Mobile phone silicone shell appearance detection system based on optical image sensor

Abstract

The application discloses a mobile phone silica gel shell appearance detection system based on an optical image sensor, relates to the technical field of appearance defect detection and intelligent visual identification, and has the advantages that a high-quality multi-channel image is acquired through multi-angle synchronous acquisition by an image acquisition module; adaptive adjustment of a regional light source is realized based on reflectivity perception and a gray gradient change rate by a dynamic light field regulation module; pixel-level defect classification is realized by fusing edge features, texture changes and a deep learning model by an image processing and defect identification module; a material identification and shielding module identifies non-silica gel regions through multi-dimensional spectral feature recognition and excludes interference; a time sequence defect evolution analysis module models a defect evolution trajectory; a trend prediction module realizes expansion trend prediction of potential cracks and fatigue regions; and a control feedback module realizes intelligent judgment and response in linkage with a manufacturing system. The system can realize high-precision, high-robustness automatic defect detection and risk prediction and is suitable for online quality monitoring of large batches of flexible products.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of appearance defect detection and intelligent visual recognition, and particularly relates to a mobile phone silica gel shell appearance detection system based on an optical image sensor. BACKGROUND

[0002] With the large-scale popularization of smart phones, the market has higher requirements for the appearance quality of mobile phone accessory products, especially mobile phone protective cases. Mobile phone silica gel shells have become the mainstream protective case material due to their softness, impact resistance and easy processability. However, the appearance detection of mobile phone silica gel shells still mainly relies on manual visual inspection, which is not only low in efficiency and high in cost, but also limited by manual experience, and has problems such as missed detection and false detection, which is difficult to meet the quality control requirements of consistency and accuracy in large-scale production.

[0003] Although some existing automatic detection methods introduce image processing technology, they are mostly limited to static photography and simple image comparison, and lack accurate recognition ability for complex backgrounds, multi-angle deformation and small defects (such as bubbles, scratches, black spots, etc.), resulting in unstable good product rate evaluation. In addition, the existing systems do not fully adapt to the optical properties of silica gel materials, and the image acquisition quality is limited, which further affects the recognition accuracy.

[0004] Therefore, an automatic detection system based on an optical image sensor is urgently needed, which can combine multi-light source image enhancement, intelligent defect recognition algorithm and multi-dimensional data fusion strategy to realize efficient and accurate detection of appearance defects of mobile phone silica gel shells, improve the overall quality control level, and meet the actual needs of industrial production. SUMMARY

[0005] The purpose of the present application is to provide a mobile phone silica gel shell appearance detection system based on an optical image sensor to solve the problems in the background art.

[0006] In order to achieve the above purpose, the present application provides the following technical scheme: a mobile phone silica gel shell appearance detection system based on an optical image sensor, comprising:

[0007] A dynamic light field regulation module is used to adaptively adjust the light emitting intensity and incident angle of the multi-light source according to the light reflection characteristics of different parts of the silica gel shell to enhance the imaging contrast of small defects in the image;

[0008] An image acquisition module acquires multi-channel image data of the silica gel shell to be tested in a multi-angle synchronous capture manner under the regulation of the dynamic light field;

[0009] A material recognition and shielding module is used to construct a silica gel material recognition model based on material spectral characteristics, and to exclude or mark non-silica gel materials in the image;

[0010] a time-series defect evolution analysis module that models the change trend of the same position in consecutive detection batches, identifies the evolution path of the micro-defects and determines whether it has reached the warning threshold;

[0011] an image processing and defect identification module that extracts edge features, shape information and texture changes in the standardized image, and identifies the defect type based on a deep learning model;

[0012] a trend prediction module that sub-pixel level locates the identified defects in the image, and judges the formation trend of potential cracks and material fatigue areas in combination with historical features;

[0013] a control feedback module that transmits the defect detection results to the manufacturing control system in real time to realize online judgment, marking and dynamic repair suggestion output.

