Touch screen defect detection method and system based on machine vision

Abstract

The invention discloses a touch screen defect detection method and system based on machine vision, and relates to the technical field of defect detection.During operation of the system, a defect sample library is constructed based on defect distribution data in historical detection results, and a pre-constructed basic detection model is trained until a preset condition is met; the trained basic detection model is determined as a touch screen defect detection model, the touch screen defect detection model comprises a clustering network and a trajectory analysis network, the clustering network is a deep residual network containing a multi-scale feature extraction mechanism, and the trajectory analysis network is a dynamic tracking module containing a time sequence correlation structure; and inputting touch screen defect data collected in real time into the touch screen defect detection model, and outputting defect early warning information and process offset suggestions.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, specifically to a method and system for detecting defects in touchscreens based on machine vision. Background Technology

[0002] With the rapid development of the smart terminal industry, touchscreens, as the core carrier of human-computer interaction, directly determine user experience and product yield through their appearance and optical quality. In modern manufacturing, machine vision inspection systems have become an indispensable quality monitoring link on touchscreen production lines, especially in processes such as optical bonding, polarizer attachment, and conductive mesh formation. Through multi-light source, multi-angle imaging and algorithm recognition, the system can quickly detect visible defects such as scratches, bubbles, dark spots, residual adhesive, and mesh imprints under high-rate conditions. However, while traditional inspection systems have a high defect detection rate on single samples, they often lack the ability to perform global analysis across batches and time periods, failing to identify hidden patterns such as "repeated occurrences of similar defects in similar areas." Consequently, in the field of touchscreen inspection, there is a growing demand for "machine vision-based defect evolution analysis systems," shifting inspection from "single-unit judgment" to "trend diagnosis," becoming an important direction for quality control in smart manufacturing.

[0003] In current touchscreen production lines, the LOCA (Optical Adhesive-Based Acrylic) bonding process is one of the most critical and complex steps. In this process, the thickness of the optical adhesive, curing energy, and bonding uniformity have a significant impact on display quality. In actual production, while machine vision inspection equipment can detect defects such as "bright spots" or "bubbles" after each screen is bonded, it often treats each inspection result as an independent event, lacking the ability to trace back longitudinally or compare laterally. When the dispensing trajectory or curing lamp position of a bonding device slightly deviates, the system will still issue an alarm for that single screen, without recognizing that "repeated bright spots in the same location" represent a systemic deviation in the equipment. This defect may not be significant in the short term, but as production continues, its cumulative effect can lead to uneven brightness, abnormal local reflections, or visual banding in the same area across the entire batch of touchscreens. Traditional detection algorithms cannot detect this trend of "repeated spatial patterns," resulting in production line personnel being able to identify individual defects but struggling to promptly identify the root cause of "continuous anomalies," leading to misjudgments, misplacement, or even the scrapping of the entire batch. Therefore, an innovative mechanism with spatial clustering and temporal evolution analysis capabilities is needed to dynamically self-cluster and track the defect distribution, thereby identifying equipment offset and process drift at an early stage. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a touchscreen defect detection method and system based on machine vision, which solves the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a touchscreen defect detection method based on machine vision, comprising the following steps: constructing a defect sample library based on defect distribution data from historical detection results; training a pre-constructed basic detection model based on the defect sample library until preset conditions are met, and determining the trained basic detection model as the touchscreen defect detection model; wherein, the basic detection model includes: a clustering network and a trajectory analysis network; the clustering network is a deep residual network containing a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module containing a time-series correlation structure; inputting real-time collected touchscreen defect data into the touchscreen defect detection model, and outputting defect warning information and process offset suggestions.

[0006] Preferably, the defect sample library includes: a clustering sample library and a trajectory sample library; the step of constructing the defect sample library based on defect distribution data in historical detection results includes: acquiring defect distribution data in historical detection results; performing sliding sampling on the defect distribution data based on a spatial window of a preset size to obtain a local defect distribution map; marking recurring defect areas in the local defect distribution map to obtain a labeled data map; determining the labeled data map with a first label and the local defect distribution map with a second label as samples in the clustering sample library; determining the labeled data map and the corresponding time series data as samples in the trajectory sample library; wherein, the labeled data map is a feature, and the time series data is a label.

[0007] Preferably, the recurring defect areas include: bright spot areas, bubble areas, and dark spot areas; the area, location, and time interval of the marking of the recurring defect areas are all randomly generated within a preset range.

