A VCSEL quality screening method based on an image fusion algorithm

Abstract

The present application relates to a kind of based on improved RFN-Nest image fusion algorithm, and the algorithm is applied in VCSEL's EL defect image enhancement, and finally applied to the defect detection field of VCSEL.The method is based on RFN-Nest model Grad Cross-Nest fusion network is proposed;The Grad Cross-Nest fusion network includes encoder, fusion network and decoder.The method is according to VCSEL defect development mechanism, creatively introduces image fusion algorithm into VCSEL's quality screening scheme when.The algorithm is free from the dependence of traditional manual design fusion strategy, can self-adaptively learn the feature information of input image, realizes end-to-end fusion process.Compared with traditional screening means, the method not only low in cost, and have higher efficiency, can promote VCSEL manufacturing and defect cause analysis research.

Description

Technical Field

[0001] This invention relates to an improved RFN-Nest image fusion algorithm and applies it to the enhancement of VCSEL defect images. This method can be used in the VCSEL quality screening process. Background Technology

[0002] Vertical-cavity surface-emitting lasers (VCSELs) have become quite popular in data communications and are increasingly favored in fields such as optical imaging, detection, consumer electronics, robotics, and autonomous driving. Compared to edge-emitting lasers (DFB), VCSELs offer longer lifetimes, lower manufacturing costs, and better beam profiles. To improve their reliability, research into different failure mechanisms is essential.

[0003] The complex manufacturing process of VCSELs leads to various types of defects. Research by Agilent Technologies' team indicates these defects include epitaxial defects, scratches, surface contamination, electrostatic discharge (ESD), and electrical overstress. Numerous technologies are applied to the VCSEL quality screening process to address these defects. Currently, the mainstream methods for VCSEL defect detection include reverse IV testing, electroluminescence imaging (EL), focused ion beam / transmission electron microscopy (FIB / TEM), thermal voltage variation (TIVA), and scanning electron microscopy (SEM). TIVA cannot pinpoint the defect source and requires polishing the back of the VCSEL, making it a destructive testing technique. While FIB / TEM and SEM offer high precision and deep imaging depth, they are also destructive testing methods, primarily used to detect internal microscopic defects caused by ESD and EOS. Furthermore, their high labor and cost make them unsuitable for routine use. IV testing requires thousands of hours of lifetime testing to observe changes in leakage current and reverse bias breakdown voltage to determine damage, resulting in significant time costs. The most useful conventional technique is electroluminescent imaging (EL). EL technology is a time-saving, low-cost, and routine inspection method that does not require surface polishing or sample preparation, making it the most commonly used inspection method by manufacturers in the VCSEL quality screening stage. However, in order to obtain a clear infrared image of dark defects in the emission aperture, the aperture and infrared light source intensity need to be adjusted appropriately. This causes the VCSEL's EL infrared image to lose most of the surface defect features. This not only leads to engineers repeating the inspection work on damaged chips, but also makes it difficult for engineers to infer the extent to which VCSEL surface defects affect the dark defects in the emission aperture. Summary of the Invention

[0004] This invention proposes an RFN-Nest-based image fusion algorithm to fuse images of VCSEL surface defects and dark defect images at the aperture, aiming to help engineers efficiently complete the quality screening process of VCSELs.

[0005] This invention is achieved using the following technical solution: a VCSEL quality screening method based on an image fusion algorithm, which proposes a Grad Cross-Nest fusion network based on the RFN-Nest model; the Grad Cross-Nest fusion network includes an encoder, a fusion network, and a decoder; Pairs of VCSEL infrared and visible light images are input into the encoder, which decomposes each pair of input images into multiple pairs of images with different resolutions and inputs them into the fusion network. The fusion network employs a gradient cross-fusion strategy. This network includes two tightly connected convolutional neural network modules and a gradient cross-path module. The two tightly connected convolutional neural networks extract pixel intensity information from visible light and infrared light images, respectively. The gradient cross-path module simultaneously extracts gradient cross-information from both visible light and infrared light images. The resulting gradient cross-information contains gradient information from both visible light and infrared light images. This information is then added to the feature maps of the visible light and infrared light images extracted from pixel intensity information, respectively, to initially form a multi-channel pre-fused feature map. The decoder upsamples the multi-channel pre-fused feature maps obtained by the fusion network to reconstruct the image, resulting in a fused image for quality screening.