[0014] Preferably, the dynamic light field regulation module specifically comprises:

[0015] a plurality of annular adjustable light source arrays are configured and symmetrically arranged around the image acquisition area, and the light source array supports independent control of light intensity and angle direction in different regions;

[0016] a surface reflectivity sensing unit based on a spectral feedback mechanism is provided to acquire the diffuse reflection and specular reflection ratio of different regions of the silica shell in real time;

[0017] a regional light field adaptation model is constructed based on the reflection ratio, and the illumination parameters of the corresponding region light source are dynamically adjusted by calculating the local maximum gray gradient change rate.

[0018] Preferably, the construction of the regional light field adaptation model comprises:

[0019] the surface reflectivity values collected by the sensing unit are regionally divided, and a reflectivity spatial distribution matrix is established;

[0020] for each to-be-tested region, the gray change surface of the corresponding region in the continuous image frames is extracted, and the local maximum gray gradient change rate is calculated;

[0021] the local maximum gray gradient change rate is taken as the dimming driving factor, the irradiation angle and brightness intensity of the light source in the region are corrected in real time combined with the preset illumination response curve, and a regional level dynamic illumination regulation strategy is formed.

[0022] Preferably, the material identification and shielding module specifically comprises:

[0023] the spectral response characteristics of each pixel region in the multi-channel image are obtained, and a corresponding multi-dimensional material reflection feature vector is constructed;

[0024] Based on the obtained silica gel material spectrum library, a high-dimensional feature clustering algorithm is used to classify the material of the image region, and the region with low matching degree with the silica gel feature is identified;

[0025] The non-silica gel region with a matching degree lower than the set threshold is automatically marked or shielded.

[0026] Preferably, the timing defect evolution analysis module comprises:

[0027] A timing defect data association model based on position index is established to number and align the defect images at the same physical coordinates in consecutive detection batches;

[0028] The morphological feature vector and gray scale distribution parameter of the defect region are extracted to construct a defect evolution trajectory sequence;

[0029] The defect change trend is modeled by a timing convolution network with residual structure, and the defect evolution speed is calculated by combining the feature offset in the time window;

[0030] When the defect growth rate exceeds the set evolution threshold, the early warning mechanism is triggered and the potential failure risk position is output.

[0031] Preferably, the image processing and defect identification module comprises:

[0032] The edges of the normalized image are enhanced and the background is suppressed for pretreatment, and the edge features and local shape contours in the image are extracted;

[0033] The texture change information at different resolutions is extracted using a multi-scale image pyramid to generate a fusion feature map;

[0034] The fusion feature map is input into a convolutional neural network optimized based on an attention mechanism to identify the defect region and perform pixel-level classification, and output the defect type label.

[0035] Preferably, the trend prediction module comprises:

[0036] Based on a high-resolution image interpolation reconstruction method, the identified defect edge is sub-pixel level positioned, and a high-precision defect contour model is generated;

[0037] The morphological features at the same position in the defect history image sequence are separated into long-term and short-term trends, and the evolution direction and growth rate features are extracted;

[0038] A timing prediction network that combines position encoding and defect intensity weight is constructed to regress model the future development trend of potential cracks or material fatigue regions;

[0039] The defect future expansion probability map is output, and whether it constitutes a structural risk region is judged in combination with the system tolerance threshold.

[0040] Preferably, wherein the trend prediction module further comprises: based on multi-source image features, detecting environmental parameters and historical defect behaviors, constructing an improved Bayesian belief network; the Bayesian network introduces an attention weight mechanism to dynamically adjust the prior contribution of multi-input features; combined with the defect morphology evolution trajectory and environmental disturbance factor, joint probability inference is carried out to output the defect future expansion probability map.

[0041] In the above technical solution, the technical effects and advantages provided by the present application are:

[0042] 1. By introducing dynamic light field regulation, multi-channel image processing, deep learning recognition and time series defect evolution modeling, the present application realizes high-precision and automatic detection of appearance defects of mobile phone silica gel shells. Compared with traditional manual visual or static image analysis methods, the present application has significant improvement in edge recognition, weak defect appearance, multi-angle adaptability and material shielding ability, and is especially suitable for detection scenes with low contrast, complex texture and various deformations.

[0043] 2. The present application constructs a defect trend prediction model by combining time series convolution network and Bayesian belief network, breaks through the limitation of the prior art that can only "detect current defects", and for the first time realizes defect evolution path modeling and future expansion probability prediction, has the ability to early warn potential cracks and material fatigue areas, significantly enhances the intelligent, forward-looking and closed-loop correction ability of quality control, and is suitable for online detection needs in high-reliability electronic product manufacturing. BRIEF DESCRIPTION OF DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art based on these drawings.

