An industrial visual recognition method and system based on deep learning

Abstract

This invention belongs to the field of quality inspection technology and discloses a deep learning-based industrial visual recognition method and system. The method includes: constructing a high-quality welding image sequence through multi-angle image acquisition, multi-scale noise reduction optimization, and weld feature enhancement processing; using adaptive feature extraction and cross-scale feature fusion technology to generate a fused welding feature tensor; utilizing attention mechanism enhancement and automatic segmentation technology to achieve accurate extraction of weld area and enhanced identification of microscopic defects; combining historical welding quality assessment data to conduct dynamic feature correlation analysis and construct a quality assessment model; establishing a welding quality diagnosis strategy through multi-dimensional defect type identification and adaptive quality grade assessment; and finally achieving dynamic quality prediction and real-time monitoring of weld defects. The present invention achieves accurate assessment and real-time monitoring of welding quality, effectively reducing quality costs and rework rates.

Description

Technical Field

[0001] The present invention relates to the field of quality inspection technology, and more specifically, to an industrial visual recognition method and system based on deep learning. Background Art

[0002] With the rapid development of the electronics manufacturing industry and the increasing trend toward product miniaturization, wire bonding, a critical process for connecting electronic components, is becoming increasingly crucial for quality control, impacting product reliability and performance. Traditional wire bonding quality inspection relies primarily on manual visual inspection, which is inefficient and significantly influenced by the inspector's experience and subjective judgment. This makes it difficult to meet the high-precision, high-consistency quality control requirements of the modern electronics manufacturing industry.

[0003] In recent years, machine vision technology has been widely used in industrial inspection, with automated welding quality inspection systems based on image processing gradually replacing manual inspection. Currently, common welding visual inspection systems on the market primarily use traditional image processing algorithms, such as edge detection, morphological analysis, and threshold segmentation, to identify and evaluate welds. While these systems can achieve basic defect detection under ideal conditions, they face numerous challenges in real-world production environments.

[0004] Traditional visual inspection systems perform poorly in complex industrial environments, especially when faced with actual working conditions such as lighting changes, surface reflections, and environmental interference. Detection accuracy drops significantly, resulting in a large number of minor defects being missed. On high-speed production lines, the inspection system is unable to adapt in real time to subtle changes in the shape and position of solder joints, resulting in a high rate of misjudgment and necessitating reliance on manual re-inspection, which significantly increases labor costs and production cycles. Particularly for the micro-solder joints of high-density electronic components, existing technologies have low accuracy when identifying key quality issues such as false solder joints, cold solder joints, and insufficient solder joints. This allows a large number of potential quality risks to flow into downstream processes and even end users, causing product reliability risks and brand reputation losses. In addition, existing systems lack the ability to deeply mine and learn from historical data, and are unable to establish correlation analysis between defects and process parameters, resulting in frequent and recurring quality issues that cannot be resolved.

[0005] In view of this, the present invention proposes an industrial visual recognition method and system based on deep learning to solve the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned objectives, the present invention provides the following technical solution: an industrial visual recognition method based on deep learning, comprising:

[0007] Step S1: Acquire a core wire welding multi-angle image dataset; perform multi-scale noise reduction optimization on the core wire welding multi-angle image dataset and perform weld feature enhancement processing to obtain an enhanced core wire welding image sequence;

[0008] Step S2: performing adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; performing cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor;

[0009] Step S3: Automatically segmenting the weld area of ​​the fused welding feature tensor to obtain a weld area feature matrix; performing microscopic defect enhancement recognition on the weld area feature matrix to obtain a weld defect feature space;

[0010] Step S4: Acquire historical welding quality assessment data; perform dynamic feature correlation analysis on the historical welding quality assessment data to construct a welding quality assessment model;

[0011] Step S5: Using the welding quality assessment model to perform multi-dimensional defect type identification on the weld defect feature space, and perform adaptive quality grade assessment to construct a welding quality diagnosis strategy;

[0012] Step S6: Based on the welding quality diagnosis strategy, dynamic quality prediction is performed on the weld defect feature space, and a real-time monitoring model for core wire welding quality is constructed.

[0013] Preferably, the specific steps of step S1 are:

[0014] Step S11: Acquire a core wire welding multi-angle image dataset;

[0015] Step S12: performing illumination imbalance correction on the core wire welding multi-angle image data set to obtain illumination balanced image data;

[0016] Step S13: performing multi-scale wavelet transform noise reduction on the illumination-equalized image data to identify image noise interference areas;

[0017] Step S14: performing adaptive filtering on the illumination-equalized image data according to the noise interference area to obtain noise-reduced image data;

[0018] Step S15: performing solder joint edge enhancement processing on the noise reduction image data to generate edge enhanced image data;

[0019] Step S16: performing contrast adaptive adjustment processing on the edge enhanced image data to obtain an enhanced core wire welding image sequence.

[0020] Preferably, the specific steps of step S15 are:

[0021] Perform multi-directional gradient calculation on the denoised image data to obtain the edge gradient features of the solder joint;

[0022] Extract edge candidate regions based on the edge gradient features of solder joints to obtain solder joint edge candidate regions;

[0023] Perform edge continuity analysis on the candidate areas of the solder joint edge and generate edge continuity scores;

[0024] Performing parametric edge enhancement on the denoised image data based on the edge continuity score to generate an edge sharpening parameter range;

[0025] Adaptively sharpen the denoised image data based on the edge sharpening parameter range and mark key edge feature points;

[0026] Nonlinear edge enhancement optimization is performed on key edge feature points to generate edge-enhanced image data.

[0027] Preferably, the specific steps of step S2 are:

[0028] Step S21: performing multi-scale convolutional network processing on the enhanced core wire welding image sequence to generate feature maps of different scales;

[0029] Step S22: performing solder joint feature recognition analysis on feature images of different scales to obtain a multi-level solder joint feature representation;

[0030] Step S23: performing adaptive feature channel weighting based on the multi-level weld feature representation to generate a multi-level welding feature map;

[0031] Step S24: performing feature similarity calculation on the multi-level welding feature maps to obtain a feature similarity matrix;

[0032] Step S25: normalizing the feature level differences of the multi-level welding feature map to generate a standardized multi-level feature map;

[0033] Step S26: performing cross-scale feature fusion on the standardized multi-level feature map based on the feature similarity matrix to generate a fused welding feature tensor.

[0034] Preferably, the specific steps of step S3 are:

[0035] Step S31: performing attention mechanism enhancement on the fused welding feature tensor to obtain a weld spot attention area tensor;

[0036] Step S32: Automatically segmenting the solder joint region tensor to obtain a solder joint region feature matrix;

[0037] Step S33: performing high-frequency detail enhancement processing on the solder joint area feature matrix to obtain a microscopic detail enhancement feature matrix;

[0038] Step S34: performing micro-defect enhancement identification on the micro-detail enhancement feature matrix to obtain a solder joint defect feature space.