[0008] Preferably, the clustering network includes three base layers and a global pooling layer connected in sequence; each base layer includes two feature extraction blocks connected in sequence; each feature extraction block includes a convolutional layer, a multi-scale feature fusion layer and a spatial mapping layer connected in sequence.

[0009] Preferably, the trajectory analysis network includes a time encoder and a trend decoder; the time encoder includes three consecutive feature extraction blocks, each feature extraction block including two parallel channels, the first channel including two consecutive 3×1 convolutional layers, a normalization layer and an activation function; the second channel is a time-related structure including a 1×1 convolutional layer; the trend decoder includes two consecutive 3×1 convolutional layers and a flattening operation layer.

[0010] Preferably, the loss function of the trajectory analysis network is based on a comprehensive evaluation of the temporal consistency constraint and the spatial distribution overlap of each data sample.

[0011] Preferably, the touchscreen defect detection model includes: a trained clustering network and a trained trajectory analysis network; the step of inputting real-time collected touchscreen defect data into the touchscreen defect detection model and outputting defect warning information and process deviation suggestions includes: inputting real-time collected touchscreen defect data into the trained clustering network to obtain a local defect distribution map and a labeled data map; inputting the labeled data map into the trained trajectory analysis network to obtain a time series trend map; and integrating the local defect distribution map and the time series trend map to determine the defect warning information and process deviation suggestions.

[0012] Preferably, before inputting the real-time acquired touchscreen defect data into the touchscreen defect detection model, the method further includes: performing multi-scale enhancement and noise suppression processing on the real-time acquired touchscreen image data; wherein, the multi-scale enhancement processing includes: brightness normalization based on the regional gradient change rate and structural enhancement based on the local contrast factor; the noise suppression processing includes: dynamically adjusting the image smoothing parameters according to the feature extraction results of the clustering network, so as to improve the overall signal-to-noise ratio without losing the subtle defect features.

[0013] Preferably, after the steps of outputting defect warning information and process deviation suggestions, the system further includes: generating a defect distribution evolution report based on the defect warning information; the defect distribution evolution report includes: a defect occurrence frequency distribution map, a spatial clustering coefficient change curve, and a process deviation trend map; the report is used to analyze the recurring areas and location deviation trajectories of the same type of defect within a preset time window; when the defect clustering coefficient is detected to be continuously increasing or the deviation trajectory exceeds a threshold, the system automatically triggers equipment maintenance or process parameter calibration commands to achieve closed-loop quality control of the production line.

[0014] A machine vision-based touchscreen defect detection system includes a sample library construction module, a model training module, and a detection module. The system includes a sample library construction module for building a defect sample library based on defect distribution data from historical inspection results; a model training module for training a pre-built basic detection model based on the defect sample library until preset conditions are met, and then determining the trained basic detection model as the touchscreen defect detection model; wherein, the basic detection model includes: a clustering network and a trajectory analysis network; the clustering network is a deep residual network containing a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module containing a time series correlation structure; and a detection module for inputting real-time collected touchscreen defect data into the touchscreen defect detection model and outputting defect warning information and process offset suggestions.

[0015] This invention provides a touchscreen defect detection method and system based on machine vision, which has the following beneficial effects: (1) When the system is running, a defect sample library is constructed based on the defect distribution data in the historical detection results, and a pre-constructed basic detection model is trained until the preset conditions are met. The trained basic detection model is then determined as the touch screen defect detection model, which includes a clustering network and a trajectory analysis network. The clustering network is a deep residual network containing a multi-scale feature extraction mechanism, and the trajectory analysis network is a dynamic tracking module containing a time series correlation structure. The real-time collected touch screen defect data is input into the touch screen defect detection model, and defect warning information and process offset suggestions are output.

[0016] (2) The present invention provides a machine vision-based touchscreen defect detection method and system. By constructing a defect sample database including a clustering sample library and a trajectory sample library, and combining a clustering network with multi-scale feature extraction and a trajectory analysis network with time series correlation, a closed-loop process from historical detection data to real-time defect analysis is realized. This method can automatically identify and classify repetitive defects such as bright spots, dark spots, and bubbles in the online detection environment of touchscreen production lines, and complete the intelligent identification task of defect spatial distribution, temporal evolution trend, and process deviation correlation. The system not only outputs real-time early warning information of defects, but also generates process adjustment suggestions and trend analysis reports at the same time, thereby realizing the functional leap from "single-piece detection" to "dynamic process monitoring", significantly improving the automation and intelligence level of touchscreen production quality monitoring.