[0045] Figure 1 The system module mind map of the present application. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0047] Embodiments, please refer to Figure 1As shown, the mobile phone silica gel shell appearance detection system based on an optical image sensor in the embodiment comprises:

[0048] A dynamic light field regulation module is configured to adaptively adjust the light intensity and incident angle of the multi-light source according to the light reflection characteristics of different parts of the silica gel shell, so as to enhance the imaging contrast of the image to the micro defects.

[0049] An image acquisition module is configured to acquire multi-channel image data of the silica gel shell to be detected in a multi-angle synchronous capture manner under the regulation of the dynamic light field.

[0050] A material identification and shielding module is configured to construct a silica gel material identification model based on the material spectral characteristics, and to exclude or mark the non-silica gel material in the image.

[0051] A time sequence defect evolution analysis module is configured to model the defect change trend of the same position in the continuous detection batches, identify the evolution path of the micro defects, and judge whether the micro defects reach the warning threshold.

[0052] An image processing and defect identification module is configured to extract edge features, shape information and texture changes in the standardized image, and identify the defect type based on a deep learning model.

[0053] A trend prediction module is configured to sub-pixel level position the identified defects in the image, and judge the formation trend of the potential cracks and material fatigue areas in combination with the historical characteristics.

[0054] A control feedback module is configured to transmit the defect detection results to the manufacturing control system in real time, so as to realize online judgment, marking and dynamic repair suggestion output.

[0055] In a preferred embodiment of the present application, the dynamic light field regulation module is designed to adapt to the imaging quality fluctuation problem caused by the changes of the surface reflection characteristics of the silica gel shell with the area, angle and curvature. Since the silica gel shell of the mobile phone is usually flexible, semi-transparent or matte, it has obvious mixed characteristics of diffuse reflection and specular reflection on the surface, and there are slight differences in material, color, texture, etc. between different batches of products, which often causes problems such as insufficient imaging contrast, blurred edges, unclear defects, etc. Therefore, the dynamic light field regulation technology driven by the spectral feedback mechanism can improve the imaging recognition ability of the image to the micro defects (such as light-colored scratches, low-contrast bubbles, small impurities, etc.) while ensuring uniform illumination, thereby greatly improving the stability and sensitivity of the detection system.

[0056] The module is preferably configured with multiple sets of annular adjustable light source arrays, which are symmetrically arranged around the image acquisition area to ensure controllable illumination from multiple angles. Each set of light source array is composed of several high-brightness LED light sources, and integrates a current driving circuit supporting independent control. Each array is divided into multiple sub-regions, and the light intensity of each sub-region can be independently adjusted according to the control instructions. The arrangement structure ensures that the light can be customized according to the local characteristics of the surface of the silicone shell being detected during the detection process to avoid overexposure or shadow coverage.

[0057] In order to realize the adaptability of light field regulation, the present application designs a surface reflectivity sensing unit based on a spectral feedback mechanism. The unit includes a short-wave infrared (SWIR) and visible light composite spectrum receiver, which can acquire the reflected light intensity of each detection area of the measured silicone shell in real time. By comparing the diffuse reflection component and the specular reflection component at different wavelengths, the system can calculate the reflectance ratio (reflectivity vector) of each detection area as the basis for formulating subsequent light regulation strategies.

[0058] Specifically, the system synchronously acquires the reflectance spectrum in the pre-scanning stage, uses the Fourier transform method to extract the ratio between the diffuse reflection main peak and the specular reflection peak value, and thus constructs the light reaction characteristics of each detection area. The ratio data will be used as an input factor for region modeling.

[0059] In order to convert the above reflectivity information into a dynamic lighting strategy, the present application further proposes a regional light field adaptation model, which is realized by the following steps:

[0060] Firstly, the reflectance ratio data acquired by the surface reflectivity sensing unit is divided according to the spatial coordinates of the detection image, and a two-dimensional spatial mapping relationship is formed. The system divides the image acquisition area into several sub-regions (such as 5x5 pixels per sub-block), and generates a corresponding reflectivity spatial distribution matrix for each region, wherein each element represents the reflectance ratio of the sub-block.

