queues, supporting the addition of missing target boxes, deletion of erroneous candidate boxes, and adjustment of the coordinate size of the labeled boxes. After all the review and corrections are completed, an expanded training dataset with version identifiers is uniformly generated. In this embodiment, the total number of images in the expanded dataset is approximately 1,000.

[0047] Step 4: Training the Formal Industrial Testing Model

[0048] Using the publicly available YOLOv8n pre-trained weight file best.pt as initial parameters, an expanded dataset was used to train a target detection model specifically for industrial production lines. The training hardware consisted of NVIDIA series GPUs. Key training parameters were: input size 640×640, total number of iterations 100, fixed batch size 16, optimizer AdamW, and initial learning rate 0.001. Mixed precision AMP acceleration was enabled throughout the training process, and Mosaic data augmentation was disabled in the last 10 iterations to enhance the model's ability to fit small local metal features at terminals.

[0049] Step 5: Production Line Reasoning, Screening, and Judgment Logic

[0050] The trained detection model is deployed to the industrial control computer on the production line to perform real-time inference, and outputs all candidate target boxes within the ROI area of ​​the circular port; it sequentially performs preset detection area filtering, confidence threshold screening, and NMS (non-maximum suppression) overlap suppression triple screening to extract independent and effective local metal structures in the image; in this embodiment, the inference confidence threshold is fixed at 0.5 and the NMS intersection-union ratio (IoU) threshold is 0.5.

[0051] Judgment rules: If the number of valid targets after screening is equal to 2, an "OK" judgment is output; if the number of valid targets is 0, 1, or greater than 2, an "NG" or "manual review pending" judgment is output. The system synchronously stores all detection metadata, including the original end face image, ROI cropping image, coordinates of all candidate boxes, confidence scores, valid target count, judgment results, production batch, and current model version, providing a complete basis for subsequent error rework and model iteration traceability.

[0052] Example 3: Detection Model Training and Faulty Instance Regression Model Iteration

[0053] This embodiment covers the complete closed-loop process from online production line detection and judgment to production error reprocessing and model iteration fine-tuning. See [link to documentation]. Figure 5 .

[0054] During online inspection on the production line, the system first acquires an image of the cable end face containing the circular port and performs ROI extraction to eliminate background interference and pinpoint the target detection area. The preprocessed ROI area is then input into the target detection model, which outputs candidate bounding boxes and confidence scores. Subsequently, a filtering process is performed to remove duplicate candidate targets corresponding to the same local metal structure, obtaining valid targets that form independent and identifiable structures in the end face image. The system counts the valid targets after NMS overlap suppression. If the number of valid targets is 2, the system outputs an OK (OK) pass / fail judgment; if the number of valid targets is 0, 1, or greater than 2, it outputs an NG (Not Acceptable) / Pending Review judgment.

[0055] Regardless of whether the production line determines the image as qualified or abnormal, the system supports feeding back some images to the dataset, especially NG (Not Found), images awaiting re-inspection, and confirmed false positives, missed positives, and difficult images. The collected fed-back images are manually re-annotated, and the image source, anomaly type, and corresponding model version are recorded. The re-annotated images are added to an expanded training dataset with version identifiers and combined with the original training dataset at a preset ratio to prepare for subsequent fine-tuning updates.

[0056] When the model is updated and fine-tuned, the system loads the weights of the old model currently in use on the production line and performs offline fine-tuning training. To meet the accuracy requirements of industrial inspection, the AdamW optimizer is used during the fine-tuning phase, with an initial learning rate set to 0.001, and Automatic Mixed Precision (AMP) training is enabled to improve training efficiency. The normal fine-tuning epochs are set to 100, the input size remains at 640×640, and the batch size is set to 16. Specifically, in the last 10 epochs of training, Mosaic data augmentation is disabled (close_mosaic=10) to enhance the model's ability to regress and locate small local features on the end face.

[0057] After each fine-tuning training, a new dataset version and model iteration version are generated. The system performs independent regression validation on a fixed regression test set and a newly added error test set. Only when the product judgment accuracy, anomaly detection index, and single-frame inference performance of the new model fully meet the preset update thresholds will the system officially update the fine-tuned model as the production model. If the new model does not meet the thresholds, the update is rejected, and the original production version is retained, thereby strictly controlling the model iteration risk in industrial settings.

[0058] Example 4: Industrial cable terminal integrity testing system

[0059] This embodiment is a hardware and software integrated system adapted to the aforementioned detection method and model iteration method, corresponding to the present invention. Figure 6 The system architecture diagram shows that the system is divided into a hardware layer and a software functional module layer.

[0060] The hardware layer includes an industrial CCD image acquisition unit, a ring-shaped supplementary light source, an industrial control processor, a GPU inference unit, a data storage unit, and a human-machine interaction terminal. The image acquisition unit is fixedly arranged at a fixed workstation on the terminal production line, continuously acquiring images of the circular port end face of the terminal and transmitting them to the industrial control computer for real-time processing.

[0061] The software layer contains multiple decoupled functional modules, specifically including:

[0062] The image preprocessing module is used to perform ROI port location, size normalization, brightness and contrast correction, and background interference removal on the original end face image.