[0017] (3) Compared with existing touchscreen inspection technologies that rely solely on static image analysis, this invention introduces defect clustering and time series tracking mechanisms at the method level, enabling the inspection model to identify "recurring spatial defect patterns," thus overcoming the limitation of traditional methods that can only identify "single-time anomalies." The multi-scale feature fusion structure of the clustering network can accommodate defect feature extraction under different sizes and lighting conditions; while the time encoder and trend decoder of the trajectory analysis network enable the system to continuously monitor the evolution and shift of defect positions, detecting potential equipment drift or process deviations. Furthermore, the newly added multi-scale image enhancement and noise suppression steps effectively improve the signal-to-noise ratio of the inspection image, and the defect distribution evolution report and automatic trigger feedback mechanism upgrade the inspection system from a simple judgment tool to a quality control unit capable of proactively intervening in the production process.

[0018] (4) Through the synergistic effect of the above-mentioned innovative structure and algorithm, this invention achieves a significant improvement in the accuracy and stability of touch screen detection. Compared with traditional detection systems, this method can improve the accuracy of repetitive defect identification by about 30% under the same hardware conditions, and can predict about 80% of process deviation trends in advance, greatly reducing the risk of batch scrapping and the cost of manual re-inspection. At the same time, the process deviation suggestions and evolution reports output by the system can be directly used for the self-calibration or parameter adjustment of production equipment, realizing an automated closed loop of detection-analysis-control. Overall, this invention, through technological innovations such as cluster recognition, time tracking, feature enhancement and feedback linkage, constructs a highly robust, high-precision, and evolvable machine vision detection system, providing a sustainable optimization technical foundation for intelligent quality control of touch screen production lines. Attached Figure Description

[0019] Figure 1 This is a block diagram of a touchscreen defect detection system based on machine vision according to the present invention. Figure 2 This is a schematic diagram illustrating the steps of a machine vision-based touchscreen defect detection method according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 This invention provides a machine vision-based method for detecting defects in touchscreens. Please refer to [link / reference]. Figure 1 The process includes the following steps: constructing a defect sample library based on defect distribution data from historical inspection results; training a pre-built basic inspection model based on the defect sample library until preset conditions are met, and determining the trained basic inspection model as the touchscreen defect inspection model; wherein, the basic inspection model includes: a clustering network and a trajectory analysis network; the clustering network is a deep residual network containing a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module containing a time-series correlation structure; inputting real-time collected touchscreen defect data into the touchscreen defect inspection model, and outputting defect warning information and process offset suggestions.

[0022] In this embodiment, the present invention constructs a defect sample library based on defect distribution data from historical detection results, and uses this sample library to train a preset basic detection model, ultimately forming a touchscreen defect detection model with self-learning capabilities, realizing the transformation from "experience-based detection" to "data-driven recognition." The basic detection model simultaneously incorporates two structures: a clustering network and a trajectory analysis network. The former, through a deep residual network with a multi-scale feature extraction mechanism, can accurately identify defects of different sizes and shapes under complex lighting and multi-layered structures. The latter, through a dynamic tracking module with a time-series correlation structure, can analyze the recurrence and positional shift of defects over time, identifying potential process fluctuation trends. This method can directly output defect warning information and process deviation suggestions during the real-time detection stage, enabling the system to achieve coordinated linkage between defect identification and process monitoring without changing hardware conditions. Compared with existing traditional detection methods that rely solely on single-frame image judgment, the present invention significantly improves the stability and trend analysis capabilities of detection, effectively reducing the missed detection rate of recurring defects, providing early warnings of production deviations, and thus improving the overall production line yield and process consistency.

[0023] Example 2 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the defect sample library includes a clustering sample library and a trajectory sample library. The steps for constructing the defect sample library based on defect distribution data in historical detection results include: acquiring defect distribution data in historical detection results; performing sliding sampling on the defect distribution data based on a spatial window of a preset size to obtain a local defect distribution map; marking recurring defect areas in the local defect distribution map to obtain a labeled data map; determining the labeled data map with a first label and the local defect distribution map with a second label as samples in the clustering sample library; and determining the labeled data map and the corresponding time series data as samples in the trajectory sample library; wherein, the labeled data map is a feature, and the time series data is a label.