[0061] In the continuous image frames obtained by the image acquisition module, the gray value distribution of the corresponding region is extracted, and a three-dimensional gray surface is constructed. The derivative of the change of the gray surface on the time axis is calculated, and the local maximum gray gradient change rate is extracted. The gradient change rate represents the sensitivity of the edge response of the region, which is used to judge the strength of the current light on the appearance of the small defects.

[0062] The above gray gradient change rate is used as the driving factor of light regulation and input into the light field control engine. The engine determines the most suitable light source irradiation angle and brightness intensity for the current region according to the preset light response curve model, and sends light adjustment instructions to the corresponding light source array sub-region in real time, thereby forming a dynamic light regulation strategy at the regional level.

[0063] Through the above technical means, the system can dynamically adjust the incident light angle and intensity of each region according to the local optical behavior of the surface of the silica gel shell, so that the weak edge features in the image have more significant gray level transitions, and the signal-to-noise ratio of defect recognition is improved.

[0064] In a preferred embodiment of the present application, the image acquisition module cooperates with the dynamic light field regulation module to work synchronously, and adopts a multi-angle synchronous capture mode to acquire multi-channel image data of the silica gel shell to be detected. Due to the flexible concave-convex, transparent or semi-transparent structure on the surface of the silica gel shell, imaging under a single angle or fixed lighting condition is prone to shadow obstruction, glare interference or surface detail loss. Therefore, the module is designed through a spatial multi-view layout and a time synchronous triggering mechanism.

[0065] Specifically, the image acquisition module includes a plurality of high-definition industrial cameras, which are arranged at a certain angle around the detection area to form a multi-view imaging array. Each camera has multi-channel image acquisition capability and can simultaneously acquire images of different spectral channels such as RGB, infrared (IR) and ultraviolet (UV) to enhance the level expression capability of surface features.

[0066] During detection, the image acquisition module receives the lighting state signal from the dynamic light field regulation module to ensure that each camera is synchronously exposed and collected under the current lighting strategy. Through hardware triggering or time controller synchronous driving, the image time consistency is ensured, and feature deviation caused by phase difference is avoided. The multi-channel images collected will be used as the input basis for subsequent defect recognition, reflectivity modeling and image fusion, providing high-resolution, high-consistency and low-noise visual data support for the system.

[0067] In the present application, in order to solve the problems of foreign matter interference, background mixing and non-target material misrecognition of mobile phone silica gel shells during detection, a material identification and shielding module is specially provided, which has the following main functions: based on the material spectral response characteristics of each pixel or region in the image, accurately identifying the image region corresponding to the silica gel material, and automatically marking or shielding the non-silica gel material (such as dust, oil stains, foreign matter debris, transportation residues or background objects, etc.), thereby improving the accuracy and stability of the subsequent defect recognition algorithm.

[0068] Traditional defect detection methods based on image processing mostly rely on gray level, texture or edge features to analyze target regions. However, in actual production and detection scenarios, the following problems often exist:

[0069] The surface of a mobile phone silica gel shell is prone to attach small foreign matters such as dust and fluff fibers, which have different reflection characteristics from silica gel and are prone to misjudgment;

[0070] Part of the detection image contains non-target regions such as detection clamps, pipeline backgrounds, etc. If not removed, it will mislead the feature extraction of the deep learning model.

[0071] Some foreign matters are similar in appearance to defects, and it is difficult to distinguish them by conventional visual features, such as transparent plastic fragments or residual glue adhered to the surface.

[0072] Therefore, it is difficult to exclude interference from the material level by relying only on geometric or brightness features. The present application introduces the spectral feature dimension at the image acquisition level by constructing a silica gel material recognition model, and realizes the automatic screening and shielding of the target material region in the image from the material attribute.

[0073] The material recognition and shielding module according to the present application specifically comprises the following sub-function components:

[0074] The spectral feature acquisition unit extracts the light intensity response value of each pixel or region in the image under different wavebands based on the multi-channel image data (such as visible light RGB, near-infrared NIR, short-wave infrared SWIR, etc.) provided by the image acquisition module, forming a multi-dimensional spectral response vector reflecting the intrinsic properties of the material. For example, the combination of the reflection intensities of the same pixel under the RGB, IR, and SWIR channels constitutes its feature vector [R, G, B, IR, SWIR], which has high distinguishability on different materials.

