Plated metal quality real-time monitoring method based on deep learning

Abstract

The invention relates to an electroplated metal quality real-time monitoring method based on deep learning, and belongs to the technical field of industrial automation and intelligent detection. Comprising the following steps: step 1, image acquisition and preprocessing; step 2, deep learning of a feature extraction and detection framework; step 3, carrying out integrated detection without NMS post-treatment; and step 4, defect classification and real-time feedback. According to the invention, through multi-scale feature fusion and large kernel convolution, the capturing capability of tiny defects is significantly improved, high detection precision can be maintained in complex illumination and strong reflection environments, and the method has the characteristics of high precision, high efficiency and high adaptability, and is suitable for various industrial scenes.

Description

A Deep Learning-Based Real-Time Monitoring Method for Electroplated Metal Quality Technical Field

[0001] This invention belongs to the field of industrial automation and intelligent detection technology, specifically relating to a real-time monitoring method for electroplated metal quality based on deep learning. Background Technology

[0002] Industrial electroplating is a key metal surface treatment process, and its quality directly affects the performance and lifespan of products. In industries such as aerospace, automotive manufacturing, and electronic equipment, metal electroplating not only needs to enhance the corrosion resistance of materials but also improve their conductivity and mechanical properties. Therefore, the inspection of the surface quality of electroplated metals is particularly important.

[0003] Existing inspection methods mainly fall into two categories: manual inspection and traditional machine vision inspection. Manual inspection is time-consuming and inefficient, failing to meet the demands of high-paced industrial production lines. Furthermore, due to subjective experience and visual fatigue, it's difficult to guarantee consistent results. While traditional machine vision inspection offers a degree of automation, it lacks sensitivity to subtle defects such as pinholes, bubbles, and particles, performing poorly especially under varying lighting conditions or with reflective interference. In recent years, deep learning-based object detection algorithms have shown great potential in industrial inspection; however, existing algorithms still suffer from significant missed detections in subtle defect scenarios. Moreover, because current algorithm frameworks rely on post-processing steps, they cannot meet end-to-end real-time inspection requirements. Summary of the Invention

[0004] (I) Technical problem to be solved The technical problem to be solved by the present invention is to provide a real-time monitoring method for electroplated metal quality based on deep learning to solve the problem that existing algorithms still have obvious missed detections in the scenario of detecting minute defects.

[0005] (II) Technical Solution To solve the above technical problems, the present invention provides a real-time monitoring method for electroplated metal quality based on deep learning, including the following steps: Step 1: Image acquisition and preprocessing; Step 2: Deep learning feature extraction and detection framework; Step 3: Integrated detection without NMS post-processing; Step 4: Defect classification and real-time feedback.

[0006] Specifically, step 1 includes: Step 11: Image acquisition, deploying a high-resolution industrial camera and an LED ring light source on the production line, installing a polarizer in the light source path, adding a linear polarizing filter in front of the camera lens, capturing images at regular intervals via software control, and adjusting the exposure time according to the production line's rhythm to obtain an input image with stability and contrast; Step 12: Image preprocessing; Step 121: Denoising the image using a Gaussian filter, and smoothing the image by weighted averaging of each pixel and its neighborhood; Step 122: Enhancing the image contrast through histogram equalization, making the grayscale distribution in the image more uniform; Step 123: Normalizing the image to ensure that the pixel values ​​of the input image are within the same range, guaranteeing consistency of features when inputting into the model.

[0007] In step 2, the preprocessed image is input into a customized deep learning network. The deep learning network includes multiple convolutional layers, downsampling layers, and some self-attention modules. The downsampling strategy with spatial-channel decoupling preserves key information while compressing the feature map size.

[0008] In step 3, consistency detection is completed through a dual allocation strategy of "one-to-one" and "one-to-many" matching, enabling the detection framework to directly output high-quality prediction results.

[0009] In step 4, the defect category probability is output through the Softmax function, and minor defects are highlighted in the detection results. This is linked with the production line control system to issue an audible and visual alarm and dynamically adjust production parameters when a serious defect occurs.

[0010] In step 2, some self-attention modules adaptively highlight areas that may have defects in the feature map through an attention mechanism, thereby improving the detectability of minor flaws.

[0011] (III) Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects: 1) High real-time performance: The vision-based real-time monitoring method for industrial metal electroplating quality proposed in this invention significantly reduces inference latency by using a consistent dual allocation strategy, thus meeting the high efficiency requirements of industrial production lines.