[0024] The recurring defect areas include: bright spot areas, bubble areas, and dark spot areas; the area, location, and time interval of the recurring defect areas are all randomly generated within a preset range.

[0025] The clustering network comprises three base layers and a global pooling layer connected in sequence; each base layer comprises two feature extraction blocks connected in sequence; each feature extraction block comprises a convolutional layer, a multi-scale feature fusion layer and a spatial mapping layer connected in sequence.

[0026] In this embodiment, the present invention achieves systematic management and feature association of touchscreen defect distribution information by constructing a defect sample system comprising a clustering sample library and a trajectory sample library. The clustering sample library uses local defect distribution maps and labeled data maps as core samples, enabling spatial clustering feature extraction of areas such as bright spots, bubbles, and dark spots that repeatedly appear in historical inspections, thereby forming a stable defect type identification benchmark. The trajectory sample library pairs labeled data maps with time-series data to construct a dynamic evolution model of defects over the production time dimension. Through sliding sampling and random labeling mechanisms, this method can expand the sample coverage without relying on manual intervention, enhancing the model's generalization ability and anti-overfitting performance. Simultaneously, the clustering network employs a deep residual structure containing multi-scale feature fusion layers and spatial mapping layers, enabling accurate feature extraction and morphological preservation of defect regions under different scales and brightness conditions, thereby effectively distinguishing surface noise from real defects. This structure enables the detection model to maintain high detection accuracy and stability even when facing complex backgrounds, uneven brightness, and multi-layer bonding structures. It significantly improves the system's adaptability and detection reliability in multiple scenarios, and provides a more comprehensive and intelligent algorithmic foundation for long-term quality monitoring in the touch screen production process.

[0027] Example 3 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically: the trajectory analysis network includes a time encoder and a trend decoder; the time encoder includes three consecutive feature extraction blocks, each of which includes two parallel channels. The first channel includes two consecutive 3×1 convolutional layers, a normalization layer, and an activation function; the second channel is a time-related structure, including a 1×1 convolutional layer; the trend decoder includes two consecutive 3×1 convolutional layers and a flattening operation layer.

[0028] The loss function of the trajectory analysis network is based on a comprehensive evaluation of the temporal consistency constraint and spatial distribution overlap of each data sample.

[0029] The touchscreen defect detection model includes: a trained clustering network and a trained trajectory analysis network. The steps of inputting real-time collected touchscreen defect data into the touchscreen defect detection model and outputting defect warning information and process deviation suggestions include: inputting the real-time collected touchscreen defect data into the trained clustering network to obtain a local defect distribution map and an annotated data map; inputting the annotated data map into the trained trajectory analysis network to obtain a time series trend map; and integrating the local defect distribution map and the time series trend map to determine the defect warning information and process deviation suggestions.

[0030] Before inputting the real-time acquired touchscreen defect data into the touchscreen defect detection model, the method further includes: performing multi-scale enhancement and noise suppression processing on the real-time acquired touchscreen image data; wherein, the multi-scale enhancement processing includes: brightness normalization based on the regional gradient change rate and structural enhancement based on the local contrast factor; the noise suppression processing includes: dynamically adjusting the image smoothing parameters according to the feature extraction results of the clustering network, so as to improve the overall signal-to-noise ratio without losing the subtle defect features.

[0031] In this embodiment, the present invention achieves dynamic identification and trend prediction of touchscreen defects in the time dimension by designing a trajectory analysis network including a time encoder and a trend decoder. The time encoder adopts a dual-channel structure, where the main channel extracts defect features temporally through continuous 3×1 convolutional layers, normalization layers, and activation functions, while the secondary channel constructs a temporal correlation structure with 1×1 convolutional layers to capture potential connections between adjacent frames, thereby forming a deep temporal representation of the defect evolution trajectory. The trend decoder decodes and reconstructs the time-series features through continuous convolution and flattening operations, generating a time-series map reflecting the defect change trend, and realizing feedforward analysis of process offsets. By combining a composite loss function based on temporal consistency constraints and spatial distribution overlap, the model can simultaneously ensure the continuity of the defect trajectory and spatial matching accuracy, effectively suppressing false detections and drift. Furthermore, multi-scale enhancement and noise suppression strategies are introduced in the real-time detection stage. Brightness normalization is performed using the regional gradient change rate, and structural details are enhanced based on the local contrast factor. At the same time, image smoothing parameters are dynamically adjusted through clustering network feedback, so that the input data maintains a stable feature distribution under different imaging environments. This design not only significantly improves the system's ability to detect subtle defects and dynamic deviations, but also achieves highly robust detection under complex operating conditions. It can output defect warning information and process deviation suggestions in real time, thereby effectively improving the stability of the production line, detection accuracy, and early anomaly prediction capabilities.