[0075] The silica gel material modeling unit establishes a statistical feature model (such as mean vector and covariance matrix) of silica gel material under each channel spectrum by offline acquisition of multiple typical silica gel shell samples, forming a standard silica gel material spectral template library. In actual detection, the system compares the spectral vector of the region to be detected with the silica gel samples in the template library through high-dimensional vector similarity algorithms (such as cosine similarity or Mahalanobis distance) to calculate the matching degree.

[0076] The material classification and mask generation unit divides the detection image into several sub-regions and classifies them according to the matching degree with the silica gel template. If the matching degree of a certain region is lower than a set threshold (such as 90%), it is determined as a non-silica gel material region, and the region is marked in the mask image. The mask image is in binary form, with silica gel regions marked as "1" and non-silica gel regions marked as "0". This mask will be used for feature extraction and shielding in the subsequent defect recognition stage, i.e. only the effective regions in the mask are processed, and the non-target material regions are ignored.

[0077] The material shielding and image fusion unit finally generates a material mask image that is pixel-level fused with the original image to form a structured input image. Before inputting the image into the defect recognition model, most of the material interference regions have been removed by the shielding mechanism, avoiding the model from learning irrelevant objects and improving the consistency and generalization ability of the training and recognition stages.

[0078] In the mobile phone silica gel shell appearance detection system based on an optical image sensor provided by the present application, in order to realize accurate modeling and risk prediction of the dynamic change trend of the micro defect, a time sequence defect evolution analysis module is specially set. The module breaks through the limitation of traditional defect detection which is limited to single-frame static judgment, and realizes the trend judgment and early warning of the defect in the production process by combining image time sequence modeling, feature extraction, neural network analysis and early warning logic.

[0079] The module is suitable for tracking the continuous detection data of the same product area between production batches, analyzing whether the defect has growth, expansion or evolution trend, and providing data support for quality control and defect intervention.

[0080] In order to ensure the comparability of defect images in different detection batches, a time sequence defect data correlation model based on position index is first established. Based on the two-dimensional coordinates provided by the image acquisition module and the product positioning mechanism, the defect images of the same physical position (such as a certain edge corner, hole periphery or key groove) are numbered and spatially aligned.

[0081] Specifically, the system uses image coordinate calibration and product feature point anchoring technology, combines template matching and affine transformation algorithm, realizes image registration between batches. By constructing a position index table, the index and calling of images of the same position defect at different time points are realized, providing basic data support for subsequent time sequence modeling.

[0082] After obtaining the aligned defect image sequence, the system extracts multi-dimensional features of the defect area at each time point to form a unified defect representation vector. The representation includes but is not limited to the following contents:

[0083] Morphological feature vector: including the area, aspect ratio, boundary curvature, edge sharpness of the defect area, etc.

[0084] Gray distribution parameters: including the average gray value, gray variance, gray gradient direction distribution of the region, etc.

[0085] Structural texture features: such as LBP (Local Binary Pattern), HOG (Histogram of Oriented Gradients) and other texture descriptors.

[0086] All features are arranged in time sequence to form a set of defect evolution trajectory sequence tensors, i.e. time as the first dimension, features as the second dimension, forming a three-dimensional data structure for deep model input.

[0087] In order to identify the evolution trend of the defect from the time sequence features, the present application adopts a time sequence convolution network (Temporal Convolutional Network, TCN) with residual structure for modeling.

[0088] The model has the following characteristics:

[0089] Causal convolution structure: maintain time order, ensure that the current prediction only depends on historical information;

[0090] Multi-scale convolution kernel: capture local and global change patterns;

[0091] Residual connection mechanism: improve the stability and gradient transmission efficiency of deep network, and avoid overfitting;

[0092] Sliding window structure: make the model can be updated in real time in the production process, has good online analysis ability.

[0093] The model input is the defect evolution trajectory tensor, which is processed by multi-layer one-dimensional convolution, ReLU activation and normalization, and the output is the defect evolution trend function, that is, the fitting curve of the defect severity in the time dimension.

