CN-121998938-A - Method and system for detecting dispensing quality of dispensing detector

CN121998938ACN 121998938 ACN121998938 ACN 121998938ACN-121998938-A

Abstract

The invention relates to the technical field of industrial machine vision detection, and discloses a method and a system for detecting dispensing quality of a dispensing detector. The method comprises the steps of collecting multispectral images of a dispensing area, fusing based on colloid transmission characteristics to obtain a multiband image set, executing median filtering and illumination normalization on the multiband image set to obtain a correction image set, executing principal component analysis on the correction image set to extract a band difference vector, inputting a convolution neural network by combining texture gradients to obtain a fusion feature map, executing threshold comparison and morphological screening on the fusion feature map to obtain a preliminary defect mark, intercepting candidate image blocks, extracting reflectivity and texture features, determining final defect information by using a support vector machine, and adjusting preprocessing parameters in a closed loop based on detection confidence variance. The invention can overcome the interference of complex illumination and material changeable environments, and realize the high-precision and high-stability detection of the micro-dispensing defects.

Inventors

WEN BING
ZHAO LIHONG
ZHOU HUI

Assignees

惠州市精而美科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260126

Claims (9)

1. The method for detecting the dispensing quality of the dispensing detector is characterized by comprising the following steps of: collecting multispectral images of the dispensing area to obtain original photoelectric data, and carrying out fusion processing on the original photoelectric data according to the transmission characteristics of the colloid material obtained in advance to obtain a multiband image set containing reflection and absorption characteristics; performing median filtering and illumination normalization processing on the multiband image set to obtain a corrected image set; Traversing the corrected image set to construct a multidimensional spectrum vector, executing principal component analysis dimension reduction processing on the multidimensional spectrum vector, calculating the projection amplitude of principal component characteristics after the dimension reduction processing, and combining to obtain a wave band difference vector matrix reflecting material characteristics; constructing a texture gradient matrix based on the band difference vector matrix, splicing the texture gradient matrix and the band difference vector matrix, inputting the texture gradient matrix and the band difference vector matrix into a preset convolutional neural network model, and extracting features through a convolutional layer to obtain a fusion feature map; comparing the pixel values of the fusion feature map with a preset colloid distribution abnormal threshold to generate an abnormal distribution matrix, and performing morphological operation and connected domain screening on the abnormal distribution matrix to obtain a preliminary defect mark; And intercepting the candidate defect image blocks according to the preliminary defect marks, extracting reflectivity curves and texture feature vectors of the candidate defect image blocks, inputting the reflectivity curves and the texture feature vectors into a preset support vector machine classifier, and determining final defect information.
2. The method for detecting the dispensing quality of a dispensing inspection machine according to claim 1, wherein the collecting the multispectral image of the dispensing area to obtain the original photoelectric data, and performing fusion processing on the original photoelectric data according to the transmission characteristics of the pre-obtained colloid material to obtain the multiband image set including the reflection and absorption characteristics, includes: Scanning the dispensing area by a multispectral camera according to preset exposure time to obtain original photoelectric data; Converting the original photoelectric data into a single-band two-dimensional matrix sequence according to a sensor wavelength response curve of the multispectral camera; Calculating the reflection intensity and absorption coefficient of each pixel position by combining a preset surface curvature model aiming at the single-band two-dimensional matrix sequence; and according to the transmission characteristics of the colloid material, fusing the characteristic data blocks with different reflection intensities and absorption coefficients in the single-band two-dimensional matrix sequence to obtain a multiband image set containing reflection absorption characteristics.
3. The method for detecting the dispensing quality of the dispenser according to claim 1, wherein the performing median filtering and illumination normalization on the multiband image set to obtain a corrected image set includes: Constructing a sliding neighborhood window for each wave band image in the multi-wave band image set, taking the median of gray values in the window, and generating a smooth gray matrix; extracting low-frequency components from the smooth gray matrix to obtain background illumination field data; performing element-by-element division operation on the smooth gray matrix and the background illumination field data to obtain reflection component data; and carrying out gray stretching on the reflection component data to obtain a corrected image set.
4. The method of claim 1, wherein traversing the corrected image set to construct a multi-dimensional spectral vector, performing principal component analysis dimension reduction processing on the multi-dimensional spectral vector, calculating projection amplitudes of principal component features after the dimension reduction processing, and combining to obtain a band difference vector matrix reflecting material characteristics, comprises: extracting gray values of the same space coordinate position of all wave bands in the corrected image set, and constructing a multidimensional spectrum vector; Calculating covariance matrixes of the multidimensional spectrum vectors at all positions, and carrying out feature decomposition on the covariance matrixes to obtain feature values and feature vectors; selecting a part of the feature vector corresponding to the feature value, wherein the accumulated contribution rate of the part exceeds a preset contribution rate threshold value, and forming a projection matrix; projecting all the multidimensional spectrum vectors by using the projection matrix to obtain low-dimensional main component characteristic data; and calculating the projection amplitude of each main component characteristic data, and combining according to preset weights to generate a wave band difference vector matrix.
