Optical mirror surface defect detection method based on deep learning

Abstract

The invention relates to the field of computer vision and industrial detection, and provides an optical mirror surface defect detection method based on deep learning in order to solve the technical problems of low optical mirror surface defect detection precision and weak generalization ability in a small sample scene. The method comprises the steps that firstly, a space-channel (SC) collaborative data enhancement strategy is adopted, and a training sample is expanded to be 9 times of the original scale through spatial dimension rotation, mirror image overturning and channel dimension gradient feature fusion; then constructing a deep network architecture comprising 34 layers of cascade modules, adopting a residual connection mechanism and channel decoupling convolution collaborative design, and embedding a spatial Dropout layer; and finally, inputting the enhanced data set into the network for training to realize classified detection of three defects of pocking marks, scratches and dust on the optical mirror surface. Small sample statistical deviation is relieved through SC enhancement, deep network gradient attenuation is inhibited by means of a dynamic residual structure, the feature cross-layer transmission efficiency is improved, the generalization ability of the model is enhanced in combination with a Dropout layer, the accuracy and robustness of optical mirror surface defect detection are effectively improved, and the industrial-grade detection requirements of optical elements in a high-power laser device are met.

Description

Technical Field

[0001] This invention relates to the intersection of computer vision and industrial optical inspection, specifically a method for accurately detecting three types of defects—pockmarks, scratches, and dust—on the surface of large-size optical mirrors in high-power laser devices, based on deep learning technology, suitable for small sample scenarios. This method can meet the dual requirements of industrial-grade optical component quality control for both detection accuracy and efficiency. Background Technology

[0002] As a core component of high-power solid-state laser systems (such as inertial confinement fusion ICF devices), optical mirrors are susceptible to damage from surface defects (scratches, pits, dust, etc.), which can directly alter the energy distribution of the laser beam, causing a sudden increase in local energy density, leading to thermal deformation or even cracking of the mirror, seriously threatening the system's operational stability and laser output performance. Even micron-level scratches can cause fatal malfunctions.

[0003] With the surge in demand for large-aperture optical components from ICF devices in my country, existing detection technologies have gradually revealed significant shortcomings: traditional manual inspection is inefficient, lacks standardized judgment criteria, cannot meet the needs of industrialized batch inspection, and is insufficient in identifying micron-level defects; in traditional machine vision technology, large-size mirrors require high-resolution scanning, resulting in low image acquisition efficiency; traditional classification algorithms (such as Support Vector Machine (SVM) and Canny operator combined with Relevance Vector Machine (RVM)) rely on manual feature extraction, which is difficult to adapt to the irregularity and size differences of defects, easily leading to missed detections and misjudgments; at the same time, the sparse distribution of defects on large-size mirrors makes feature quantification difficult, further restricting detection accuracy; although deep learning has matured in industrial defect detection such as road surfaces and railway tracks, research on optical mirrors has mostly focused on small-aperture components, lacking dedicated public datasets and models.

[0004] While early deep learning models (such as ICFNet) achieved 96.7% accuracy on the small-sample ICF-90 dataset, their symmetric dual-block structure suffers from issues such as overlapping functions between shallow and deep feature extractors, limited regularization effect of a single BN layer, and gradient vanishing and slow convergence caused by the TanH activation function. These limitations make it difficult to meet the generalization and efficiency requirements of industrial-grade detection. In summary, the contradiction between high accuracy and high efficiency in detecting large-size optical mirror defects in small-sample scenarios is prominent, necessitating a dedicated detection method that integrates data augmentation and deep network optimization to overcome the current technological bottlenecks. Summary of the Invention

[0005] To address the issues of weak generalization ability, gradient decay in deep networks, and overfitting in optical mirror defect detection under small sample scenarios in existing technologies, this invention provides a deep learning-based method for optical mirror defect detection. Through spatial-channel collaborative data augmentation, dynamic residual network architecture construction, and spatial Dropout regularization design, it achieves high-precision and high-efficiency detection of optical mirror defects, comprising the following steps: Step S1: Obtain the original dataset of optical mirror defects and process it using a spatial-channel (SC) collaborative data augmentation strategy. The spatial dimension enhances sample diversity through rotation, mirror flipping, and brightness adjustment. The channel dimension calculates the gradients along the x and y axes using the Sobel operator and concatenates them with the original grayscale image to construct a three-channel gradient feature space. This expands the training samples to nine times their original size, mitigating statistical bias from small samples and enhancing the feature representation of weak edge defects such as scratches.

