A method and device for detecting wind spot of tempered glass based on brightness perception

Abstract

This application provides a method and apparatus for detecting wind spots on tempered glass based on brightness perception. The method includes: pre-acquiring a standard optical image of the wind spots; amplifying the standard optical image based on normal illuminance fluctuations and day-night illuminance fluctuations; training a preset deep network to obtain a detection model using the amplified optical image; acquiring an image of the glass to be tested; and inputting the image of the glass to be tested into the prediction model. The technical solution provided by this application can improve the recognition accuracy of wind spots on tempered glass.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for detecting wind spots on tempered glass based on brightness perception. Background Technology

[0002] Wind spots on tempered glass are stress-induced birefringence optical distortions caused by uneven cooling during the tempering process. In most conventional scenarios, wind spots on tempered glass are not considered defects. However, in some special scenarios, such as covering the exterior walls of office buildings or solar panels with tempered glass, some wind spots may be considered defects. How to identify these wind spots that are considered defects has become a concern in the industry.

[0003] Currently, the industry mainly employs three types of technologies for detecting weathering spots on tempered glass: manual visual inspection, traditional machine vision inspection, and conventional deep learning inspection. Manual visual inspection, assisted by a polarizing light box, involves inspectors visually observing the weathering spots on the glass surface and subjectively judging whether it is acceptable. This is the mainstream offline sampling inspection method for small and medium-sized production lines. Traditional machine vision inspection uses a CCD / CMOS camera with polarized imaging. It employs traditional algorithms such as grayscale threshold segmentation, edge detection, and grayscale co-occurrence matrix texture extraction. Weathering spot features are manually designed and fixed thresholds are set to achieve defect classification. This method is mostly used for semi-automated offline inspection. Conventional deep learning inspection directly uses a general-purpose deep convolutional network, combined with general data augmentation strategies such as random rotation and hue / saturation adjustment, along with default training parameters. It is then fine-tuned on the weathering spot dataset to achieve defect recognition.

[0004] However, existing technologies are only designed for identifying wind marks on tempered glass in conventional scenarios, and their accuracy cannot meet the requirements of special scenarios. Summary of the Invention

[0005] To address the aforementioned technical issues, this application proposes a method and apparatus for detecting wind spots on tempered glass based on brightness perception, in order to improve the recognition accuracy of wind spots on tempered glass in special scenarios.

[0006] In a first aspect, one or more embodiments of this application provide a method for detecting wind spots on tempered glass based on brightness perception, comprising: pre-acquiring a standard optical image of the wind spots; amplifying the standard optical sample based on conventional illuminance fluctuations and day-night illuminance fluctuations; training a preset deep network to obtain a detection model using the amplified optical image; acquiring an image of the glass to be tested; and inputting the image of the glass to be tested into the prediction model.

[0007] Furthermore, the amplification of the standard optical sample based on conventional illuminance fluctuations includes: acquiring the standard illuminance of the production site; setting an illuminance fluctuation range according to the standard illuminance; and inputting the illuminance fluctuation range and the standard optical image into a preset data augmentation model for amplification, wherein the data augmentation model amplifies the standard optical image without changing the hue and saturation.

[0008] Furthermore, the data augmentation model includes an image luminance channel; the augmentation of the standard optical image is performed in the image luminance channel.

[0009] Furthermore, the amplification of the standard optical sample based on diurnal illuminance fluctuations includes: collecting diurnal standard brightness; setting a brightness variation gradient and contrast scaling range according to the diurnal standard brightness and the brightness of the standard optical image; inputting the brightness variation gradient, the contrast scaling, and the standard optical image into a preset data augmentation model for amplification, wherein the data augmentation model amplifies the standard optical image without changing the hue and saturation.

[0010] Furthermore, the method also includes: performing image enhancement on the standard optical image in multiple ways; the multiple ways include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-scale translation and scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%.

[0011] Secondly, one or more embodiments of this application provide a tempered glass wind spot detection device based on brightness perception, comprising: a model training module for pre-acquiring standard optical images of wind spots; amplifying the standard optical samples based on conventional illuminance fluctuations and day-night illuminance fluctuations; training a preset depth network to obtain a detection model using the amplified optical images; an acquisition module for acquiring images of the glass to be tested; and an image processing module for inputting the images of the glass to be tested into the prediction model.