[0039] Preferably, the specific steps of step S4 are:

[0040] Step S41: Acquire historical welding quality assessment data and corresponding image annotation data;

[0041] Step S42: performing defect type distribution analysis on historical welding quality assessment data to generate a defect type distribution map;

[0042] Step S43: performing dynamic feature correlation analysis on the defect type distribution map to generate welding quality assessment features;

[0043] Step S44: Perform multi-level neural network training on the welding quality assessment features to construct a welding quality assessment model.

[0044] Preferably, the specific steps of step S43 are:

[0045] Extract multiple defect characteristic parameters based on defect type distribution map;

[0046] Calculate the defect severity of historical welding quality assessment data based on multiple defect characteristic parameters to obtain a defect severity index;

[0047] Perform defect space distribution analysis on the defect type distribution map to obtain defect space distribution characteristics;

[0048] Conduct correlation analysis on the defect spatial distribution characteristics based on the defect severity index to obtain defect correlation feature data;

[0049] Identify factors affecting welding process parameters based on defect type distribution maps;

[0050] Perform defect cause correlation analysis based on the process parameter influencing factors to obtain process parameter correlation data;

[0051] Dynamic feature fusion analysis is performed on process parameter correlation data and defect correlation feature data to generate welding quality assessment features.

[0052] Preferably, the specific steps of step S5 are:

[0053] Step S51: using the welding quality assessment model to perform multi-category defect recognition on the weld defect feature space and extract potential defect feature data;

[0054] Step S52: performing defect morphology analysis on the solder joint defect feature space to obtain a defect morphology feature link;

[0055] Step S53: performing multi-dimensional defect type identification on the potential defect feature data based on the defect morphology feature link to obtain a defect type identification result;

[0056] Step S54: quantifying the degree of quality impact of the defect type identification result to generate a welding quality assessment value;

[0057] Step S55: performing adaptive quality grade evaluation based on the welding quality evaluation value and constructing a welding quality diagnosis strategy.

[0058] Preferably, the specific steps of step S6 are:

[0059] Step S61: constructing a quality prediction model for the weld defect feature space to generate a welding quality prediction framework;

[0060] Step S62: mapping the welding quality prediction framework to diagnostic rules based on the welding quality diagnosis strategy to construct a core wire welding quality diagnosis framework;

[0061] Step S63: Perform dynamic quality prediction on the core wire welding quality diagnosis framework and construct a real-time monitoring model for the core wire welding quality.

[0062] A deep learning-based industrial visual recognition system, which is used to implement the deep learning-based industrial visual recognition method, includes:

[0063] A processing module is used to obtain a core wire welding multi-angle image dataset; perform multi-scale noise reduction optimization on the core wire welding multi-angle image dataset, and perform weld point feature enhancement processing to obtain an enhanced core wire welding image sequence;

[0064] A deep hierarchical module is used to perform adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; and to perform cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor;

[0065] The regional space module is used to automatically segment the weld area of ​​the fused welding feature tensor to obtain the weld area feature matrix; and to enhance the recognition of microscopic defects on the weld area feature matrix to obtain the weld defect feature space;

[0066] Model building module, used to obtain historical welding quality assessment data; perform dynamic feature correlation analysis on historical welding quality assessment data to build a welding quality assessment model;

[0067] Strategy fitting module, used to use welding quality assessment model to identify multi-dimensional defect types in weld defect feature space, and to perform adaptive quality grade assessment to build welding quality diagnosis strategy;

[0068] The integrated module performs dynamic quality prediction on the weld defect feature space based on the welding quality diagnosis strategy and builds a real-time monitoring model for core wire welding quality; each module is connected via wired and / or wireless means.

[0069] The technical effects and advantages of the industrial visual recognition method and system based on deep learning of the present invention are as follows:

[0070] The beneficial effect of the present invention is that it significantly improves the overall performance and practical value of core wire welding quality detection. It solves the pain points in traditional detection, greatly improves the accuracy and sensitivity of solder joint defect identification, and especially achieves a qualitative leap in the ability to detect microscopic and subtle defects. It realizes a comprehensive and multi-angle quality assessment system, which can not only accurately distinguish different types of defects, but also quantitatively assess their impact on product quality. Thanks to the application of adaptive technology, it has strong adaptability and robustness to various complex working conditions and abnormal situations, effectively reducing the misjudgment rate and missed detection rate. The real-time monitoring and early warning mechanism enables the production line to detect and respond to potential quality risks in a timely manner, prevent the generation of defective products from the source, and significantly reduce quality costs and rework rates. In addition, by exploring the intrinsic connection between defects and process parameters, it provides data support and decision-making basis for process optimization and continuous quality improvement, and ultimately achieves a dual improvement in production efficiency and product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of an industrial visual recognition method based on deep learning according to the present invention;

[0072] Figure 2 Schematic diagram of an industrial visual recognition system based on deep learning of the present invention. DETAILED DESCRIPTION

[0073] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0074] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In the embodiments of the present application, an embodiment of the industrial visual recognition method based on deep learning includes:

[0075] Step S1: Acquire a core wire welding multi-angle image dataset; perform multi-scale noise reduction optimization on the core wire welding multi-angle image dataset and perform weld feature enhancement processing to obtain an enhanced core wire welding image sequence;

[0076] Step S2: performing adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; performing cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor;

[0077] Step S3: Automatically segmenting the weld area of ​​the fused welding feature tensor to obtain a weld area feature matrix; performing microscopic defect enhancement recognition on the weld area feature matrix to obtain a weld defect feature space;

[0078] Step S4: Acquire historical welding quality assessment data; perform dynamic feature correlation analysis on the historical welding quality assessment data to construct a welding quality assessment model;

[0079] Step S5: Using the welding quality assessment model to perform multi-dimensional defect type identification on the weld defect feature space, and perform adaptive quality grade assessment to construct a welding quality diagnosis strategy;

[0080] Step S6: Based on the welding quality diagnosis strategy, dynamic quality prediction is performed on the weld defect feature space, and a real-time monitoring model for core wire welding quality is constructed.

[0081] It is understandable that the execution subject of this application can be an industrial visual recognition system based on deep learning, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0082] Specifically, a multi-angle image dataset of core wire welding is acquired, including core wire welding images captured from different angles. Illumination imbalance correction is performed on the multi-angle image dataset to generate illumination-balanced image data. Illumination imbalance correction uses methods such as histogram equalization or adaptive gamma correction to eliminate image quality issues caused by uneven illumination. Multi-scale wavelet transform is then used to reduce noise on the illumination-balanced image data to identify areas of image noise interference. Multi-scale wavelet transform can decompose an image into different scales and identify and process noise components at each scale. Adaptive filtering is then performed on the illumination-balanced image data based on the areas of image noise interference to generate de-noised image data. Adaptive filtering dynamically adjusts filter parameters based on noise characteristics, effectively removing noise while preserving image detail. Weld spot edge enhancement is then performed on the de-noised image data to generate edge-enhanced image data. This weld spot edge enhancement improves the clarity and recognizability of weld spot edges through multi-directional gradient calculation and edge continuity analysis. Adaptive contrast adjustment is then performed on the edge-enhanced image data to generate an enhanced core wire welding image sequence. Adaptive contrast adjustment dynamically adjusts contrast parameters based on image content to improve the distinction between the solder joint area and the background.