[0012] 2) High precision and high robustness: This invention significantly improves the ability to capture tiny defects through multi-scale feature fusion and large kernel convolution, and can maintain high detection accuracy in complex lighting and strong reflective environments. It has the characteristics of high precision, high efficiency and high adaptability, and is suitable for various industrial scenarios. Attached Figure Description

[0013] Figure 1 is a detection block diagram of the present invention; Figure 2 is a schematic diagram of the network structure of the method of the present invention. Detailed Implementation

[0014] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0015] Example 1 This example provides a real-time monitoring method for electroplated metal quality based on deep learning, including: Step 1: High-quality image acquisition and preprocessing: By using hardware configurations such as industrial cameras, LED ring light sources, and polarizers, along with adaptive image enhancement technology, more stable and contrasting input images are obtained.

[0016] Step 2: Deep learning feature extraction and detection framework: A multi-scale feature extraction structure and a spatial-channel decoupled downsampling strategy are adopted, and a large kernel convolution and a partial self-attention (PSA) module are introduced to enhance the ability to capture subtle defect features.

[0017] Step 3: Integrated detection strategy without NMS post-processing: Consistency detection is achieved by using a dual allocation strategy of "one-to-one" and "one-to-many" matching, which enables the detection framework to directly output high-quality prediction results without post-processing steps, greatly improving the real-time performance of detection while ensuring accuracy.

[0018] Step 4: Defect Classification and Real-time Feedback: The Softmax algorithm outputs the probability of each defect category, and minor defects (such as pinholes, scratches, and peeling points) are highlighted in the detection results. Integrated with the production line control system, it can issue audible and visual alarms and dynamically adjust production parameters when serious defects occur.

[0019] Example 2 This example includes: Step 1: Image Acquisition A high-resolution industrial camera (such as 4K or higher resolution) and an LED ring light source are deployed on the production line to ensure image clarity and stable illumination. To reduce the effect of specular reflection from metal surfaces, polarizers can be installed in the light source path, and a linear polarizing filter can be added in front of the camera lens to minimize interference from bright reflective points. Images are captured at timed intervals via software control, and the exposure time is automatically or manually adjusted according to the production line's pace.

[0020] Step 2: Image Preprocessing. First, a Gaussian filter is used to reduce noise in the image. This is done by weighting each pixel and its neighborhood, using the formula... The calculations are performed to smooth the image, reduce the impact of noise, and ensure the reliability of image features in subsequent processing.

[0021] Secondly, histogram equalization is used to enhance image contrast, resulting in a more uniform grayscale distribution. To cope with changes in lighting conditions, adaptive histogram equalization can be employed, using the formula... This can be used to locally enhance the contrast of different areas. Furthermore, dynamic contrast enhancement technology can be employed to further enhance local details.

[0022] Subsequently, the image is normalized, mainly to ensure that the pixel values ​​of the input image are within the same range, thus guaranteeing the consistency of each feature when input into the model.

[0023] Finally, for images with special materials or complex textures, conventional filtering methods may not be effective in removing background noise. In such cases, guided filtering can be used, through the formula... This is used to process images and further remove complex background noise.

[0024] The above preprocessing steps can effectively remove noise, enhance image details, and provide stable and consistent input data for feature extraction in deep learning models, thereby improving model performance.

[0025] Step 3: Feature Extraction. The preprocessed image is input into a customized deep learning network. This network contains multiple convolutional layers, downsampling layers, and a portion of a self-attention module (PSA). A spatially-channel decoupled downsampling strategy compresses the feature map size while preserving key information. Large kernel convolutions help extract global information within a larger receptive field, thereby capturing the surrounding contextual features of subtle defects. The PSA module adaptively highlights potentially defective areas in the feature map through an attention mechanism, improving the detectability of minor flaws.

[0026] Assuming the preprocessed input image is Feature maps are extracted using a deep convolutional neural network: ,in This represents a feature extraction network, including multi-layer convolutions, downsampling, and non-linear activation. It employs large-kernel convolutions (such as...). (or larger) Increase the receptive field: .

[0027] The PSA module highlights potentially defective regions through an attention mechanism. Let the feature mapping be... Mapping via query (Q), key (K), and value (V): , , ,in , , The parameters are trainable, and the attention distribution is... PSA can compute the above attention distribution within a subset of channels and regions of the feature map, thereby highlighting minute defect features more effectively.

[0028] Step 4: Target detection employs a consistent dual-assignment strategy, matching candidate locations in the feature map with real defect targets in a "one-to-one" and "one-to-many" manner. This strategy improves the recall rate of defect detection, avoids redundant candidate boxes, and outputs stable prediction results. Thanks to this strategy, this invention eliminates the need for traditional post-processing steps such as NMS to obtain accurate, deduplicated detection results, achieving true end-to-end real-time detection.