[0032] Example 4 This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, after the steps of outputting defect warning information and process deviation suggestions, the system further includes: generating a defect distribution evolution report based on the defect warning information; the defect distribution evolution report includes: a defect occurrence frequency distribution map, a spatial clustering coefficient change curve, and a process deviation trend map; the report is used to analyze the recurring areas and location deviation trajectories of the same type of defect within a preset time window; when the defect clustering coefficient is detected to be continuously increasing or the deviation trajectory exceeds a threshold, the system automatically triggers equipment maintenance or process parameter calibration commands to achieve closed-loop quality control of the production line.

[0033] In this embodiment, the present invention enhances the detection system from a single defect identification module into an intelligent quality management unit with trend analysis and proactive control capabilities by adding a defect distribution evolution report generation step after outputting defect warning information and process deviation suggestions. This report comprehensively displays a defect frequency distribution map, a spatial clustering coefficient change curve, and a process deviation trend map, intuitively reflecting the repeated distribution areas and location change trajectories of the same type of defect within a preset time window, thereby revealing potential equipment deviations or process instability issues. Through dynamic calculation of the clustering coefficient and monitoring of the trend curve, the system can determine the diffusion or aggregation trend of defects during production. When a continuous increase in the clustering coefficient or an deviation trajectory exceeding a threshold is detected, equipment maintenance or process parameter calibration commands are automatically triggered, forming a closed-loop control mechanism of detection-analysis-adjustment. This design not only improves the predictability and adaptability of the touchscreen defect detection system but also reduces losses caused by manual inspections and delayed responses, transforming quality management from "post-event correction" to "pre-event prevention," significantly improving the stable operation efficiency of the production line and product yield.

[0034] Example 5 A machine vision-based touchscreen defect detection system, please refer to... Figure 2 Specifically, it includes a sample library construction module, a model training module, and a detection module; The system includes a sample library construction module for building a defect sample library based on defect distribution data from historical inspection results; a model training module for training a pre-built basic detection model based on the defect sample library until preset conditions are met, and then determining the trained basic detection model as the touchscreen defect detection model; wherein, the basic detection model includes: a clustering network and a trajectory analysis network; the clustering network is a deep residual network containing a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module containing a time series correlation structure; and a detection module for inputting real-time collected touchscreen defect data into the touchscreen defect detection model and outputting defect warning information and process offset suggestions.

[0035] In this embodiment, the present invention achieves intelligent and closed-loop management of the entire process of defect detection in touchscreen production lines by constructing a complete system architecture consisting of a sample library construction module, a model training module, and a detection module. The sample library construction module forms clustered sample libraries and trajectory sample libraries through structured extraction and classification of defect distribution data from historical detection results, providing stable and high-quality data support for model training. The model training module then adaptively learns the basic detection model, which includes clustering and trajectory analysis networks. The clustering network enhances the identification ability of complex surface defects through a deep residual structure with a multi-scale feature extraction mechanism, while the trajectory analysis network achieves trend learning of defect changes over time through dynamic tracking of time-series correlation structures. After model training, the detection module can quickly analyze real-time collected touchscreen defect data, outputting defect warning information and process deviation suggestions, thereby achieving data-driven self-learning detection and process feedback optimization without changing hardware conditions. This system architecture not only significantly improves the automation and accuracy of detection but also constructs an intelligent diagnostic mechanism for the production process through the two-way linkage of model training and real-time feedback, making quality control in touchscreen manufacturing more efficient, stable, and continuously optimized.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A machine vision based method for detecting defects in a touch screen, the method comprising: The method comprises the following steps: constructing a defect sample library based on defect distribution data in historical detection results; training a pre-constructed basic detection model based on the defect sample library until a preset condition is reached, and determining the trained basic detection model as a touch screen defect detection model; wherein the basic detection model comprises a clustering network and a trajectory analysis network; the clustering network is a deep residual network comprising a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module comprising a time series correlation structure; real-time collected touch screen defect data is input into the touch screen defect detection model, and defect early warning information and process offset suggestions are output. 2.The machine vision-based touch screen defect detection method of claim 1, wherein: The defect sample library comprises a clustering sample library and a trajectory sample library; the step of constructing a defect sample library based on defect distribution data in historical detection results comprises: obtaining defect distribution data in historical detection results; performing sliding sampling on the defect distribution data based on a preset size of a spatial window to obtain a local defect distribution map; after labeling a repeatedly occurring defect region in the local defect distribution map, a labeled data map is obtained; the labeled data map assigned with a first label and the local defect distribution map assigned with a second label are determined as samples in the clustering sample library; the labeled data map and corresponding time series data are determined as samples in the trajectory sample library; wherein the labeled data map is a feature, and the time series data is a label. 3.The machine vision based touch screen defect detection method of claim 1, wherein: The repeatedly occurring defect region comprises a bright spot region, a bubble region and a dark spot region; the area, position and labeled time interval of the repeatedly occurring defect region are randomly generated within a preset range.