[0083] A multi-scale convolutional network is applied to the enhanced core wire welding image sequence to generate feature maps at different scales. The multi-scale convolutional network captures the multi-scale feature information of the weld joint using convolution kernels of different sizes. Weld joint feature recognition and analysis are performed on the feature maps at different scales to obtain a multi-level weld joint feature representation. This recognition and analysis uses a deep learning model to extract features such as weld joint shape, texture, and structure. Adaptive feature channel weighting is performed based on the multi-level weld joint feature representation to generate a multi-level weld feature map. Feature channel weighting assigns different weights to each channel based on its importance to weld joint recognition, highlighting key features. Feature similarity is calculated on the multi-level weld feature map to generate a feature similarity matrix. Feature similarity calculation assesses the correlation between features at different levels, providing a basis for subsequent fusion. Feature level differences are normalized on the multi-level weld feature map to generate a standardized multi-level feature map. Feature level difference normalization eliminates scale differences between features at different levels, making feature fusion more effective. Cross-scale feature fusion is performed on the standardized multi-level feature map based on the feature similarity matrix to generate a fused weld feature tensor. Cross-scale feature fusion integrates multi-level feature information to form a more comprehensive welding feature representation.

[0084] The fused weld feature tensor is enhanced with an attention mechanism to obtain a solder joint attention region tensor. This attention mechanism enhances the model by learning the importance weights of the solder joint regions, guiding it to focus on key areas of the solder joint. The solder joint attention region tensor is automatically segmented to obtain a solder joint region feature matrix. Automatic solder joint region segmentation uses a semantic segmentation network to accurately locate the solder joint regions and eliminate background interference. High-frequency detail enhancement is performed on the solder joint region feature matrix to obtain a micro-detail enhancement feature matrix. High-frequency detail enhancement highlights the microstructural features of the solder joint, facilitating subsequent defect identification. Micro-defect enhancement recognition is performed on the micro-detail enhancement feature matrix to obtain a solder joint defect feature space. Micro-defect enhancement recognition uses a feature enhancement network to highlight potential defect features, forming a defect feature space representation.

[0085] Obtain historical welding quality assessment data and corresponding image annotation data, including expert assessment results and defect annotation information. Perform defect type distribution analysis on the historical welding quality assessment data to generate a defect type distribution map. This defect type distribution analysis statistically analyzes the frequency and distribution characteristics of various defects to form a defect distribution map. Perform dynamic feature association analysis on the defect type distribution map to generate welding quality assessment features. Dynamic feature association analysis explores the correlation between defect features and welding quality, extracting key features for quality assessment. Perform multi-level neural network training on the welding quality assessment features to construct a welding quality assessment model. This multi-level neural network training uses a deep learning algorithm to establish a mapping between defect features and quality assessment, forming a quality assessment model.

[0086] A welding quality assessment model is used to perform multi-category defect recognition on the weld defect feature space and extract potential defect feature data. Multi-category defect recognition uses a trained deep learning model to identify various weld defects, including porosity, cracks, and lack of fusion. Defect morphology analysis is performed on the weld defect feature space to generate a defect morphology feature chain. Defect morphology analysis studies the geometry, size, and distribution of defects to construct a defect morphology feature chain. Based on the defect morphology feature chain, multi-dimensional defect type recognition is performed on the potential defect feature data to generate defect type recognition results. Multi-dimensional defect type recognition comprehensively considers factors such as defect morphology, location, and severity to accurately determine the defect type. The defect type recognition results are quantified to determine the degree of quality impact and generate a welding quality assessment value. This quality impact quantification calculates a comprehensive quality assessment value based on the weight of the impact of different defect types on welding quality. Based on the welding quality assessment value, adaptive quality grade assessment is performed to develop a welding quality diagnosis strategy. Adaptive quality grade assessment maps quality assessment values ​​to different quality grades based on industry standards and specific application requirements, and formulates a corresponding diagnosis strategy.

[0087] A quality prediction model is constructed for the weld spot defect feature space to generate a welding quality prediction framework. This quality prediction model uses a time-series deep learning network to establish a mapping relationship between defect characteristics and quality change trends. Based on the welding quality diagnosis strategy, diagnostic rules are mapped to the welding quality prediction framework to construct a core wire welding quality diagnosis framework. This diagnostic rule mapping transforms the quality diagnosis strategy into an executable rule set to guide real-time quality monitoring. Dynamic quality prediction is performed on the core wire welding quality diagnosis framework to construct a real-time core wire welding quality monitoring model. Dynamic quality prediction continuously analyzes changes in defect characteristics during the welding process, enabling real-time monitoring and early warning of welding quality.

[0088] In the embodiment of the present application, the quality of the original image is effectively improved through multi-scale noise reduction optimization and weld feature enhancement processing, laying the foundation for subsequent analysis. Through adaptive feature extraction and cross-scale feature fusion, the comprehensive capture and effective integration of weld features are achieved. Through automatic segmentation of weld areas and enhanced identification of microscopic defects, the weld areas are accurately located and the potential defect features are highlighted. Through dynamic feature association analysis and multi-level neural network training, an accurate and reliable welding quality assessment model is constructed. Through multi-dimensional defect type identification and adaptive quality grade assessment, accurate diagnosis and grading of welding quality are achieved. Through dynamic quality prediction and real-time monitoring model construction, continuous monitoring and early warning of welding quality are achieved, improving production efficiency and product quality.

[0089] In a specific embodiment, the process of executing step S1 specifically includes the following steps:

[0090] Step S11: Acquire a core wire welding multi-angle image dataset;

[0091] Step S12: performing illumination imbalance correction on the core wire welding multi-angle image data set to obtain illumination balanced image data;

[0092] Step S13: performing multi-scale wavelet transform noise reduction on the illumination-equalized image data to identify image noise interference areas;

[0093] Step S14: performing adaptive filtering on the illumination-equalized image data according to the noise interference area to obtain noise-reduced image data;

[0094] Step S15: performing solder joint edge enhancement processing on the noise reduction image data to generate edge enhanced image data;

[0095] Step S16: performing contrast adaptive adjustment processing on the edge enhanced image data to obtain an enhanced core wire welding image sequence.