[0094] In order to quantify the degree of defect evolution, the system calculates the change rate of defect features in each time window. Specifically:

[0095] Set the time window length to n frames, extract the sequence of defect morphology features in the window;

[0096] Calculate the Euclidean distance or cosine distance for each pair of adjacent frames to form a feature offset sequence;

[0097] Take the first derivative of the sequence as the estimated value of the defect evolution speed.

[0098] The evolution speed is used as a quantitative indicator to measure the growth rate of defects, and is compared with the preset threshold. If the speed value in consecutive frames is continuously higher than the threshold, it is considered as abnormal growth, and the system will trigger the warning mechanism, mark the position as a potential failure risk area, and output to the manufacturing execution system (MES) or artificial review module.

[0099] In the mobile phone silicone shell appearance detection system based on optical image sensor proposed in the present application, the image processing and defect recognition module undertakes the core image analysis and defect positioning function. Considering that there are complex visual features such as flexible deformation, low contrast texture, shallow scratches and the like on the surface of the silicone shell, traditional static image filtering and edge detection methods are often difficult to stably extract the defect area, therefore, this module combines multi-scale image structure extraction, deep feature modeling and pixel-level classification algorithm, effectively improving the robustness and detection accuracy of the system.

[0100] Firstly, the standardized images collected are preprocessed to highlight the structural information in the images and suppress the interfering background in the non-defect areas. An improved Canny edge detection algorithm combined with guided filtering is used to enhance the image edges, and an adaptive background suppression algorithm based on histogram equalization is used to compress the global brightness of the image, highlighting the edge profile and local shape features. This preprocessing step can improve the boundary clarity without damaging the image details, optimizing the subsequent feature extraction effect.

[0101] Since the defects of the mobile phone silica gel shell may exhibit differences at different scales (such as micro-bubbles and large-area scratches have different spatial distribution characteristics), this module uses a multi-scale image pyramid structure to process the images. The specific method includes downsampling the image to generate multiple resolution levels, and independently extracting texture change information such as gradient direction, gray difference, and frequency domain texture response in each level. Then, using a feature fusion mechanism, the information extracted at each scale is merged by channel to construct a fused feature map, achieving a balance between full-size perception and local detail preservation.

[0102] The fused feature map is sent as input to a convolutional neural network model (CNN) optimized based on attention mechanism. This model introduces channel attention and spatial attention modules based on traditional convolutional layers to enhance the response weight of the defect area and suppress the interference contribution of the background area. The network structure includes multiple layers of convolution, normalization, ReLU activation, and pooling operations, and finally outputs a pixel-level classification map that marks the specific boundary position and type label (such as scratch, impurity, bubble, crack, etc.) of the defect.

[0103] During model training, a cross-entropy loss function is used and a Dice coefficient optimization strategy is introduced to address the detection bias caused by class imbalance. The model can be adapted to different manufacturers, colors, and texture features of silica gel shells through transfer learning and continuous training, achieving wide adaptability.

[0104] Table 1: Comparison of defect recognition accuracy between the present invention and traditional detection methods

[0105]

[0106] In the mobile phone silica gel shell appearance detection system based on optical image sensors proposed in the present invention, the trend prediction module is used for deeper dynamic analysis and evolution trend judgment of the identified defects, especially suitable for detecting cracks, material fatigue, and other latent, extended structural defects. This module not only evaluates the current state of the defect, but also combines historical image information and environmental factors to model its possible development path, assisting the system in risk prediction and maintenance decision-making, which is a key technical component for building a transition from "detection" to "prediction".

[0107] Firstly, sub-pixel level positioning is performed on the defect area output by the detection module. The process is based on a high-resolution image interpolation reconstruction method, and the original image is interpolated and reconstructed by a Laplacian pyramid to improve the image resolution. Then, fine profile extraction is performed on the defect edge in the enhanced image, and a high-precision defect profile model is generated by combining the Canny edge operator and sub-pixel fitting curve (such as ellipse fitting, Bezier curve interpolation).

[0108] This profile model serves as an important basis for subsequent trend modeling, ensuring the consistency and traceability of the defect boundary in the time series during the analysis process.