5. The method for detecting the dispensing quality of the dispensing detector of claim 1, wherein the constructing a texture gradient matrix based on the band difference vector matrix, splicing the texture gradient matrix and the band difference vector matrix, inputting the texture gradient matrix and the band difference vector matrix into a convolutional neural network model, and extracting features through a convolutional layer to obtain a fusion feature map comprises: Constructing a local sliding window by taking pixel points in the band difference vector matrix as the center, and calculating gray level difference statistics of pixels in the local sliding window to generate a texture gradient matrix; Splicing the texture gradient matrix and the band difference vector matrix in the channel dimension to construct a composite characteristic tensor; And inputting the composite characteristic tensor into a convolutional neural network model, extracting the joint characteristics of the space and the spectrum by using a plurality of layers of convolutional kernels, and performing dimension reduction mapping to obtain a fusion characteristic diagram.
6. The method for detecting the dispensing quality of the dispensing inspection machine according to claim 1, wherein comparing the pixel values of the fused feature image with a preset threshold value to generate an abnormal distribution matrix, and performing morphological operation and connected domain screening on the abnormal distribution matrix to obtain a preliminary defect mark, comprises: Comparing the pixel value of each pixel in the fusion feature map with a preset colloid distribution abnormal threshold, assigning pixel points with pixel values larger than the colloid distribution abnormal threshold as a first mark value, and assigning other pixel points as a second mark value to generate an initial abnormal distribution matrix; Sequentially executing morphological corrosion and expansion operation on the initial abnormal distribution matrix, eliminating noise points, and obtaining a purified abnormal distribution matrix; Carrying out connected domain analysis on the purifying abnormal distribution matrix, and marking out independent abnormal plaques; and screening abnormal plaques with pixel areas exceeding the preset minimum defect area, extracting contour coordinates and mapping the contour coordinates to an original image coordinate system to obtain a preliminary defect mark.
7. The method for detecting the dispensing quality of a dispenser according to claim 1, wherein the capturing candidate defect image blocks according to the preliminary defect mark, extracting a reflectivity curve and a texture feature vector of the candidate defect image blocks, inputting the reflectivity curve and the texture feature vector into a preset support vector machine classifier, and determining final defect information comprises: mapping the preliminary defect mark back to the multiband image set, and intercepting a corresponding candidate defect image block; calculating the average gray values of the candidate defect image blocks under different wave bands to obtain a reflectivity curve, and calculating gray level co-occurrence matrix characteristics of the candidate defect image blocks to obtain texture feature vectors; splicing the reflectivity curve and the texture feature vector to generate a combined feature vector; And inputting the combined feature vector into a preset support vector machine classifier, and determining the final defect type and position information of the candidate defect image block.
8. The method for inspecting glue dispensing quality of a glue dispensing inspection machine according to claim 7, wherein the capturing candidate defect image blocks according to the preliminary defect mark, extracting reflectivity curves and texture feature vectors of the candidate defect image blocks, inputting the reflectivity curves and texture feature vectors into a preset support vector machine classifier, and determining final defect information, further comprises: Counting the classification confidence corresponding to the final defect information obtained by continuous multi-frame image detection, and calculating the confidence value variance to obtain a confidence variance sequence; judging whether the confidence coefficient variance sequence exceeds a preset stable interval or not; If the gray level histogram exceeds the local contrast matrix, calculating a gray level histogram of the image corresponding to the original photoelectric data; According to the distribution characteristics of the gray level histogram and the statistical value of the local contrast matrix, calculating to obtain a filter kernel size adjustment gradient and a gain coefficient correction value; Performing median filtering by using the adjusted kernel size, performing illumination normalization by using the corrected gain coefficient, and performing optimization processing on the subsequently acquired multispectral image; If not, no adjustment is necessary.
9. A dispensing quality detection system for a dispensing detector, comprising: The image acquisition module is used for acquiring multispectral images of the dispensing area to obtain original photoelectric data, and carrying out fusion processing on the original photoelectric data according to the transmission characteristics of the colloid material acquired in advance to obtain a multiband image set containing reflection and absorption characteristics; The image preprocessing module is used for carrying out median filtering and illumination normalization processing on the multiband image set to obtain a corrected image set; The feature extraction module is used for traversing the correction image set to construct a multi-dimensional spectrum vector, performing principal component analysis and dimension reduction on the multi-dimensional spectrum vector, calculating the projection amplitude of principal component features after the dimension reduction, and combining to obtain a wave band difference vector matrix reflecting the material characteristics; The feature fusion module is used for constructing a texture gradient matrix based on the wave band difference vector matrix, splicing the texture gradient matrix and the wave band difference vector matrix, inputting the texture gradient matrix and the wave band difference vector matrix into a preset convolutional neural network model, and extracting features through a convolutional layer to obtain a fusion feature map; The preliminary defect marking module is used for comparing the pixel values of the fusion feature images with a preset colloid distribution abnormal threshold value to generate an abnormal distribution matrix, and performing morphological operation and connected domain screening on the abnormal distribution matrix to obtain a preliminary defect mark; And the defect classification confirming module is used for intercepting the candidate defect image blocks according to the preliminary defect marks, extracting the reflectivity curve and the texture feature vector of the candidate defect image blocks, inputting the reflectivity curve and the texture feature vector into a preset support vector machine classifier, and determining final defect information.