[0096] Specifically, a multi-angle image dataset of core wire welding is acquired. An industrial camera array is used to capture core wire welding images from different angles, ensuring full capture of weld features. Multi-angle acquisition helps reduce occlusion issues associated with a single viewpoint and improves the integrity of feature extraction. Lighting imbalance correction is performed on the multi-angle image dataset of core wire welding to obtain illumination-balanced image data. This illumination imbalance correction uses histogram equalization technology to adjust the image pixel intensity distribution and improve image contrast. For example, for image I(x, y), its histogram equalization representation is:

[0097] I_eq(x, y) = T[I(x, y)]; where T[] is the transformation function, and x and y are the coordinate indices of the pixel, mapping the grayscale values ​​of the original image to the new grayscale value range. Multiscale wavelet transform is performed on the illumination-balanced image data to reduce noise and identify areas of image noise interference. Multiscale wavelet transform decomposes the image into different frequency components, which can be expressed as:

[0098] Where D_j(x, y) represents the detail component at scale j, and A_J(x, y) represents the approximate component at scale j. By thresholding the wavelet coefficients at each scale, noise can be effectively identified and suppressed. Adaptive filtering is performed on the illumination-equalized image data according to the noise interference area to obtain the denoised image data. Adaptive filtering dynamically adjusts the filter parameters according to the local noise characteristics, which is expressed as:

[0099] I_d(x, y) = I_e(x, y) × h(x, y, σ(x, y)); where h() is the filter kernel function, and σ(x, y) is a parameter adaptively adjusted based on the local noise level. Weld edge enhancement is performed on the denoised image data to generate edge-enhanced image data. This process includes multi-directional gradient calculation, edge candidate region extraction, and edge continuity analysis, effectively improving the clarity of weld edges. Contrast adaptive adjustment is performed on the edge-enhanced image data to produce an enhanced core wire welding image sequence. Contrast adaptive adjustment dynamically adjusts the contrast parameter based on the image content, expressed as:

[0100] I_e(x, y) = α(x, y) × I_b(x, y) + β(x, y); where α(x, y) and β(x, y) are gain and bias parameters adaptively adjusted based on local image features, I_b(x, y) is the edge-enhanced image, and I_e(x, y) is the final enhanced image. This series of processing yields a significantly improved sequence of enhanced core wire welding images, providing a high-quality data foundation for subsequent analysis.

[0101] In a specific embodiment, the process of executing step S15 specifically includes the following steps:

[0102] Perform multi-directional gradient calculation on the denoised image data to obtain the edge gradient features of the solder joint;

[0103] Extract edge candidate regions based on the edge gradient features of solder joints to obtain solder joint edge candidate regions;

[0104] Perform edge continuity analysis on the candidate areas of the solder joint edge and generate edge continuity scores;

[0105] Performing parametric edge enhancement on the denoised image data based on the edge continuity score to generate an edge sharpening parameter range;

[0106] Adaptively sharpen the denoised image data based on the edge sharpening parameter range and mark key edge feature points;

[0107] Nonlinear edge enhancement optimization is performed on key edge feature points to generate edge-enhanced image data.

[0108] Specifically, multi-directional gradient calculation is performed on the denoised image data to obtain the edge gradient features of the solder joint. Multi-directional gradient calculation calculates image gradients in different directions using operators such as Sobel, Prewitt, or Canny. Through multi-directional gradient calculation, the directional features of the solder joint edge can be fully captured. Edge candidate regions are extracted based on the solder joint edge gradient features to obtain solder joint edge candidate regions. Edge candidate region extraction is performed through gradient amplitude threshold processing, and regions with gradient amplitudes greater than the threshold are selected as edge candidate regions. Edge continuity analysis is performed on the edge candidate regions of the solder joint to generate an edge continuity score. Edge continuity analysis evaluates the degree of edge continuity by calculating the connectivity and curvature characteristics of the edge candidate regions, which is expressed as:

[0109] C(i)=f(L(i),K(i)); where L(i) is the length of the i-th edge region, K(i) is its curvature characteristic, f() is the scoring function, and C(i) is the continuity score of the i-th edge region.

[0110] ; Where L(i) is the length of the i-th edge region, L_max is the maximum length of all edge regions, K_var(i) is the degree of curvature change (curvature variance or volatility) of the i-th edge region, K_max is the maximum curvature change of all edge regions, w1 and w2 are weight coefficients, usually w1+w2=1.

[0111] Based on the edge continuity score, the denoised image data is subjected to parametric edge enhancement processing to generate an edge sharpening parameter range. The parametric edge enhancement processing adaptively determines the enhancement parameter range according to the continuity score, which is expressed as:

[0112] P_range(i)=[P_min+α×C(i), P_max-β×(1-C(i))]; where P_range(i) is the sharpening parameter range of the i-th edge region, α and β are adjustment parameters, and the preset range interval is [P_min, P_max]. Based on the edge sharpening parameter range, the denoised image data is adaptively sharpened and key edge feature points are marked. Adaptive sharpening performs non-uniform sharpening on the image according to the sharpening parameter range to highlight key edge features, which can be expressed as:

[0113] I_s(x, y) = I_d(x, y) + λ(x, y) × (I_d(x, y) - I_d(x, y) × h_blur(x, y)); where I_s(x, y) is the image after adaptive sharpening, h_blur is the blur kernel, and λ(x, y) is the local sharpening strength determined according to the sharpening parameter range, which varies with position. The entire formula represents adding the extracted edge details back to the original image according to the local sharpening strength, thereby enhancing the edge features. Nonlinear edge enhancement optimization is performed on the key edge feature points to generate edge-enhanced image data. Nonlinear edge enhancement optimization further enhances edge features through a nonlinear mapping function, expressed as:

[0114] I_b(x, y)=g(I_s(x, y)); where g() is a nonlinear mapping function (such as a Sigmoid function), and I_b(x, y) is the final edge-enhanced image data.

[0115] In a specific embodiment, the process of executing step S2 may specifically include the following steps:

[0116] Step S21: performing multi-scale convolutional network processing on the enhanced core wire welding image sequence to generate feature maps of different scales;

[0117] Step S22: performing solder joint feature recognition analysis on feature images of different scales to obtain a multi-level solder joint feature representation;

[0118] Step S23: performing adaptive feature channel weighting based on the multi-level weld feature representation to generate a multi-level welding feature map;

[0119] Step S24: performing feature similarity calculation on the multi-level welding feature maps to obtain a feature similarity matrix;

[0120] Step S25: normalizing the feature level differences of the multi-level welding feature map to generate a standardized multi-level feature map;

[0121] Step S26: performing cross-scale feature fusion on the standardized multi-level feature map based on the feature similarity matrix to generate a fused welding feature tensor.

[0122] Specifically, the enhanced core wire welding image sequence is processed by a multi-scale convolutional network to generate feature maps of different scales. The multi-scale convolutional network captures the multi-scale feature information of the weld through convolution kernels of different sizes and pooling operations, which can be expressed as:

[0123] F_l=CNN_l(I_e); where CNN_l represents the lth layer of the convolutional network, and F_l is the generated feature map. Through multi-layer convolutional networks, multi-scale feature representations from low-level textures to high-level semantics can be obtained. Solder joint feature recognition and analysis are performed on feature maps of different scales to obtain a multi-level solder joint feature representation. Solder joint feature recognition and analysis uses a feature enhancement network to extract features such as the shape, texture, and structure of the solder joint, which can be expressed as: F_l_e=FEN(F_l); where FEN is the feature enhancement network, which is used to strengthen and refine the features extracted by the original convolutional network. It includes a residual connection module (to enhance feature propagation), an attention module (to highlight important areas), dilated convolution (to expand the receptive field), and feature recalibration (to adjust feature distribution); F_l_e is the enhanced feature representation. Based on the multi-level solder joint feature representation, adaptive feature channel weighting is performed to generate a multi-level welding feature map. Feature channel weighting assigns different weights to feature channels through the attention mechanism to highlight key features, which is expressed as:

[0124] F_l_w=F_l_e×W_l; where W_l is the channel weight learned through the attention mechanism, a vector whose elements correspond to the importance weight of a feature channel, and F_l_w is the weighted feature map of the lth layer. Feature similarity calculation is performed on the multi-level welding feature map to obtain a feature similarity matrix. Feature similarity calculation evaluates the degree of correlation between features at different levels and is measured using cosine similarity or Pearson correlation coefficient. Feature level difference normalization is performed on the multi-level welding feature map to generate a standardized multi-level feature map. Feature level difference normalization eliminates scale differences between features at different levels and is expressed as:

[0125] Where μ_l and σ_l are the mean and standard deviation of the l-th layer features, respectively, and F_l_n is the standardized feature map, which eliminates the scale differences between features at different levels. Based on the feature similarity matrix, cross-scale feature fusion is performed on the standardized multi-level feature map to generate a fused welding feature tensor. Cross-scale feature fusion integrates multi-level feature information and is expressed as:

[0126] ; where w_l is the fusion weight calculated based on the similarity matrix, which determines the importance of the l-th layer feature in the final fusion result, up() is the upsampling operation to align feature maps of different scales, and F_fusion is the final fusion feature tensor.

[0127] In a specific embodiment, the process of executing step S3 may specifically include the following steps:

[0128] Step S31: performing attention mechanism enhancement on the fused welding feature tensor to obtain a weld spot attention area tensor;

[0129] Step S32: Automatically segmenting the solder joint region tensor to obtain a solder joint region feature matrix;

[0130] Step S33: performing high-frequency detail enhancement processing on the solder joint area feature matrix to obtain a microscopic detail enhancement feature matrix;

[0131] Step S34: performing micro-defect enhancement identification on the micro-detail enhancement feature matrix to obtain a solder joint defect feature space.

[0132] Specifically, the attention mechanism is enhanced on the fused welding feature tensor to obtain the solder joint attention area tensor. The attention mechanism enhancement combines spatial attention and channel attention to highlight the key areas of the solder joint, which can be expressed as:

[0133] A_spatial=σ(f_s(F_fusion));

[0134] A_channel=σ(f_c(F_fusion));

[0135] F_att=F_fusion⊙A_spatial⊙A_channel; where A_spatial is the spatial attention map, representing the importance weight of each spatial position on the feature map; f_s is the spatial attention calculation function, which usually calculates the importance of each position on the feature map through convolution operations; A_channel is the channel attention vector, representing the importance weight of each feature channel; f_c is the channel attention calculation function, which usually calculates the importance of each feature channel through global pooling and fully connected layers; σ is the activation function, usually a Sigmoid function, which maps values ​​to the range of 0-1; F_att is the feature tensor after attention enhancement. The solder joint area tensor is automatically segmented to obtain the solder joint area feature matrix. Automatic solder joint area segmentation accurately locates the solder joint area through a semantic segmentation network, expressed as:

[0136] M=Seg(F_att);

[0137] F_roi = F_att⊙M; where Seg is a segmentation network, such as U-Net or DeepLab, used to segment the image into different semantic regions. M is the segmentation mask, a binary image where regions with a value of 1 represent solder joints and regions with a value of 0 represent background. ⊙ represents element-wise multiplication. F_roi is the solder joint region feature matrix, which retains only the features of the solder joint region and eliminates background interference. High-frequency detail enhancement is performed on the solder joint region feature matrix to obtain a micro-detail enhancement feature matrix. High-frequency detail enhancement highlights the microstructural features of the solder joint, expressed as:

[0138] F_detail = F_roi + λ1 × HF(F_roi); where HF is a high-frequency extraction operation, which can be a high-pass filter, Laplace operator, or wavelet transform, used to extract high-frequency details in the image. λ1 is the enhancement coefficient, which controls the intensity of high-frequency detail enhancement. F_detail is the feature matrix after detail enhancement, highlighting the microstructural features of the solder joint. Microdefect enhancement recognition is performed on the microdetail enhancement feature matrix to obtain the solder joint defect feature space. Microdefect enhancement recognition highlights potential defect features through a feature enhancement network, expressed as:

[0139] F_defect = DefectNet(F_detail); DefectNet is a defect enhancement network, a neural network used to identify and enhance solder joint defect features. It includes a residual module (to improve feature extraction capabilities), an attention mechanism (to focus on potential defect areas), multi-scale feature fusion (to capture defect features at different scales), and a contrastive learning module (to enhance the difference between defects and normal areas). F_defect is the final solder joint defect feature space and the feature representation ultimately used for defect analysis and quality assessment.

[0140] In a specific embodiment, the process of executing step S4 specifically includes the following steps:

[0141] Step S41: Acquire historical welding quality assessment data and corresponding image annotation data;

[0142] Step S42: performing defect type distribution analysis on historical welding quality assessment data to generate a defect type distribution map;

[0143] Step S43: performing dynamic feature correlation analysis on the defect type distribution map to generate welding quality assessment features;

[0144] Step S44: Perform multi-level neural network training on the welding quality assessment features to construct a welding quality assessment model.

[0145] Specifically, historical welding quality assessment data and corresponding image annotation data are obtained. The historical data includes expert assessment results, defect annotation information, and related process parameters, providing annotated samples for model training. Defect type distribution analysis is performed on the historical welding quality assessment data to generate a defect type distribution map. The defect type distribution analysis statistics the occurrence frequency and distribution characteristics of various defects, expressed as:

[0146] ;

[0147] ; Wherein, defect_i is the i-th type of defect, such as pores, cracks, lack of fusion and other specific defect types, count(defect_i) is the number of times the i-th type of defect appears in the historical data, total is the total number of samples in the historical data, D(i) is the frequency distribution of the i-th type of defect, indicating the probability of this type of defect appearing in all samples, P(i, j) is the conditional probability distribution of the i-th type of defect at the j-th position, indicating the probability of this type of defect appearing in a specific position, location_j is the j-th location area, such as the center, edge, heat-affected zone of the weld, etc., count(defect_i, location_j) is the number of times the i-th type of defect appears in the j-th location area. Dynamic feature association analysis is performed on the defect type distribution map to generate welding quality assessment features. Dynamic feature association analysis explores the correlation between defect features and welding quality, which is expressed as:

[0148] R(i, q)=corr(defect_i, quality_q);

[0149] F_quality=fl({R(i,q)},{D(i)},{P(i,j)}); where R(i,q) is the correlation between the i-th defect and the quality indicator q, indicating the degree of impact of the defect on quality, quality_q is the q-th quality indicator, such as strength, durability, conductivity, etc.; corr() is a correlation calculation function, such as the Pearson correlation coefficient or the Spearman rank correlation coefficient; F_quality is the extracted quality assessment feature, which integrates information on defect type, distribution, and quality impact; {R(i,q)} is the set of correlation coefficients between all defect types and quality indicators; {D(i)} is the set of frequency distributions of all defect types; {P(i,j)} is the set of conditional probability distributions of all defect types at each location; and fl() is a feature fusion function, which may be a weighted combination, nonlinear transformation, or neural network. Multi-level neural network training is performed on the welding quality assessment features to construct a welding quality assessment model. Multi-level neural network training uses a deep learning algorithm to establish a mapping relationship between defect features and quality assessment, which can be expressed as:

[0150] Q=DNN(F_quality); DNN is a deep neural network containing multiple hidden layers, convolutional layers, and fully connected layers, which is used to learn the complex mapping relationship between defect characteristics and quality assessment. Q is the predicted quality assessment result, which is the quality grade, pass rate, or specific quality score.