[0012] Furthermore, the model training module is used to obtain the standard illuminance of the production site; set the illuminance fluctuation range according to the standard illuminance; input the illuminance fluctuation range and the standard optical image into a preset data augmentation model for amplification, wherein the data augmentation model amplifies the standard optical image without changing the hue and saturation.

[0013] Furthermore, the model training module is used to collect day and night standard brightness; based on the day and night standard brightness and the brightness of the standard optical image, a brightness change gradient and a contrast scaling range are set; the brightness change gradient, the contrast scaling, and the standard optical image are input into a preset data augmentation model for amplification, and the data augmentation model amplifies the standard optical image without changing the hue and saturation.

[0014] Furthermore, the model training module is also used to perform image enhancement on the standard optical image in multiple ways; the multiple ways include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-scale translation and scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%.

[0015] Thirdly, one or more embodiments of this application provide a storage medium for storing computer-executable instructions, which, when executed, perform the following process: pre-acquiring a standard optical image of a wind spot; amplifying the standard optical sample based on conventional illuminance fluctuations and diurnal illuminance fluctuations; training a preset deep network to obtain a detection model using the amplified optical image; acquiring an image of the glass to be tested; and inputting the image of the glass to be tested into the prediction model.

[0016] Compared with the prior art, this application can achieve at least the following technical effects: This application addresses the problem of scarce wind spot samples in tempered glass by pre-acquiring standard optical images of wind spots, providing a foundation of positive and negative samples for subsequent model training. By amplifying these standard optical samples based on normal and diurnal illuminance fluctuations, the model fully learns the differentiated characteristics of compliant and non-compliant wind spot defects under different brightness conditions, solving the problems of existing general data augmentation methods that easily damage the inherent optical characteristics of wind spots and have poor illumination robustness. Furthermore, by training a pre-defined deep network using the amplified optical images to obtain a detection model, the deep convolutional network automatically mines weak contrast defect information, addressing the insufficient ability of traditional algorithms to identify weak, blurred-edge wind spots. Finally, by acquiring images of the glass under test and inputting them into the prediction model, end-to-end automated discrimination is achieved, overcoming the problems of high subjectivity and low efficiency in manual inspection. The synergistic combination of these techniques enables wind spot detection accuracy to meet the high-precision identification requirements of tempered glass in special scenarios such as office building exteriors and solar panels. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in one or more embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a tempered glass wind spot detection method based on brightness perception, provided for one or more embodiments of this application; Figure 2 This is a schematic diagram of a tempered glass wind spot detection device based on brightness perception, provided for one or more embodiments of this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this application, the technical solutions in one or more embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on one or more embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this application.

[0020] In existing technologies, wind spots are not considered significant defects, thus the required detection accuracy is not high, and therefore, existing technologies do not deeply explore the characteristics of wind spots. Wind spots are monochromatic, so the wind spot area and the non-wind spot area can form a clearly defined interface on a grayscale image. However, wind spots are sensitive to light; changes in environmental illumination can affect the boundary between the wind spot area and the non-wind spot area.

[0021] Based on the characteristics of wind spots, such as Figure 1 As shown, this application provides a method for detecting wind spots on tempered glass based on brightness perception. The overall process is divided into two main parts: an offline model training stage and an online detection stage. The specific steps are as follows: Step 1: Acquire a standard optical image of the wind spot beforehand.

[0022] Standard optical images of wind spots refer to a set of tempered glass images, including positive and negative samples, acquired using a standardized imaging scheme. Positive samples are glass images containing wind spot defects that need to be detected, while negative samples are normal glass images containing only compliant and acceptable light spots or no spots at all. The standard optical image generation scheme involves displaying equal-width and equidistant stripes on a horizontally oriented display screen, and using a fixed-position camera to acquire images of the stripes mapped onto the tempered glass surface. If wind spot defects exist on the glass surface, the stripes will exhibit corresponding bending deformation in the wind spot area, thus transforming the invisible optical distortion caused by stress birefringence into a visible feature of stripe geometric deformation.

[0023] When acquiring standard optical images, sample preprocessing is required: background cropping and debris removal are performed on the acquired raw images to remove irrelevant interference areas such as edge supports and environmental reflections. Positive and negative samples with uniform backgrounds and complete defect features are selected to construct a balanced wind spot sample dataset. For example, for tempered glass used in construction, an effective detection area is set in the imaging image to remove imaging interference such as glass conveyor rollers and side metal positioning blocks, ensuring that the glass substrate area occupies no less than 95% of the effective pixels in the input image. This preprocessing step eliminates irrelevant noise interference from the input, enabling the subsequent deep network to focus on the wind spot features of the glass itself. This solves the problem of high background interference and low effective feature ratio in the original image due to the complex acquisition environment, achieving the technical effect of improving model training stability and convergence efficiency.