[0109] The system calls the images extracted from the same defect area in multiple historical detection batches to construct a defect image sequence. To improve the modeling accuracy, the historical images are first spatially aligned and intensity normalized. Then, the morphological features of the defect are extracted in the sequence, including area change, edge gradient, concave-convex depth, and boundary morphological change rate.

[0110] Next, the long-short-term trend separation algorithm is used to decouple the above feature sequence. Based on the improved Holt-Winters filter, the morphological change is decomposed into:

[0111] Short-term change component: captures minor disturbances or occasional effects in the production process;

[0112] Long-term change component: reveals the structural trend of defect continuous evolution, such as crack expansion and fatigue deformation.

[0113] The evolution direction vector and growth rate index extracted finally are used as the main features of trend analysis, providing structured data for the next step of time series modeling.

[0114] To predict the potential cracks or fatigue areas, the invention constructs a time series prediction network that integrates structural features and spatial location information. The network is based on LSTM (Long Short-Term Memory) structure and introduces the following two innovations:

[0115] Position encoding mechanism: encode the position vector of the defect area in the image coordinates into a position attention matrix to guide the network to focus on the possible stress concentration areas;

[0116] Defect intensity weight embedding: sort the morphological evolution features by importance, embed weight coding to enhance the model's learning ability for high-risk defect dimensions.

[0117] The model output is a defect future expansion probability map, which reflects the direction and intensity of crack possible expansion in the form of a pixel-level heat map. Combined with the structural tolerance threshold set by the system, it can be directly used for failure prediction and product screening.

[0118] On the basis of the trend prediction network, in order to introduce uncertainty modeling and multi-factor fusion capability, the application further designs an improved Bayesian belief network for joint reasoning and risk quantitative evaluation.

[0119] The Bayesian network has the following technical features:

[0120] Multi-source feature input: the input includes defect evolution trajectory, image feature, environmental disturbance factor (such as temperature, vibration, humidity), product batch information, etc.

[0121] Dynamic prior adjustment mechanism: the attention mechanism is used to learn and update the prior weight of the input feature, realizing adaptive judgment in different detection scenes.

[0122] Joint probability reasoning model: combined with Bayesian rules and conditional probability inference, a reasoning link structure with "defect→evolution path→risk level" as the core is constructed.

[0123] Output probability graph: the model finally outputs a defect expansion probability graph, and labels the structural risk level of each region, supporting intelligent early warning and maintenance intervention strategy formulation of the system.

[0124] Table 2: Comparison of the defect evolution risk warning capability of the trend prediction module of the application

[0125]

[0126] In the application, the control feedback module serves as a bridge between the system and the manufacturing execution unit, and undertakes the information interaction and closed-loop control function of the defect detection result. Based on high-speed data bus and communication protocol (such as Modbus, OPCUA or industrial Ethernet), the module transmits the recognized defect type, position coordinates and severity in the form of structured data to the upper manufacturing control system in real time. The system can automatically judge whether the product is qualified according to the set threshold, and link the sorting mechanism or alarm device, realizing fast response and intelligent shunting.

[0127] In addition, the control feedback module has a built-in repair strategy suggestion library, which can output targeted processing suggestions such as marking, repair, reinforcement, mold replacement, etc. according to the defect type and evolution trend. The module also supports the feedback results to be input into the detection system model in reverse, for optimizing the subsequent detection strategy, realizing adaptive learning and dynamic adjustment of parameters, and thus building a real detection-control-optimization closed-loop system.

[0128] Table 3: Comparison of the material identification and shielding capability of the system of the application in complex background

[0129]