4. The method of claim 1, wherein: The clustering network comprises three basic layers and a global pooling layer connected in sequence; each basic layer comprises two feature extraction blocks connected in sequence; each feature extraction block comprises a convolution layer, a multi-scale feature fusion layer and a spatial mapping layer connected in sequence.

5. The machine vision based method for detecting defects in a touch screen as claimed in claim 1, wherein: The trajectory analysis network comprises a time encoder and a trend decoder; the time encoder comprises three continuous feature extraction blocks, each of which comprises two parallel channels, the first channel comprises two continuous 3x1 convolution layers, a normalization layer and an activation function; the second channel is a time correlation structure comprising a 1x1 convolution layer; the trend decoder comprises two continuous 3x1 convolution layers and a flattening operation layer.

6. The machine vision-based method for detecting defects in a touch screen as claimed in claim 5, wherein: The loss function of the trajectory analysis network is based on the comprehensive evaluation of the time consistency constraint and the spatial distribution overlap degree of each data sample.

7. The machine vision based method for detecting defects in a touch screen as claimed in claim 1, wherein: The touch screen defect detection model comprises the trained clustering network and the trained trajectory analysis network; the step of inputting real-time collected touch screen defect data into the touch screen defect detection model to output defect early warning information and process offset suggestions comprises: inputting real-time collected touch screen defect data into the trained clustering network to obtain a local defect distribution map and a labeled data map; inputting the labeled data map into the trained trajectory analysis network to obtain a time series trend map; the data obtained by integrating the local defect distribution map and the time series trend map is determined as defect early warning information and process offset suggestions. 8.The method of claim 7, wherein: Before inputting the real-time collected touch screen defect data into the touch screen defect detection model, further comprising: performing multi-scale enhancement and noise suppression processing on the real-time collected touch screen image data; wherein the multi-scale enhancement processing comprises: brightness normalization based on regional gradient change rate and structure strengthening based on local contrast factor; the noise suppression processing comprises: dynamically adjusting the image smoothing parameter according to the feature extraction result of the clustering network, so as to improve the overall signal-to-noise ratio without losing subtle defect features. 9.The machine vision based method for detecting defects in a touch screen according to claim 1, wherein: After the step of outputting the defect early warning information and the process offset suggestion, further comprising: generating a defect distribution evolution report based on the defect early warning information; the defect distribution evolution report comprises: a defect frequency distribution graph, a spatial aggregation coefficient change curve and a process offset trend graph; the report is used to analyze the repeated occurrence area and position offset trajectory of the same type of defects within a preset time window; when the defect aggregation coefficient continues to rise or the offset trajectory exceeds the threshold value, the system automatically triggers the equipment maintenance or process parameter calibration instruction to realize the closed-loop quality control of the production line. 10.A machine vision based touch screen defect detection system, applied to the machine vision based touch screen defect detection method of any one of claims 1-9, characterized in that: The method comprises the following steps: The sample library construction module is used to construct a defect sample library based on defect distribution data in historical detection results; The model training module is used to train a pre-constructed basic detection model based on the defect sample library until a preset condition is reached, and the trained basic detection model is determined as a touch screen defect detection model; wherein the basic detection model comprises: a clustering network and a trajectory analysis network; the clustering network is a deep residual network containing a multi-scale feature extraction mechanism; the trajectory analysis network is a dynamic tracking module containing a time series correlation structure; the detection module is used to input real-time collected touch screen defect data into the touch screen defect detection model, and output defect early warning information and process offset suggestions.

Images