[0024] Step 2: Amplify the standard optical samples based on normal illuminance fluctuations and diurnal illuminance fluctuations.

[0025] The standard optical samples are the wind spot sample dataset constructed in step 1. This step, taking into account the inherent characteristics of wind spots—monochromatic color and light sensitivity—custom-designed a brightness-aware data enhancement strategy. Its core idea is to use the image brightness channel as the only adjustment object, and to disable random perturbation operations on hue and saturation throughout the process to avoid destroying the inherent grayscale optical characteristics of wind spots.

[0026] Furthermore, the standard optical sample is augmented based on the normal illuminance fluctuation, including: obtaining the standard illuminance of the production site; setting the illuminance fluctuation range according to the standard illuminance; inputting the illuminance fluctuation range and the standard optical image into a preset data augmentation model for augmentation, wherein the data augmentation model augments the standard optical image without changing the hue and saturation.

[0027] Regular illuminance fluctuations refer to the continuous changes in illuminance over a short timescale in industrial production lines caused by factors such as automatic camera exposure adjustment, fluctuations in light source intensity, and external shading. The standard illuminance at the production site can be obtained in real-time by an illuminance sensor deployed next to the camera's imaging plane. For example, if the average standard illuminance collected during normal production in a tempered glass production workshop is L0 = 1200 lux, then the illuminance fluctuation range set based on this value can cover ±25% of the range, i.e., L ∈ [900 lux, 1500 lux]. This illuminance fluctuation range directly causes overall image brightness shifts and local contrast fluctuations, resulting in dynamic changes in the grayscale difference Δg between the wind spot area and the normal glass substrate. Especially for edge wind spots with already weak contrast (e.g., Δg is only 5~8 gray levels), their distortion characteristics are easily masked by brightness fluctuations.

[0028] Furthermore, the data augmentation model includes an image luminance channel; augmentation of standard optical images is performed within this luminance channel. The image luminance channel can be extracted from the V component in the HSV color space or from the Y component in the YUV color space. Strictly confining the augmentation operation to the luminance channel ensures that the hue component H and saturation component S of the image remain unchanged before and after augmentation, thus fully preserving the inherent grayscale optical distortion characteristics of windblown defects and avoiding feature distortion caused by hue disturbances. For example, if the luminance of the original image in HSV space is V(x,y), the new luminance value after augmentation can be calculated using the formula V'(x,y)=V(x,y)·(1+ΔL / L0), where ΔL is the illuminance fluctuation and L0 is the standard illuminance reference value; H(x,y) and S(x,y) remain unchanged throughout the calculation process. This approach solves the problem of random adjustment of hue / saturation destroying the optical characteristics of wind spots in existing general deep learning detection, making the enhanced image highly consistent with the imaging performance under the actual illumination drift of the production line, and achieving the technical effect of improving the model's anti-interference ability to short-timescale illumination fluctuations.

[0029] Furthermore, the standard optical sample is augmented based on the diurnal illuminance fluctuation, including: collecting diurnal standard brightness; setting brightness variation gradient and contrast scaling range according to the diurnal standard brightness and the brightness of the standard optical image; inputting the brightness variation gradient, contrast scaling and the standard optical image into a preset data augmentation model for augmentation, the data augmentation model augments the standard optical image without changing the hue and saturation.

[0030] Diurnal illuminance fluctuation refers to the continuous and slow drift of illuminance on the production line as it transitions from day to night and vice versa over time. This change is not a transient jump, but a continuous and gradual process. Standard daytime and nighttime illuminance values ​​can be obtained by sampling the illuminance values ​​at the same location on the production line for 24 consecutive hours. For example, the standard illuminance at 12:00 noon is denoted as Lday = 1500 lux, and the standard illuminance at 2:00 am is denoted as Lnight = 300 lux; these two values ​​constitute the extreme values ​​of daytime and nighttime illuminance. Based on the above-mentioned standard day and night brightness and the reference brightness V0 of the standard optical image, the brightness change gradient gV and the contrast scaling interval [Cmin, Cmax] can be set: the brightness change gradient is used to simulate the gradual increase or decrease of illuminance during the day-night transition, for example, the brightness fluctuation range is controlled within ±30% of the original image brightness, that is, V'(x,y)∈[0.7·V0, 1.3·V0]; the contrast scaling interval is used to simulate the migration of local brightness and darkness contrast relationship under different illuminance, for example, the interval is set to C∈[0.8, 1.2] times.