[0130] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A mobile phone silicone case appearance detection system based on an optical image sensor, characterized in that: Comprise: A dynamic light field regulation module for adaptively adjusting the intensity and angle of incidence of multiple light sources based on the light reflection characteristics of different parts of the silica shell to enhance the imaging contrast of images to minor defects; An image acquisition module that acquires multi-channel image data of the silica shell under test in a multi-angle synchronous capture mode under the dynamic light field regulation; A material identification and shielding module for constructing a silica material identification model based on material spectral characteristics to exclude or mark non-silica materials in the image; An image processing and defect identification module for extracting edge features, shape information, and texture changes in the standardized image and identifying defect types based on a deep learning model; A trend prediction module for sub-pixel level positioning of identified defects in the image and judging the formation trend of potential cracks and material fatigue areas, including: Based on a high-resolution image interpolation reconstruction method, the edges of the identified defects are located at the sub-pixel level, and a high-precision defect contour model is generated; Separate long-term and short-term trends of morphological features at the same position in the defect history image sequence to extract evolution direction and growth rate features; Construct a time series prediction network that combines position encoding and defect intensity weight, and based on the extracted evolution direction and growth rate features, regress model the future development trend of potential cracks or material fatigue areas; Output the defect future expansion probability map and judge whether it constitutes a structural risk area based on the system tolerance threshold; A time series defect evolution analysis module that models the defect trend at the same position in consecutive detection batches, identifies the evolution path of minor defects, and judges whether it has reached the warning threshold; A control feedback module for real-time transmission of defect detection results to the manufacturing control system to realize online judgment, marking, and dynamic repair suggestion output.

2. The optical image sensor based cell phone silicone case appearance inspection system of claim 1, wherein: The dynamic light field regulation module specifically includes: Configured with multiple sets of annular adjustable light source arrays arranged symmetrically around the image acquisition area, the light source arrays support independent control of light intensity and angle direction in different zones; Provided with a surface reflectivity perception unit based on a spectral feedback mechanism for real-time acquisition of the diffuse reflectance and specular reflectance ratio of different regions of the silica shell; Based on the reflectance ratio, a regional light field adaptation model is constructed to dynamically adjust the lighting parameters of the corresponding region light source by calculating the local maximum gray gradient change rate.

3. The optical image sensor based cell phone silicone case appearance inspection system of claim 2, wherein: The construction of the regional light field adaptation model includes: Divide the surface reflectance ratio collected by the perception unit into regions and establish a reflectance spatial distribution matrix; For each region to be tested, extract the gray change surface of the corresponding region in consecutive image frames and calculate its local maximum gray gradient change rate; Use the local maximum gray gradient change rate as the dimming driving factor to real-time correct the illumination angle and brightness intensity of the light source in the region based on the preset light response curve to form a dynamic light regulation strategy at the regional level.

4. The optical image sensor based cell phone silicone case appearance inspection system of claim 1, wherein: The material identification and shielding module specifically includes: Acquire the spectral response characteristics of each pixel region in the multi-channel image and construct the corresponding multi-dimensional material reflection feature vector; Based on the obtained silica gel material spectrum library, a high-dimensional feature clustering algorithm is used to classify the material of the image region, and the region with low matching degree with the silica gel feature is identified; The non-silica gel region with a matching degree lower than the set threshold is automatically marked or shielded.

5. The optical image sensor based cell phone silicone case appearance inspection system of claim 1, wherein: The time sequence defect evolution analysis module includes: A time sequence defect data association model based on position index is established to number and align the defect images at the same physical coordinates in consecutive detection batches; The morphological feature vector and gray distribution parameter of the defect region are extracted to construct a defect evolution trajectory sequence; Through a time sequence convolution network with residual structure, the defect change trend is modeled, and the defect evolution speed is calculated by combining the feature offset in the time window. When the defect growth rate exceeds the set evolution threshold, the early warning mechanism is triggered and the potential failure risk position is output.

6. The optical image sensor based cell phone silicone case appearance inspection system of claim 1, wherein: The image processing and defect identification module includes: Edge enhancement and background suppression preprocessing are performed on the standardized image to extract the edge features and local shape contours in the image; Multi-scale image pyramid is used to extract texture change information at different resolutions to generate a fusion feature map; The fusion feature map is input into a convolutional neural network optimized based on an attention mechanism to identify the defect region and perform pixel-level classification, and the defect type label is output.

7. The optical image sensor based cell phone silicone case appearance inspection system of claim 1, wherein: The trend prediction module further includes: based on multi-source image features, detection environment parameters and historical defect behaviors, an improved Bayesian belief network is constructed; the improved Bayesian belief network introduces an attention weight mechanism to dynamically adjust the prior contribution of multiple input features; combined with the defect morphological evolution trajectory and environmental disturbance factors, joint probability inference is performed to output a defect future expansion probability map.

Images