CN-121999283-A - Intelligent identification method for surface defects of single crystal material

CN121999283ACN 121999283 ACN121999283 ACN 121999283ACN-121999283-A

Abstract

The application provides an intelligent identification method of single crystal material surface defects, which comprises the steps of collecting single crystal material image data with defects on a plurality of surfaces to obtain a single crystal material image data set, establishing an initial YOLOv s model based on YOLOv s model, embedding an attention module into the initial YOLOv s model, adjusting characteristic channel weights in training set data to obtain YOLOv s-SE model, carrying out learning training on the YOLOv s-SE model to obtain a target detection model, detecting video streams on the surface of the single crystal material to be detected through the target detection model, identifying defect types and determining defect positions, selecting a lightweight network architecture to embed the characteristic weights into the attention module, improving the detection capability of small, medium and large size targets, setting a dynamic learning rate and an optimizer to combine with the early stop training model, realizing real-time video stream processing and result display storage, remarkably improving the real-time property and generalization capability of detection accuracy, being applicable to resource limited equipment, and ensuring production quality and efficiency.

Inventors

ZHAO WEILUN
WANG YAN
YIN CHANGYONG
PANG XINFU
YAN HONGKUI
LEI YANHUA

Assignees

沈阳工程学院

Dates

Publication Date: 20260508
Application Date: 20260117

Claims (9)

1. The intelligent identification method for the surface defects of the single crystal material is characterized by comprising the following steps of: acquiring single crystal material image data with defects on a plurality of surfaces to obtain a single crystal material image data set; Marking the defect area of each single crystal material image data, and distributing defect type labels to obtain training set data; Establishing an initial YOLOv s model based on YOLOv s model, embedding an attention module into the initial YOLOv s model, and adjusting the weight of a characteristic channel in training set data to obtain a YOLOv s-SE model; Setting training parameters, learning and training the YOLOv s-SE model according to the training parameters and training set data to obtain a target detection model, so that the video stream on the surface of the single crystal material to be detected is detected through the target detection model, the defect type is identified, and the defect position is determined.
2. The method for intelligently identifying the surface defects of the single crystal material according to claim 1, wherein the marking of the defect area of each single crystal material image data and the distribution of defect type labels are carried out to obtain training set data, specifically comprising the following steps: Introducing the single crystal material image by using a marking tool; Manually drawing a rectangular frame to cover a defect area in the single crystal material image; according to the defect type, distributing a corresponding defect type label for each rectangular frame, recording geometric shape morphological characteristics and space distribution data of defects, and generating a labeling file; And converting the annotation file into a preset format, and storing the annotation file in a designated path for model training and reading.
3. The method for intelligently identifying surface defects of single crystal materials according to claim 1, wherein the steps of establishing an initial YOLOv s model based on YOLOv s model, embedding an attention module into the initial YOLOv s model, and adjusting characteristic channel weights in training set data to obtain a YOLOv s-SE model comprise the following steps: Selecting YOLOv s model as basic layer structure, and embedding attention module in specific part; and carrying out global average pooling operation through the attention module, compressing channel characteristics into scalar quantities, inputting the channel characteristics into a full-connection layer, generating channel attention weights, adjusting characteristic distribution, strengthening important characteristic channel response, and obtaining the YOLOv s-SE model.
4. The intelligent identification method of single crystal material surface defects according to claim 3, wherein the embedding of the attention module in the specific part comprises the following steps: acquiring a level position from a multi-scale output path, processing the level position through a convolutional neural network, outputting position coordinates, and determining the position of a medium target detection branch; Inserting an attention module for the medium-size target detection branch to generate a medium-size target expression; inserting the attention module into a small target detection branch to enhance fine granularity feature focusing; inserting the attention module into a large target detection branch to strengthen large-range target perception; And inserting the attention module into a trunk feature expression layer, and inhibiting background interference through dynamic weight adjustment to improve the overall detection performance.
5. The intelligent identification method of single crystal material surface defects according to claim 3, wherein the global average pooling operation is performed by an attention module, channel characteristics are compressed into scalar quantities, the channel characteristics are input into a full connection layer, channel attention weights are generated, characteristic distribution is adjusted, and important characteristic channel responses are enhanced, and the intelligent identification method specifically comprises the following steps: Compressing the two-dimensional space features of each channel into a scalar through global average pooling operation to form a global feature description of the channel level, thereby extracting the most representative information; Inputting the compressed global feature description into a full-connection layer, generating a channel attention weight through nonlinear transformation, and adjusting the importance distribution of each channel to strengthen the response of the important feature channel.
6. The method of claim 1, wherein the training parameters include learning rate, optimizer, and Batch Size.
7. The method for intelligently identifying surface defects of single crystal material according to claim 6, wherein the setting training parameters specifically comprises: Setting a dynamic learning rate scheduling strategy, and calculating the learning rate through a cosine annealing formula; Selecting an optimizer to combine the self-adaptive learning rate mechanism and the weight attenuation strategy, determining the batch size, and calculating gradient update; and monitoring a loss function, processing positioning loss, classification loss and confidence loss, and adjusting the training progress.
8. The method for intelligently identifying surface defects of single crystal materials according to claim 7, wherein the setting of a dynamic learning rate scheduling strategy and the calculation of the learning rate by a cosine annealing formula specifically comprise: (1) Wherein the method comprises the steps of For the purpose of the rate of learning, For the minimum learning rate, T is the current round, and T is the total round number.
9. The method for intelligently identifying the surface defects of the single crystal material according to claim 1, wherein the method for detecting the video stream of the surface of the single crystal material to be detected through the target detection model, identifying the defect type and determining the defect position comprises the following steps: Collecting video streams of the surface of the single crystal material through a camera, extracting images to be detected from the video streams, and inputting the images to be detected into the target detection model; the target detection model identifies and detects a defect area in the image to be detected; After dividing the defect of the defect area by the defect type, identifying the crack refinement, identifying the scratch and the inclusion, determining the defect classification result; And transmitting the defect classification result to a user interface for real-time display, and storing the defect classification result in a storage device for subsequent analysis.