[0031] In practice, image augmentation under fluctuating day and night illumination can be performed using the following formula: V''(x,y) = C·(V'(x,y) - Vm) + Vm, where V'(x,y) is the brightness value adjusted by the brightness change gradient gV, Vm is the local neighborhood mean, and C is the contrast scaling factor, with a value range of [0.8, 1.2]. When C>1, the grayscale difference between the wind spot area and the background glass is amplified, corresponding to periods of high illumination and clear imaging; when C<1, the grayscale difference is compressed, corresponding to periods of low illumination where the wind spot features tend to blend with the background. The significance of this simulation lies in the fact that if the model is only adapted to a single illumination environment during a fixed period of time, feature extraction will inevitably fail and detection confidence will continue to decline due to illumination drift during the actual 24 / 7 operation of the production line. By amplifying the illumination fluctuations between day and night, the model can fully learn the differentiated characteristics of compliant light spots and unqualified wind spots under different brightness and contrast conditions, thereby fundamentally improving the model's anti-interference ability against large-scale illumination drift and solving the problem that existing technologies cannot distinguish the boundary between compliant light spots and unqualified wind spots and that the detection accuracy drops significantly when switching between day and night.

[0032] Furthermore, the method also includes: performing image enhancement on the standard optical image in multiple ways; the multiple ways include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-scale translation scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%.

[0033] 90° four-way random rotation refers to randomly selecting one of the four directions (0°, 90°, 180°, 270°) to rotate the image, simulating the difference in placement orientation of glass during production line transport; horizontal and vertical multi-directional flipping includes two operations: horizontal flipping along the vertical axis and vertical flipping along the horizontal axis, used to simulate the difference in orientation between the front and back or left and right sides of the glass; 0.95~1.05 times micro-scale translation and scaling refers to a scaling factor limited to a small range of 0.95 to 1.05, accompanied by random translation not exceeding 5% of the image size, used to simulate the slight offset of the glass transport position when the camera is in a fixed position; 32×32 pixel local mosaic occlusion enhancement, accounting for no more than 5%, refers to randomly selecting several 32×32 pixel square areas on the image for mosaic blurring, with the total number of pixels in the occluded area controlled within 5% of the entire image, used to simulate the occlusion effect of local dust, minor scratches, and other interference on the glass surface on the image. The selection criteria for the above-mentioned image enhancement methods are all based on real working conditions on the production line. The core purpose is to expand the sample diversity while fully preserving the weak wind spot texture deformation features, solve the pain points of scarce defect samples in industrial scenarios and the easy destruction of effective features by general enhancement methods, thereby improving the robustness of the model to non-critical variation factors such as glass placement direction, positional offset, and slight surface contamination.

[0034] Step 3: Using the amplified optical image, train a preset deep network to obtain a detection model.

[0035] The preset deep network is a classification network with a lightweight deep convolutional network as its detection backbone. Its input is a uniformly sized glass image after step 2 (e.g., scaled to 448×448 pixel resolution), and its output is a binary classification confidence score (pass / fail) indicating whether the glass has wind-mark defects. The core forward propagation process of the deep convolutional network can be represented by the following formula: O(l+1) = ReLU(BN(W(l) * O(l) + b(l))) Where O(l) is the output feature map of the l-th convolutional layer, W(l) is the weight parameter of the l-th convolutional kernel, the symbol * represents the two-dimensional discrete convolution operation, b(l) is the corresponding bias term, BN represents the batch normalization operation, and ReLU is the linear rectified activation function ReLU(x)=max(0, x). At the end of the network, global average pooling is used to compress the spatial dimension into a one-dimensional feature vector, followed by two fully connected layers and a Softmax output layer to obtain the conditional probability P(c|x) of the image corresponding to category c.