[0063] The target detection inference and decision output module is used to load the deployed ONNX detection model and complete bimetallic local structure inference, region filtering, confidence filtering, NMS overlap suppression, effective target counting and OK / NG decision output.

[0064] The sample three-way diversion and screening module is used to automatically complete the three-way queue diversion of unlabeled images and distinguish between samples with abnormal counts, low confidence, and high confidence.

[0065] The manual review and interaction module supports operators in completing tasks such as label correction, error confirmation, and sample re-inspection.

[0066] The dataset version management module is used to store seed datasets, extended datasets, and reverted error datasets, and assigns a unique version identifier to each iteration of data, associating it with the corresponding training model;

[0067] The model validation and update module has a built-in fixed regression test set and automatically detects the new model's judgment accuracy, anomaly detection rate, and inference speed. It completes the production model replacement and update only after the indicators meet the standards.

[0068] The system operates in two modes: a real-time production line detection mode and a model iteration optimization mode. In the real-time production line detection mode, only the image acquisition, preprocessing, detection inference, and judgment output modules are activated to meet the high-speed quality inspection requirements of the production line. In the model iteration optimization mode, the sample distribution, manual review, dataset management, model fine-tuning, and regression verification modules are activated to complete model iteration optimization. The hardware and software work together to achieve high-precision terminal detection, low annotation costs, and traceable, iterative closed-loop industrial quality inspection functions.

Claims

1. A method for detecting cable terminals based on counting bimetallic local structures, characterized in that, An industrial cable terminal has a circular port, within which two opposing metal local structures are provided. Positional offsets and dimensional differences are permitted between the two metal local structures. The method includes: acquiring an end-face image containing the circular port; using a target detection model to detect the two metal local structures respectively, obtaining candidate target regions and corresponding confidence levels; filtering the candidate target regions to obtain valid targets that can form independent and identifiable structures in the end-face image; counting the number of valid targets; outputting a pass / fail judgment when the number of valid targets is 2, and outputting an anomaly judgment or a judgment pending review when the number of valid targets is not 2.

2. The method according to claim 1, characterized in that, The process of acquiring the end face image containing the circular port includes: determining the region of interest corresponding to the circular port, excluding background candidate targets outside the region of interest, and performing size normalization, brightness normalization, contrast adjustment, and geometric correction on the region of interest; the image acquisition area is limited to cover the circular port and two local metal structures inside it.

3. The method according to claim 2, characterized in that, The target detection model is used to detect two local metal structures within the region of interest, and candidate target boxes and their corresponding confidence scores are output. The candidate target boxes are then filtered according to a preset confidence threshold, and duplicate candidate targets corresponding to the same local metal structure are removed by non-maximum suppression to obtain the effective targets.

4. The method according to claim 3, characterized in that, When a local metal structure has slight wear, slight deformation, or local reflection, but can still be distinguished from the background and another local metal structure and represented by an independent target box, it is counted as a valid target; when a local metal structure is completely missing, obviously broken, severely folded, or severely occluded, making it impossible to form an independent and identifiable target, it is not counted as a valid target; when outputting anomaly judgments or judgments pending review, the detection data such as the initial image, candidate target regions, confidence scores, and model versions are saved.

5. A model construction method for constructing the target detection model according to any one of claims 1 to 4, characterized in that, include: Obtain a set of seed end face images, and manually annotate each effective local metal structure in each seed end face image; A pre-labeled model is trained using a set of manually labeled seed face images; the pre-labeled model is used to infer from unlabeled face images to obtain candidate labels, the number of candidate labels, and the confidence of candidate labels; the candidate labels are manually confirmed to form an expanded training dataset. A target detection model for industrial cable terminal inspection was trained using an extended training dataset.

6. The model construction method according to claim 5, characterized in that, End face images with 2 or fewer candidate annotations are assigned to the count anomaly review queue. End face images with 2 or more candidate annotations but whose confidence level is lower than the automatic acceptance threshold are assigned to the low confidence review queue. From the end face images with 2 or more candidate annotations and whose confidence level is not lower than the automatic acceptance threshold, a portion of end face images are randomly selected according to a preset ratio and assigned to the high confidence random sampling queue. Based on the manual confirmation results of the three queues, the annotations are corrected or accepted.

7. The model construction method according to claim 5 or 6, characterized in that, The system acquires manually verified false detection images, missed detection images, or images awaiting review generated by the target detection model during production testing. These images are then re-annotated, and their source, anomaly type, and corresponding model version are recorded. The re-annotated images are added to a training dataset with version identifiers for model fine-tuning. The fine-tuned model is then validated using regression on a fixed regression dataset and a newly added error example dataset. The fine-tuned model is only updated to the production model when the product judgment index, anomaly detection index, and inference performance meet the preset update threshold.

8. An industrial cable terminal testing system, characterized in that, The system includes an image acquisition unit, a processor, and a memory. The memory stores a target detection model and a computer program. When the computer program is executed by the processor, the processor performs the industrial cable terminal detection method according to any one of claims 1 to 4. The system also includes a sample diversion interface, a manual review interface, a dataset version management module, and a model regression verification module to perform the model construction method according to any one of claims 5 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1 to 7.

Images