[0151] In a specific embodiment, the process of executing step S43 may specifically include the following steps:

[0152] Extract multiple defect characteristic parameters based on defect type distribution map;

[0153] Calculate the defect severity of historical welding quality assessment data based on multiple defect characteristic parameters to obtain a defect severity index;

[0154] Perform defect space distribution analysis on the defect type distribution map to obtain defect space distribution characteristics;

[0155] Conduct correlation analysis on the defect spatial distribution characteristics based on the defect severity index to obtain defect correlation feature data;

[0156] Identify factors affecting welding process parameters based on defect type distribution maps;

[0157] Perform defect cause correlation analysis based on the process parameter influencing factors to obtain process parameter correlation data;

[0158] Dynamic feature fusion analysis is performed on process parameter correlation data and defect correlation feature data to generate welding quality assessment features.

[0159] Specifically, a variety of defect characteristic parameters are extracted based on the defect type distribution map. The defect characteristic parameters include the size, shape, density, and distribution of the defect, which can be expressed as:

[0160] P_size(i)=avg(defect_i);

[0161] P_shape(i)=shape(defect_i);

[0162] P_density(i)=density(defect_i); where P_size(i), P_shape(i), and P_density(i) represent the size, shape, and density parameters of the i-th defect, respectively. avg() is the average size calculation function, based on area, diameter, or length. Shape() is the shape factor calculation function, which can be circularity, aspect ratio, or complexity. Density() is the density calculation function, representing the number of defects per unit area. The defect severity is calculated based on a variety of defect characteristic parameters for historical welding quality assessment data to obtain a defect severity index. The defect severity calculation comprehensively considers the characteristic parameters and location of the defect and is expressed as:

[0163] SL(i)=w_s×P_size(i)+w_p×P_shape(i)+w_d×P_density(i)+w_loc×P_loc(i); where w_s, w_p, w_d, and w_loc are the weights of each parameter, P_loc(i) is the location parameter of the i-th type defect, reflecting the importance of the defect location, and SL(i) is the severity index of the i-th type defect. Defect spatial distribution analysis is performed on the defect type distribution map to obtain the defect spatial distribution characteristics. Defect spatial distribution analysis studies the spatial distribution pattern of defects in the solder joint area, which is expressed as:

[0164] D_spatial(x,y)=∑ i D(i)×P(i, x, y);

[0165] Where D_spatial(x, y) represents the defect distribution density at position (x, y), indicating the comprehensive probability of a defect occurring at that position, and P(i, x, y) is the probability of the i-th type defect at position (x, y). Correlation analysis is performed on the defect spatial distribution characteristics based on the defect severity index to obtain defect correlation feature data. Correlation analysis studies the relationship between defect severity and spatial distribution, which is expressed as:

[0166] CL(i,j)=corr(SL(i),D_spatial_j);

[0167] F_corr={CL(i, j)}; where CL(i, j) is the correlation between the severity of the i-th defect and the j-th regional distribution, D_spatial_j is the defect distribution density of the j-th region, corr() is a correlation calculation function, such as the Pearson correlation coefficient, and F_corr is defect correlation feature data, which contains the correlation information between all defect types and regional distributions. Identify the influencing factors of welding process parameters based on the defect type distribution map. Process parameter influencing factors include welding current, voltage, speed, and shielding gas, etc. By analyzing the relationship between these parameters and defect distribution, the key influencing factors are identified. Defect cause correlation analysis is performed based on the process parameter influencing factors to obtain process parameter correlation data. Defect cause correlation analysis studies the causal relationship between process parameters and defect formation, which is expressed as:

[0168] R_param(s,j)=corr(param_s,defect_j);

[0169] F_param = {R_param(s, j)}; where R_param(s, j) is the correlation between the sth process parameter and the jth type of defect, param_s is the sth process parameter, such as welding current, voltage, speed, shielding gas, etc., and F_param is the process parameter association data, which contains the correlation information between all process parameters and defect types. Dynamic feature fusion analysis is performed on the process parameter association data and defect association feature data to generate welding quality assessment features. Dynamic feature fusion analysis integrates the association information between process parameters and defect features and can be expressed as:

[0170] F_quality=Fusion(F_corr, F_param); Fusion is a feature fusion function, which can be a weighted combination, nonlinear transformation, or neural network, and F_quality is the final welding quality assessment feature, which integrates the correlation information between defect characteristics and process parameters.

[0171] In a specific embodiment, the process of executing step S5 may specifically include the following steps:

[0172] Step S51: using the welding quality assessment model to perform multi-category defect recognition on the weld defect feature space and extract potential defect feature data;

[0173] Step S52: performing defect morphology analysis on the solder joint defect feature space to obtain a defect morphology feature link;

[0174] Step S53: performing multi-dimensional defect type identification on the potential defect feature data based on the defect morphology feature link to obtain a defect type identification result;

[0175] Step S54: quantifying the degree of quality impact of the defect type identification result to generate a welding quality assessment value;

[0176] Step S55: performing adaptive quality grade evaluation based on the welding quality evaluation value and constructing a welding quality diagnosis strategy.

[0177] Specifically, the welding quality assessment model is used to perform multi-category defect recognition on the weld defect feature space and extract potential defect feature data. Multi-category defect recognition uses a trained deep learning model to identify various types of welding defects, which can be expressed as:

[0178] P(c_i|F_defect)=QualityModel(F_defect);

[0179] F_potential = Extract(F_defect, P); where P(c_i|F_defect) is the probability that the defect feature space belongs to the i-th type of defect, and F_potential is the extracted potential defect feature data. Defect morphology analysis is performed on the solder joint defect feature space to obtain the defect morphology feature chain. Defect morphology analysis studies the geometric shape and topological characteristics of the defect, expressed as:

[0180] M_shape(i)=ShapeAnalysis(F_defect, c_i);

[0181] L_morph={M_shape(i)}; where M_shape(i) is the morphological feature of the i-th type of defect, and L_morph is the morphological feature link. Based on the defect morphological feature link, multi-dimensional defect type recognition is performed on the potential defect feature data to obtain the defect type recognition result. Multi-dimensional defect type recognition comprehensively considers factors such as defect morphology, location, and severity, and can be expressed as:

[0182] R_defect=Classify(F_potential,L_morph);

[0183] Classify is a multi-dimensional classification function, and R_defect is the defect type identification result. The quality impact of the defect type identification result is quantified to generate a welding quality assessment value. The quality impact quantification calculates the comprehensive quality assessment value based on the weight of the impact of different types of defects on welding quality, which can be expressed as:

[0184] Q_value=∑ i [w_i × count(defect_i) × severity(defect_i)]; where w_i is the weight coefficient for defect type i, count(defect_i) is the number of defects in that type, severity(defect_i) is its severity, and Q_value is the comprehensive quality assessment value. Based on the welding quality assessment values, an adaptive quality grade assessment is performed to construct a welding quality diagnosis strategy. Adaptive quality grade assessment maps quality assessment values ​​to different quality grades based on industry standards and specific application requirements, which can be expressed as:

[0185] Grade=Mapping(Q_value);

[0186] Strategy=DiagnosisRule(Grade, R_defect); where Mapping is the quality grade mapping function, DiagnosisRule is the diagnosis rule generation function, Grade is the quality grade, and Strategy is the welding quality diagnosis strategy.