Description

Intelligent identification method for surface defects of single crystal material Technical Field The invention relates to the technical field of single crystal material detection, in particular to an intelligent identification method for surface defects of a single crystal material. Background At present, the defect detection of strontium titanate monocrystal mostly adopts the traditional manual visual detection mode, quality inspection engineers need to frequently carry out visual screening work in an environment with stronger light, and as the monocrystal material has unique optical reflection and refraction characteristics, when the traditional visual detection mode is adopted to face a monocrystal ingot, the defect false alarm and the great improvement of the missing report rate can occur, so that the crystal defect is only discovered when the later crystal is sliced, and the labor cost is increased and the crystal material is wasted. Therefore, a model capable of being identified is established for single crystal defect detection, but a unique and remote technical problem exists for defect identification requirements under complex industrial environments, namely, how to realize efficient extraction and accurate classification of multi-scale defect characteristics on embedded equipment with limited resources, and meanwhile, balance between instantaneity and detection precision is considered. In particular, single crystal material has various surface defect types including micro cracks, micro scratches, large area inclusions, etc., large defect size span, complex background noise, and embedded devices generally face severe limitations of computing power and memory resources. In this case, it is often difficult for the conventional object detection model to process multi-scale features and small object detection tasks simultaneously on low-power devices, especially in real-time video streaming, where the contradiction between the model inference speed and the detection accuracy is particularly prominent. In addition, the illumination condition of the images acquired by the industrial camera is different, the difference between the defect morphology and the spatial distribution is obvious, and how to dynamically adjust the characteristic weight under the limited resources to inhibit the background interference and focus the key defect area becomes a difficult problem to be solved urgently. Disclosure of Invention The invention provides an intelligent identification method for surface defects of single crystal materials, Mainly comprises the following steps: acquiring single crystal material image data with defects on a plurality of surfaces to obtain a single crystal material image data set; Marking a defect area of each single crystal material image data, and distributing defect type labels to obtain training set data, wherein the defect type labels comprise defect geometric shape morphological characteristics and space distribution data; Establishing an initial YOLOv s model based on YOLOv s model, embedding an attention module into the initial YOLOv s model, and adjusting the weight of a characteristic channel in training set data to obtain a YOLOv s-SE model; Setting training parameters, learning and training the YOLOv s-SE model according to the training parameters and training set data to obtain a target detection model, so that the video stream on the surface of the single crystal material to be detected is detected through the target detection model, the defect type is identified, and the defect position is determined. The technical scheme provided by the embodiment of the invention can have the following beneficial effects: The invention discloses a single crystal material surface defect detection method, which aims at the problem of accurately identifying various defects such as crack scratches and mingled service scenes in real time in industrial production, and the problem is derived from logic association challenges of low detection precision and weak generalization capability caused by insufficient defect diversity and authenticity in a complex environment. The method comprises the steps of obtaining a high-quality image data set containing various defect types, dividing the high-quality image data set into training verification and test sets according to a proportion, combining a manual labeling rectangular frame and a class label to ensure accurate and various data, selecting a lightweight network architecture to embed a attention module to dynamically adjust characteristic weights, improving the detection capability of small, medium and large-size targets, setting a dynamic learning rate and an optimizer to combine an early-stop training model, realizing real-time video stream processing and result display storage, constructing a confusion matrix to calculate accuracy recall rate and average accuracy quantization performance, adjusting the balance class distribution of the data set, and op