[0029] In the detection task, to achieve post-processing without NMS, a "one-to-one" and "one-to-many" matching strategy is used to assign the correspondence between predicted bounding boxes and real defect targets. Let the set of real defects be... The set of prediction boxes is By constructing a cost matrix Define the cost of matching the prediction with the true target: ,in L1 or GloU loss can be used. Cross-entropy or focus loss can be used. The optimal match can be found in "one-to-one" and "one-to-many" scenarios using Hungarian matching. The "one-to-many" strategy can appropriately allocate multiple prediction boxes to the same real target to improve the recall rate of minor defects; the "one-to-one" strategy ensures the concentration and non-redundancy of predictions, and the final output can obtain high-precision results without NMS filtering.

[0030] Step 5: Defect Classification. The Softmax function is used to calculate class probabilities. For the detected target region, a fully connected layer and the Softmax function are used to determine the defect type. The system can be expanded to multi-class classification, including pinholes, blistering, peeling, scratches, and oxidation spots. By increasing the dataset size and the granularity of the class annotation, the accuracy of detection and classification can be continuously improved.

[0031] Extract the corresponding feature vector z from the final matched predicted bounding box, and classify the defect category using the Softmax function: ,in , For classifier parameters, That is, defects are categorized. The probability of.

[0032] Step 6: Results Output. The inspection system can highlight the defect location on the display terminal in the form of a bounding box, and display the defect category, confidence level, and quantity statistics in the supplementary information. Simultaneously, a quality assessment report is generated, summarizing the inspection results for the day or a certain period, so that quality management personnel can analyze the overall production quality trend.

[0033] Step 7: Real-time Feedback. The detection results are transmitted to the production line control system in real time via industrial control bus (such as EtherCAT, PROFINET) or API interface. When a serious defect is detected, the production line can automatically take measures, such as pausing production or changing electroplating parameters for rapid correction. The system can also automatically adjust the hyperparameters or screening strategies of the deep learning model based on performance indicators such as false detection rate and false negative rate to continuously optimize the detection effect.

[0034] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of electroplated metal quality based on deep learning, characterized in that, The process includes the following steps: Step 1: Image acquisition and preprocessing; Step 2: Deep learning feature extraction and detection framework; Step 3: Integrated detection without NMS post-processing; Step 4: Defect classification and real-time feedback.

2. The method for real-time monitoring of electroplated metal quality based on deep learning as described in claim 1, characterized in that, Step 1 specifically includes: Step 11: Image acquisition, deploying a high-resolution industrial camera and an LED ring light source on the production line, installing a polarizer in the light source path, adding a linear polarizing filter in front of the camera lens, capturing images at timed intervals via software control, and adjusting the exposure time according to the production line's rhythm to obtain an input image with stability and contrast; Step 12: Image preprocessing; Step 121: Denoising the image using a Gaussian filter, and smoothing the image by weighted averaging of each pixel and its neighborhood; Step 122: Enhancing the image contrast through histogram equalization to make the grayscale distribution in the image more uniform; Step 123: Normalizing the image to ensure that the pixel values ​​of the input image are within the same range, guaranteeing consistency of features when inputting into the model.

3. The method for real-time monitoring of electroplated metal quality based on deep learning as described in claim 2, characterized in that, In step 2, the preprocessed image is input into a customized deep learning network. The deep learning network contains multiple convolutional layers, downsampling layers, and some self-attention modules. The downsampling strategy with spatial-channel decoupling preserves key information while compressing the feature map size.

4. The method for real-time monitoring of electroplated metal quality based on deep learning as described in claim 3, characterized in that, In step 3, consistency detection is completed through a dual allocation strategy of "one-to-one" and "one-to-many" matching, enabling the detection framework to directly output high-quality prediction results.

5. The method for real-time monitoring of electroplated metal quality based on deep learning as described in claim 4, characterized in that, In step 4, the defect category probability is output through the Softmax function, and minor defects are highlighted in the detection results. This is linked with the production line control system to issue an audible and visual alarm and dynamically adjust production parameters when a serious defect occurs.

6. The method for real-time monitoring of electroplated metal quality based on deep learning as described in claim 5, characterized in that, In step 2, some self-attention modules adaptively highlight areas that may have defects in the feature map through an attention mechanism, thereby improving the detectability of minor flaws.

Images