[0036] To optimize training performance in scenarios with small sample sizes, multiple refined training mechanisms are required during the training phase. Firstly, a learning rate warm-up mechanism is implemented in the early stages of training: a total of 5 warm-up cycles are set up, with an initial learning rate ηstart = 1 × 10⁻ 5increases linearly with the training epoch k to the base learning rate ηbase=1×10⁻³, and the learning rate for each epoch can be calculated by the formula η(k)=ηstart+(ηbase-ηstart)·(k-1) / 4 (k≤5); when k>5, the learning rate is gradually attenuated according to the cosine annealing strategy. The function of this mechanism is to avoid drastic oscillation of the weight gradient in the initial iteration stage of the model, ensure stable convergence of the network, and solve the problem of poor convergence point caused by unstable initial values of model weights under small samples.

[0037] Secondly, the adaptive batch size adjustment is enabled throughout the training process: the dynamic adjustment interval of the batch size bs is set to [8,32]; when the fluctuation amplitude δJ of the single-epoch training loss value J(θ) exceeds 15% (that is, δJ=|J(k)-J(k-1)| / J(k-1)>0.15), the batch size is automatically reduced to 8~16 to increase the weight update frequency of the weak wind spot features; when the loss fluctuation is stabilized within 5% (that is, δJ<0.05), the batch size is gradually increased to 24~32 to balance the training stability and the overall convergence speed. By weighing the gradient stability and step size of single-step update, this mechanism solves the problem of insufficient model weight update for such features when the proportion of weak wind spot samples is low.

[0038] Thirdly, an early stopping mechanism and optimal weight backtracking based on the F1 value of the validation set are enabled: the F1 value of the validation set is taken as the core monitoring indicator, and the patience threshold p is set to 15 epochs; if the F1 value of the validation set does not show effective improvement for 15 consecutive epochs (that is, F1(k)<F1_best), the training iteration is automatically terminated to prevent the model from overfitting the noise of the training set; at the same time, an automatic optimal weight backtracking mechanism is matched, the model weight file θ* corresponding to the F1 peak value of the validation set is saved synchronously after each iteration, and θ* is automatically loaded as the final inference model after early stopping is triggered, discarding the weight parameters in the later overfitting stage. The F1 value is comprehensively calculated from the precision Precision and the recall Recall according to F1=2·Precision·Recall / (Precision+Recall). This series of mechanisms improves the generalization ability of the model to different glass thicknesses and lighting conditions from the training process level, enhances the feature extraction effect for low-contrast weak wind spots, and makes up for the defects of conventional deep learning models such as extensive training strategies, insufficient generalization and detection accuracy.

[0039] Step 4, collect the image of the glass to be tested.

[0040] During the online detection phase, when acquiring images of the glass under test, a stripe imaging scheme completely consistent with that used in the model training phase must be employed. This involves displaying equal-width and equidistant stripes on a horizontally positioned screen and acquiring stripe mapping images using a fixed-position camera, ensuring consistency between training and inference imaging conditions from the hardware perspective. After image acquisition, only basic edge cropping and debris removal preprocessing operations, consistent with step 1, are performed; no further data augmentation is performed to ensure strict alignment between the input image to the model and the imaging style used in the training phase. This unified imaging scheme addresses issues in existing technologies such as lack of coordinated optimization between imaging hardware and algorithms, unstable camera fixation, insufficient effective samples, and numerous imaging interferences, further improving detection stability and accuracy.

[0041] Step 5: Input the image of the glass to be tested into the prediction model.

[0042] The prediction model is the deep convolutional network trained in step 3 with its weights fixed. After inputting the image to be tested, the model performs standardized end-to-end forward inference and finally outputs a binary detection result of whether the glass is qualified or unqualified: when the confidence score of the unqualified class output by the model is greater than a preset threshold (e.g., 0.5), it is determined that the glass to be tested has a wind spot defect that needs to be detected, which can trigger a production line alarm or sorting and rejection; otherwise, it is determined to be qualified glass. After optimization by a complete set of brightness perception enhancement strategies and refined training mechanisms, the model can adaptively adapt to a large range of illumination fluctuations on the production line without the need for repeated manual adjustments to grayscale features and detection thresholds. It can stably identify weak wind spots with a grayscale difference of only 5 to 8 levels between the wind spot area and the normal glass substrate, as well as edge wind spots with smooth grayscale transitions and blurred boundaries. This solves the problem of missed detection and false detection of this type of low signal-to-noise ratio, weak gradient response wind spots by traditional threshold segmentation and edge detection algorithms.