[0187] In a specific embodiment, the process of executing step S6 specifically includes the following steps:

[0188] Step S61: constructing a quality prediction model for the weld defect feature space to generate a welding quality prediction framework;

[0189] Step S62: mapping the welding quality prediction framework to diagnostic rules based on the welding quality diagnosis strategy to construct a core wire welding quality diagnosis framework;

[0190] Step S63: Perform dynamic quality prediction on the core wire welding quality diagnosis framework and construct a core wire welding quality real-time monitoring model.

[0191] Specifically, a quality prediction model is constructed for the weld defect feature space to generate a welding quality prediction framework. The quality prediction model is constructed through a time series deep learning network to establish a mapping relationship between defect features and quality change trends, which can be expressed as:

[0192] Q_t+1=fP(F_defect_t, F_defect_t-1,..., F_defect_t-n);

[0193] Frame = {fP, para}; where Q_t+1 is the predicted welding quality value at time t+1, representing the welding quality status at a future time. fP is the prediction function, typically implemented using a time series network such as an LSTM or GRU. F_defect_t represents the weld defect characteristics at time t and the current defect data. Para represents the training parameters of the prediction model, including network weights and biases. Frame is the quality prediction framework. Based on the welding quality diagnosis strategy, diagnostic rule mapping is performed on the welding quality prediction framework to construct a core wire welding quality diagnosis framework. Diagnostic rule mapping transforms the quality diagnosis strategy into an executable rule set, converting the abstract diagnosis strategy into specific rules. The prediction framework and rule set are integrated into a diagnostic framework, which integrates the prediction framework and diagnostic rules. Dynamic quality prediction is performed on the core wire welding quality diagnosis framework to construct a real-time core wire welding quality monitoring model. Dynamic quality prediction continuously analyzes changes in defect characteristics during the welding process, enabling real-time monitoring of welding quality.

[0194] The above describes the industrial visual recognition method based on deep learning in the embodiment of the present application. The following describes the industrial visual recognition system based on deep learning in the embodiment of the present application. Figure 2 In the embodiments of the present application, an industrial visual recognition system based on deep learning includes:

[0195] A processing module is used to obtain a core wire welding multi-angle image dataset; perform multi-scale noise reduction optimization on the core wire welding multi-angle image dataset, and perform weld point feature enhancement processing to obtain an enhanced core wire welding image sequence;

[0196] A deep hierarchical module is used to perform adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; and to perform cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor;

[0197] The regional space module is used to automatically segment the weld area of ​​the fused welding feature tensor to obtain the weld area feature matrix; and to enhance the recognition of microscopic defects on the weld area feature matrix to obtain the weld defect feature space;

[0198] Model building module, used to obtain historical welding quality assessment data; perform dynamic feature correlation analysis on historical welding quality assessment data to build a welding quality assessment model;

[0199] Strategy fitting module, used to use welding quality assessment model to identify multi-dimensional defect types in weld defect feature space, and to perform adaptive quality grade assessment to build welding quality diagnosis strategy;

[0200] The integrated module dynamically predicts the quality of the weld defect feature space based on the welding quality diagnosis strategy and builds a real-time monitoring model for the core wire welding quality. The modules are connected through wired and / or wireless means to realize data transmission between modules.

[0201] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the system described above and its specific working process may refer to the corresponding processes in the aforementioned method embodiments and will not be described in detail here.

[0202] In this application's examples, multi-scale noise reduction optimization and solder joint feature enhancement effectively improve the quality of the original image, laying the foundation for subsequent analysis. Multi-scale noise reduction optimization utilizes a combination of wavelet transform and adaptive filtering to effectively remove noise of varying frequencies while preserving solder joint detail. Solder joint feature enhancement emphasizes key structural features of solder joints through edge enhancement and contrast adjustment, improving the accuracy of subsequent analysis.

[0203] Through adaptive feature extraction and cross-scale feature fusion, comprehensive capture and effective integration of solder joint features are achieved. Adaptive feature extraction utilizes a multi-scale convolutional network and a feature channel weighting mechanism, dynamically adjusting feature extraction strategies based on image content to capture the multi-level features of solder joints. Cross-scale feature fusion effectively integrates feature information from different scales through feature similarity calculation and hierarchical difference normalization, forming a more comprehensive representation of solder joint features.

[0204] Through automatic solder joint segmentation and enhanced micro-defect recognition, solder joints are accurately located and potential defect features are highlighted. Automatic solder joint segmentation utilizes a combination of an attention mechanism and a semantic segmentation network to precisely locate solder joints while eliminating background interference. Enhanced micro-defect recognition emphasizes potential defects in the solder joint microstructure by enhancing high-frequency details and extracting defect features, providing critical information for subsequent defect analysis.

[0205] An accurate and reliable welding quality assessment model was constructed through dynamic feature association analysis and multi-level neural network training. Dynamic feature association analysis comprehensively considers the relationship between defect characteristics, spatial distribution, and process parameters to extract key features for quality assessment. Multi-level neural network training uses deep learning algorithms to establish a mapping between features and quality, enabling accurate welding quality assessment.

[0206] Through multi-dimensional defect type recognition and adaptive quality grade assessment, accurate diagnosis and grading of welding quality are achieved. Multi-dimensional defect type recognition comprehensively considers defect morphology, location, and severity to accurately identify various weld defects. Adaptive quality grade assessment maps quality assessment values ​​to appropriate quality grades based on industry standards and application requirements, and formulates corresponding diagnostic strategies.

[0207] By building a dynamic quality prediction and real-time monitoring model, we achieve continuous monitoring and early warning of welding quality, improving production efficiency and product quality. Dynamic quality prediction utilizes a time-series deep learning network to predict welding quality trends and promptly identify potential problems. The real-time monitoring model, through diagnostic rule mapping and continuous updates, enables real-time monitoring and early warning of welding quality, ensuring production process stability and consistent product quality.

[0208] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to modify the technical solutions described in the foregoing embodiments or to substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

[0209] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0210] In the description of the present invention, it should be understood that the terms "first", "second", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0211] In the description of the present invention, unless otherwise specified, "plurality" means two or more.

[0212] In the description of the present invention, “several” means one or more, and “a large number” means two or more.

[0213] Throughout this specification, reference to terms such as "one embodiment," "some embodiments," "examples," "specific examples," or "some examples" means that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0214] The formulas in this manual are all dimensionless and calculated using numerical values. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field based on actual conditions.