[0006] Step S2: Construct a 34-layer cascaded deep detection network (ICFNetV2), employing a residual connection mechanism and channel-decoupled convolutional co-design. Introduce 1×1 convolution operations in the jump paths of the residual structure to form a dual-stream collaborative mechanism between the main path and auxiliary paths, effectively suppressing gradient decay in deep networks and improving the efficiency of feature cross-layer transmission. Embed a spatial Dropout layer with a dropout probability of p=0.5 to force the network to establish redundant feature expression paths, achieving a regularization effect equivalent to multi-model integration and avoiding model overfitting.

[0007] Step S3: Input the enhanced dataset from Step S1 into the deep detection network constructed in Step S2, set the training rounds to 300, train the model based on the specified hardware and software environment, and complete the classification and detection of three types of defects: optical mirror pits, scratches, and dust.

[0008] The beneficial effects of the deep learning-based optical mirror defect detection method of the present invention are as follows: This method achieves accurate detection of three types of defects on large-size optical mirrors in small-sample scenarios, especially for weak-edge defects such as scratches, effectively overcoming the bottleneck of insufficient recognition accuracy in traditional methods. Experimental verification shows that the classification accuracy of this method reaches 97.4%, which is 5.2% higher than the traditional SVM method and 0.7% higher than the original ICFNet model. Moreover, the total training and testing time is only 344 seconds, balancing the requirements of detection accuracy and industrial-grade efficiency. The spatial-channel (SC) enhancement, dynamic residual structure, and spatial dropout module adopted have good reusability and can be independently adapted to other small-sample industrial defect detection scenarios such as steel strips and wind turbine blades. It is highly practical and provides a reliable technical reference for similar detection tasks. Attached Figure Description

[0009] Figure 1 This is an overall flowchart of a deep learning-based optical mirror defect detection method according to the present invention.

[0010] Figure 2 This is an ablation experiment diagram of an optical mirror defect detection method based on deep learning according to the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0012] like Figure 1 As shown, a method for detecting defects in optical mirrors based on deep learning is implemented with the following steps: Step S1: Select an example of optical mirror defect data: the ICF-90 dataset. This dataset contains three types of defects: pitting, scratches, and dust, with 30 samples for each type, totaling 90 images. The training set to test set ratio is 8:2. A spatial-channel (SC) collaborative data augmentation strategy is adopted. The spatial dimension is processed by rotation and mirroring, and the channel dimension is calculated by the Sobel operator and then concatenated with the original grayscale image, expanding the sample to 810 images, forming the ICF-810 dataset.

[0013] Step S2: Construct a 34-layer deep detection network ICFNetV2, which adopts a collaborative design of residual connections and channel decoupled convolutions, embeds a spatial dropout layer (p=0.5), contains two-stage dynamic residual blocks, and fuses dual-path features through a channel attention mechanism.

[0014] Step S3: Input the processed dataset obtained in Step S1 into the ICFNetV2 constructed in Step S2 and train it. The final classification accuracy of the three types of defects reached 97.4%, and the total training and testing time was 344 seconds.

[0015] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0016] As shown in Figure 2, to verify the effectiveness of the core module, an ablation experiment was conducted, and the steps are as follows: Step S1: Based on the environment and ICF-810 dataset of Implementation Method 1, construct 5 sets of models: full ICFNetV2, full ICFNet, Dropout layer removed, residual structure removed, and SC enhancement removed.

[0017] Step S2: Keep the training parameters consistent, train synchronously with the input dataset, and record the changes in accuracy.

[0018] Step S3: The results are shown in Table 1. The ICFNetV2 model performed best, with a significant improvement in accuracy, confirming the effectiveness of the collaborative performance enhancement of each module.