[0043] like Figure 2 As shown, this application provides a tempered glass wind spot detection device based on brightness perception, comprising: Model training module 201 is used to pre-acquire standard optical images of wind spots; amplify standard optical samples based on regular illuminance fluctuations and day-night illuminance fluctuations; and train a pre-defined deep network to obtain a detection model using the amplified optical images. This module operates in the offline deployment phase and can be further subdivided into three parts: a sample acquisition subunit, a brightness-perceived data augmentation subunit, and a deep network refinement training subunit. The sample acquisition subunit is responsible for data preparation work such as acquiring the original image, background cropping and debris removal, and balancing the distribution of positive and negative samples. The brightness-perceived data augmentation subunit is responsible for performing data augmentation operations such as regular illuminance fluctuation amplification, day-night illuminance fluctuation amplification, and multi-mode image enhancement. In this subunit, all amplification operations are performed within the image brightness channel, and random perturbations of hue and saturation are disabled throughout the amplification process. The deep network refinement training subunit is responsible for implementing training optimization mechanisms such as learning rate warm-up, adaptive batch size adjustment, early stopping of the validation set F1 score, and optimal weight backtracking.

[0044] Furthermore, the specific implementation logic of the model training module 201 when performing augmentation based on conventional illuminance fluctuations is as follows: acquiring the standard illuminance of the production site; setting the illuminance fluctuation range according to the standard illuminance; inputting the illuminance fluctuation range and the standard optical image into a preset data augmentation model for augmentation. The data augmentation model augments the standard optical image without changing hue and saturation. The corresponding technical problem it solves is that illuminance drift within a short timescale in industrial production lines leads to missed detections and misjudgments by the detection model. The achieved technical effect is to enhance the model's anti-interference ability against short-term illuminance fluctuations, allowing the model to learn the differentiated boundaries between compliant and non-compliant light spots under different brightness conditions.

[0045] Furthermore, the specific implementation logic of the model training module 201 when performing augmentation based on day-night illumination fluctuations is as follows: Standard day-night brightness is collected; based on the standard day-night brightness and the brightness of the standard optical image, a brightness change gradient and contrast scaling range are set; the brightness change gradient, contrast scaling, and the standard optical image are input into a preset data augmentation model for augmentation. The data augmentation model augments the standard optical image without changing hue and saturation. The corresponding technical problem it solves is that the slow drift of illumination during day-night switching in the production line's 24 / 7 operation leads to the failure of model feature extraction and a decrease in detection confidence. The achieved technical effect is to enable the model to adapt to a wide range of brightness gradients and contrast relationship shifts, ensuring stable and compliant detection accuracy across all time periods of day and night.

[0046] Furthermore, the model training module 201 is also used to perform image enhancement on standard optical images in various ways; these methods include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-scale translation and scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%. This combination of spatial geometric transformation and occlusion enhancement solves the problem of sample distribution differences caused by various working conditions such as inconsistent glass placement orientation on the production line, conveyor position offset, and surface dust occlusion. It effectively expands sample diversity without destroying the wind-blown texture deformation characteristics, thereby improving the model's robustness to spatial variation factors.

[0047] The acquisition module 202 is used to acquire images of the glass under test. Deployed at the online inspection station on the production line, this module consists of a horizontally positioned display screen, a fixed-position CCD / CMOS camera, an illuminance sensor, and corresponding mechanical positioning supports. The acquisition module employs a stripe imaging scheme identical to that used in the model training phase, acquiring stripe mapping images of the tempered glass under test in real time and sending the image frames to the subsequent image processing module. Before outputting the images, this module only performs basic edge cropping and debris removal preprocessing, without data augmentation, ensuring strict alignment between the online imaging style and the training data. The technical problem it solves is the lack of coordinated optimization between imaging hardware and algorithms in existing technologies, leading to inconsistent imaging conditions. The achieved technical effect is to eliminate training-inference domain bias from the data acquisition source, improving detection stability.

[0048] The image processing module 203 is used to input the image of the glass to be tested into the prediction model. This module incorporates a deep convolutional network inference engine with weights fixed in step 3. The single-frame image from the acquisition module is normalized by tensors and input into the network. Forward inference is performed to obtain a binary classification result, and a final judgment of whether the glass is qualified or unqualified is output based on a preset threshold. When the glass is determined to be unqualified, an alarm signal or sorting execution instruction can be further output to the production line control system. This module solves the problems of high subjectivity, low efficiency, and inability to adapt to the full inspection requirements of continuous production in existing technologies for manual visual inspection. It achieves automated and high-speed discrimination of wind spot defects, eliminating subjective errors and fatigue-related missed detections, eliminating the need for repeated manual adjustment of feature thresholds, significantly reducing inspection labor costs, and adapting to continuous industrial production line deployment.