[0215] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

Claims

1. An industrial visual recognition method based on deep learning, characterized in that: include: Step S1: Acquire a core wire welding multi-angle image dataset; Performing multi-scale noise reduction optimization on a core wire welding multi-angle image dataset and performing weld feature enhancement processing to obtain an enhanced core wire welding image sequence; wherein the weld feature enhancement processing includes: performing multi-directional gradient calculation on the noise reduction image data to obtain weld edge gradient features; Extract edge candidate regions based on the edge gradient features of solder joints to obtain solder joint edge candidate regions; Perform edge continuity analysis on the candidate areas of the solder joint edge and generate edge continuity scores; Performing parametric edge enhancement on the denoised image data based on the edge continuity score to generate an edge sharpening parameter range; Adaptively sharpen the denoised image data based on the edge sharpening parameter range and mark key edge feature points; Perform nonlinear edge enhancement optimization on key edge feature points to generate edge-enhanced image data; Step S2: performing adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; performing cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor; Step S3: Automatically segmenting the weld area of ​​the fused welding feature tensor to obtain a weld area feature matrix; performing microscopic defect enhancement recognition on the weld area feature matrix to obtain a weld defect feature space; Step S4: Acquire historical welding quality assessment data and corresponding image annotation data; perform defect type distribution analysis on the historical welding quality assessment data to generate a defect type distribution map; perform dynamic feature correlation analysis on the defect type distribution map to generate welding quality assessment features; perform multi-level neural network training on the welding quality assessment features to construct a welding quality assessment model; The dynamic feature association analysis includes: extracting multiple defect feature parameters based on the defect type distribution map; Calculate the defect severity of historical welding quality assessment data based on multiple defect characteristic parameters to obtain a defect severity index; Perform defect space distribution analysis on the defect type distribution map to obtain defect space distribution characteristics; Conduct correlation analysis on the defect spatial distribution characteristics based on the defect severity index to obtain defect correlation feature data; Identify factors affecting welding process parameters based on defect type distribution maps; Perform defect cause correlation analysis based on the process parameter influencing factors to obtain process parameter correlation data; Perform dynamic feature fusion analysis on process parameter correlation data and defect correlation feature data to generate welding quality assessment features; Step S5: Using the welding quality assessment model to perform multi-dimensional defect type identification on the weld defect feature space, and perform adaptive quality grade assessment to construct a welding quality diagnosis strategy; Step S6: Based on the welding quality diagnosis strategy, dynamic quality prediction is performed on the weld defect feature space, and a real-time monitoring model for core wire welding quality is constructed.

2. The industrial visual recognition method based on deep learning according to claim 1, characterized in that: The specific steps of step S1 are: Step S11: Acquire a core wire welding multi-angle image dataset; Step S12: performing illumination imbalance correction on the core wire welding multi-angle image data set to obtain illumination balanced image data; Step S13: performing multi-scale wavelet transform noise reduction on the illumination-equalized image data to identify image noise interference areas; Step S14: performing adaptive filtering on the illumination-equalized image data according to the noise interference area to obtain noise-reduced image data; Step S15: performing solder joint edge enhancement processing on the noise reduction image data to generate edge enhanced image data; Step S16: performing contrast adaptive adjustment processing on the edge enhanced image data to obtain an enhanced core wire welding image sequence.

3. The industrial visual recognition method based on deep learning according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: performing multi-scale convolutional network processing on the enhanced core wire welding image sequence to generate feature maps of different scales; Step S22: performing solder joint feature recognition analysis on feature images of different scales to obtain a multi-level solder joint feature representation; Step S23: performing adaptive feature channel weighting based on the multi-level weld feature representation to generate a multi-level welding feature map; Step S24: performing feature similarity calculation on the multi-level welding feature maps to obtain a feature similarity matrix; Step S25: normalizing the feature level differences of the multi-level welding feature map to generate a standardized multi-level feature map; Step S26: performing cross-scale feature fusion on the standardized multi-level feature map based on the feature similarity matrix to generate a fused welding feature tensor.

4. The industrial visual recognition method based on deep learning according to claim 1, characterized in that: The specific steps of step S3 are: Step S31: performing attention mechanism enhancement on the fused welding feature tensor to obtain a weld spot attention area tensor; Step S32: Automatically segmenting the solder joint region tensor to obtain a solder joint region feature matrix; Step S33: performing high-frequency detail enhancement processing on the solder joint area feature matrix to obtain a microscopic detail enhancement feature matrix; Step S34: performing micro-defect enhancement identification on the micro-detail enhancement feature matrix to obtain a solder joint defect feature space.

5. The industrial visual recognition method based on deep learning according to claim 1, characterized in that: The specific steps of step S5 are: Step S51: using the welding quality assessment model to perform multi-category defect recognition on the weld defect feature space and extract potential defect feature data; Step S52: performing defect morphology analysis on the solder joint defect feature space to obtain a defect morphology feature link; Step S53: performing multi-dimensional defect type identification on the potential defect feature data based on the defect morphology feature link to obtain a defect type identification result; Step S54: quantifying the degree of quality impact of the defect type identification result to generate a welding quality assessment value; Step S55: performing adaptive quality grade evaluation based on the welding quality evaluation value and constructing a welding quality diagnosis strategy.

6. The industrial visual recognition method based on deep learning according to claim 1, characterized in that: The specific steps of step S6 are: Step S61: constructing a quality prediction model for the weld defect feature space to generate a welding quality prediction framework; Step S62: mapping the welding quality prediction framework to diagnostic rules based on the welding quality diagnosis strategy to construct a core wire welding quality diagnosis framework; Step S63: Perform dynamic quality prediction on the core wire welding quality diagnosis framework and construct a core wire welding quality real-time monitoring model.

7. An industrial visual recognition system based on deep learning, which is used to implement the industrial visual recognition method based on deep learning according to any one of claims 1 to 6, characterized in that: include: A processing module for acquiring a core wire welding multi-angle image dataset; Multi-scale noise reduction optimization is performed on the core wire welding multi-angle image dataset, and weld point feature enhancement is performed to obtain an enhanced core wire welding image sequence; A deep hierarchical module is used to perform adaptive feature extraction on the enhanced core wire welding image sequence to generate a multi-level welding feature map; and to perform cross-scale feature fusion on the multi-level welding feature map to generate a fused welding feature tensor; The regional space module is used to automatically segment the weld area of ​​the fused welding feature tensor to obtain the weld area feature matrix; and to enhance the recognition of microscopic defects on the weld area feature matrix to obtain the weld defect feature space; Model building module, used to obtain historical welding quality assessment data; perform dynamic feature correlation analysis on historical welding quality assessment data to build a welding quality assessment model; Strategy fitting module, used to use welding quality assessment model to identify multi-dimensional defect types in weld defect feature space, and to perform adaptive quality grade assessment to build welding quality diagnosis strategy; The integrated module performs dynamic quality prediction on the weld defect feature space based on the welding quality diagnosis strategy and builds a real-time monitoring model for core wire welding quality; each module is connected via wired and / or wireless means.

Images