Description

Method and system for detecting dispensing quality of dispensing detector Technical Field The invention relates to the technical field of industrial automatic machine vision detection, in particular to a method and a system for detecting the dispensing quality of a dispensing detector. Background At present, the dispensing process is used as a key link for ensuring the tightness, the conductivity and the structural stability of a product in the electronic assembly, the manufacture of automobile parts and precision equipment, and the quality of the product directly determines the reliability of a final product. Along with the improvement of the automation degree of industrial production, the real-time and high-precision monitoring demand of the dispensing process is increasingly urgent, and the visual detection technology based on image processing becomes a core means for guaranteeing the product quality in the field. In one prior art, detection schemes based on single-band visible light imaging or conventional machine vision techniques are mainly used. According to the scheme, a two-dimensional image of a colloid area is generally acquired under standard lighting conditions, and obvious defects such as glue breaking, glue overflow or position deviation are identified by setting a contrast threshold value according to the brightness or color difference of a colloid material and a base material under a visible light wave band. This approach, which relies on single band imaging and basic feature analysis, has inherent limitations. Because the method lacks the capability of capturing the multidimensional optical characteristics of the colloid material, when the actual production environment has illumination fluctuation or the thickness, the transparency and the color of the colloid change, the contrast of the image can be obviously reduced, and the feature extraction is difficult. Particularly, when the optical reflection characteristics of the colloid and the base material are similar, the normal coverage and the local deletion of a single-band image are difficult to distinguish through simple gray level difference, and the method ignores the implicit association structure and the spatial distribution information between different spectrum bands, so that the detection stability of the colloid is poor under the complex working condition, and misjudgment or missed judgment is extremely easy to occur. Therefore, the core technical problem faced by the prior art is how to overcome the defect of insufficient information quantity of the traditional single-band imaging by deeply mining the inter-band difference characteristic and spatial context correlation of multispectral data in complex production environments with varied illumination conditions and various colloid materials, thereby realizing high-precision identification and positioning of the dispensing defect and solving the problem of low detection precision of the traditional method under complex interference. Disclosure of Invention The invention provides a method and a system for detecting the dispensing quality of a dispensing detector, which are used for solving the technical problems of difficult feature extraction and low detection precision caused by dependence on single-band imaging in complex production environments with varied illumination and various colloid materials in the prior art. In order to solve the above technical problems, the present invention provides a method for detecting a dispensing quality of a dispensing inspection machine, including: collecting multispectral images of the dispensing area to obtain original photoelectric data, and carrying out fusion processing on the original photoelectric data according to the transmission characteristics of the colloid material obtained in advance to obtain a multiband image set containing reflection and absorption characteristics; performing median filtering and illumination normalization processing on the multiband image set to obtain a corrected image set; Traversing the corrected image set to construct a multidimensional spectrum vector, executing principal component analysis dimension reduction processing on the multidimensional spectrum vector, calculating the projection amplitude of principal component characteristics after the dimension reduction processing, and combining to obtain a wave band difference vector matrix reflecting material characteristics; constructing a texture gradient matrix based on the band difference vector matrix, splicing the texture gradient matrix and the band difference vector matrix, inputting the texture gradient matrix and the band difference vector matrix into a preset convolutional neural network model, and extracting features through a convolutional layer to obtain a fusion feature map; comparing the pixel values of the fusion feature map with a preset colloid distribution abnormal threshold to generate an abnormal distribution matrix, and performing mor