[0049] Based on the same inventive concept, one or more embodiments of this application also provide a storage medium for storing computer-executable instructions. When executed, these computer-executable instructions perform the following steps: pre-acquiring a standard optical image of wind spots; amplifying the standard optical sample based on conventional illuminance fluctuations and diurnal illuminance fluctuations; training a preset deep network to obtain a detection model using the amplified optical image; acquiring an image of the glass to be tested; and inputting the image of the glass to be tested into the prediction model. The storage medium may include common computer-readable storage media such as read-only memory (ROM), random access memory (RAM), disk, or optical disk. After the program instructions stored on it are loaded and executed by the processor, they can complete offline model training and online automatic detection of tempered glass wind spots according to the aforementioned method steps, providing tempered glass manufacturers with a conveniently deployable software-based detection solution.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting wind spots in tempered glass based on brightness perception, characterized in that, Standard optical images of wind spots are acquired in advance; the standard optical samples are amplified based on conventional illuminance fluctuations and diurnal illuminance fluctuations; Using the amplified optical image, a detection model is obtained by training a preset deep network. The detection method includes: Acquire images of the glass to be tested; The image of the glass to be tested is input into the prediction model.

2. The method according to claim 1, characterized in that, The amplification of the standard optical sample based on conventional illuminance fluctuations includes: Obtain standard illuminance at the production site; The illuminance fluctuation range is set according to the standard illuminance; The illuminance fluctuation range and the standard optical image are input into a preset data augmentation model for amplification. The data augmentation model amplifies the standard optical image without changing the hue and saturation.

3. The method according to claim 2, characterized in that, The data augmentation model includes an image brightness channel; The amplification of the standard optical image is performed in the image brightness channel.

4. The method according to claim 1, characterized in that, The amplification of the standard optical sample based on diurnal illuminance fluctuations includes: Collect standard day and night brightness; Based on the day and night standard brightness and the brightness of the standard optical image, set the brightness variation gradient and contrast scaling range; The brightness variation gradient, the contrast scaling, and the standard optical image are input into a preset data augmentation model for amplification. The data augmentation model amplifies the standard optical image without changing the hue and saturation.

5. The method according to claim 1, characterized in that, The method further includes: The standard optical image is enhanced in various ways; The multiple methods include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-proportional translation and scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%.

6. A tempered glass wind spot detection device based on brightness perception, characterized in that, include: The model training module is used to pre-acquire standard optical images of wind spots; The standard optical sample was amplified based on conventional illuminance fluctuations and diurnal illuminance fluctuations. A detection model is obtained by training a pre-defined deep network using the amplified optical image; The acquisition module is used to acquire images of the glass under test. An image processing module is used to input the image of the glass to be tested into the prediction model.

7. The apparatus according to claim 6, characterized in that, The model training module is used to obtain the standard illuminance of the production site; set the illuminance fluctuation range according to the standard illuminance; input the illuminance fluctuation range and the standard optical image into a preset data augmentation model for amplification, wherein the data augmentation model amplifies the standard optical image without changing the hue and saturation.

8. The apparatus according to claim 6, characterized in that, The model training module is used to collect day and night standard brightness; based on the day and night standard brightness and the brightness of the standard optical image, it sets the brightness change gradient and contrast scaling range; The brightness variation gradient, the contrast scaling, and the standard optical image are input into a preset data augmentation model for amplification. The data augmentation model amplifies the standard optical image without changing the hue and saturation.

9. The apparatus according to claim 6, characterized in that, The model training module is also used to perform image enhancement on the standard optical image in various ways; The multiple methods include one or more of the following: 90° four-way random rotation, horizontal and vertical multi-directional flipping, 0.95~1.05 times micro-proportional translation and scaling, and 32×32 pixel local mosaic occlusion enhancement with a proportion not exceeding 5%.

10. A storage medium for storing computer-executable instructions, which, when executed, perform the following process: Standard optical images of wind spots are acquired in advance; the standard optical samples are amplified based on normal illuminance fluctuations and day-night illuminance fluctuations; and a detection model is obtained by training a preset deep network using the amplified optical images. Acquire images of the glass to be tested; The image of the glass to be tested is input into the prediction model.

Images