[0019] To verify the effectiveness of the target detection method based on color name fusion provided by the present invention, the present invention also provides the following experiments.

[0020] (1) Experimental data and experimental environment The dataset used in this invention is the ICF-90 dataset and the enhanced ICF-810 dataset, with a training set to test set ratio of 8:2. Hardware environment: Intel(R) Core(TM) i5-12600KF, 32GB RAM, NVIDIA GeForce RTX4060Ti; Software environment: Python 3.8, PyTorch 2.0.

[0021] (2) Experimental setup and evaluation indicators The ICFNetV2 network architecture was chosen to train the processed data. The batch size was set to 16, the number of training epochs was set to 300, and the learning rate was 0.0035.

[0022] Using classification accuracy (A) as the evaluation metric, the calculation formula is as follows:

[0023] (3) Experimental results and analysis Table 1 shows a performance comparison of the method of this invention with other models:

[0024] Experimental results show that the accuracy of the ICFNetV2 method is 0.7% higher than that of ICFNet and 5.2% higher than that of the traditional SVM, balancing accuracy and efficiency, and is suitable for small-sample optical mirror defect detection.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based optical mirror defect detection method, characterized in that: S1: obtaining an optical mirror defect original data set, and processing the original data set by using a space-channel (SC) cooperative data enhancement strategy to expand the training sample size; S2: constructing a deep detection network architecture, the architecture being a 34-layer cascade module, cooperatively designed and embedded with a residual connection mechanism and a channel decoupling convolution operation, and embedding a spatial Dropout layer; and S3: inputting the data set enhanced in step S1 into the deep detection network constructed in step S2 for training to complete the classification and detection of optical mirror defects. The space-channel (SC) cooperative data enhancement strategy in step S1 includes spatial dimension enhancement and channel dimension enhancement, wherein the spatial dimension enhancement processes the image by rotation, mirror flip, and brightness adjustment, the channel dimension enhancement calculates the x-axis and y-axis direction gradients by using a Sobel operator, and splices the gradient features and the original gray image to construct a three-channel gradient feature space, the training sample size is expanded to 9 times of the original data set after enhancement, the original data set is an ICF-90 data set containing three types of defects, i.e., pits, scratches, and dust, the enhanced data set is denoted as an ICF-810 data set, and the sample ratio of the training set and the test set of the enhanced data set is 8:

2. The 34-layer deep detection network architecture in step S2 includes a two-stage dynamic residual block (ResBlock), wherein the first stage (layers 1-3) is composed of a series of 3×3 convolution layers (Conv), ReLU activation functions, and 2×2 maximum pooling layers (MaxPool), the feature map is reduced by cross-step convolution, and the maximum pooling layer strengthens the brightness response of the pit area, each ResBlock in the second stage (layers 4-10) is internally arranged in parallel with a main path and an auxiliary path, the main path is of a "Conv→ReLU→Conv" structure, the auxiliary path is of a Conv structure, and the features of the two paths are dynamically fused by a channel attention mechanism. The residual connection mechanism in step S2 is an improved residual structure, a convolution operation is introduced in the jump path to form a fusion mode of "main path output + jump path output", both paths are subjected to nonlinear transformation, the jump path adopts 1×1 convolution to realize lightweight parameter expansion to ensure the feature transformation capability while avoiding the overfitting risk of a small data set. 2.The deep learning-based optical mirror defect detection method of claim 1, wherein, The dropout probability p of the spatial Dropout layer in step S2 is 0.5, random shielding of part of the neurons forces the network to establish redundant feature expression paths to realize the regularization effect equivalent to multi-model integration, and the spatial Dropout layer and the data enhancement of step S1 form double regularization, wherein the noise disturbance input space Dropout disturbs the feature space, and both cooperatively force the model to focus on the essential features of the defects. 3.The deep learning-based optical mirror defect detection method of claim 1, wherein, ​ 4.The deep learning-based optical mirror defect detection method of claim 1, wherein, ​ 5.The deep learning-based optical mirror defect detection method of claim 1, wherein, ​

Images