[0006] The screening method described in this invention is the first to introduce an image fusion algorithm into the VCSEL quality screening scheme, making a significant innovation. The gradient cross-path module enables the exchange of gradient information between two paths, preserving texture information from visible and infrared images. The pre-fused gradient feature map further enriches the detailed information of subsequent feature maps, assisting the network to converge faster and achieving better fusion performance. Applied to the VCSEL quality screening process, it avoids engineers performing repetitive inspections of damaged chips, improving inspection efficiency.

[0007] Furthermore, after the Grad Cross-Nest fusion network is built, it needs to be trained. The training process uses a VCSEL visible and infrared image dataset with a resolution of 256×256 as the training set. This network is implemented using PyTorch 1.8 and trained on a PC with an Intel CPU (i9-10900X) and an NVIDIA GPU (RTX 3090). The Adam optimizer is used to minimize the training loss. The training epochs and the number of samples taken per training cycle are both set to 4, and the learning rate is constant at 1e-4.

[0008] The beneficial effects of this invention are: (1) Based on the defect development mechanism of VCSEL, an image fusion algorithm is introduced into the quality screening scheme of VCSEL. Compared with traditional screening methods, this method is not only low in cost but also more efficient, which can promote the research on VCSEL manufacturing and defect cause analysis; (2) Based on RFN-Nest, a novel image fusion algorithm, Gradcross-Nest, is proposed. This algorithm gets rid of the dependence on the traditional manual design of fusion strategies, can adaptively learn the feature information of the input image, and realizes the end-to-end fusion process. Attached Figure Description

[0009] Figure 1 Grad Cross-Nest network structure diagram.

[0010] Figure 2 Grad Cross-Fusion network architecture diagram.

[0011] Figure 3 Grad Cross Path module structure diagram. Detailed Implementation

[0012] Example 1: A VCSEL quality screening method based on image fusion algorithm. This method proposes a Grad Cross-Nest fusion network based on the RFN-Nest model. The Grad Cross-Nest fusion network includes an encoder, a fusion layer, and a decoder. Pairs of VCSEL infrared and visible light images are input into the encoder, which decomposes each pair of input images into multiple pairs of images with different resolutions and inputs them into the fusion network. The fusion network employs a Grad Cross Fusion network, which includes two tightly connected convolutional neural network (Dense) blocks and a Grad Cross Path module. The two Dense blocks extract pixel intensity information from the visible light image and the infrared image, respectively. The Grad Cross Path module can simultaneously extract gradient cross information from the visible light image and the infrared image. The resulting gradient cross information contains gradient information from both the visible light image and the infrared image. This information is then added to the visible light image feature map and the infrared image feature map extracted from the pixel intensity information, respectively, to initially form a multi-channel pre-fused feature map. The decoder upsamples the multi-channel pre-fused feature maps obtained by the fusion network to reconstruct the image, resulting in a fused image for quality screening.

[0013] After the Grad Cross-Nest fusion network described in Example 2 is built, it needs to be trained. The training process uses a dataset of 256×256 resolution VCSEL visible and infrared images as the training set. The network is implemented using PyTorch 1.8, and the fusion model is trained on a PC with an Intel CPU (i9-10900X) and an NVIDIA GPU (RTX3090). The Adam optimizer is used to minimize the training loss. The training cycle and the batch size are both set to 4, and the learning rate is constant at 1e-4. To complete the training of the Grad Cross-Nest fusion network, 10100 pairs of 256×256 resolution VCSEL visible and infrared defect images were created as the training set.

[0014] Example 3: The Grad Cross Path module consists of a Sobel gradient extraction operator and two convolutional layers. The Sobel gradient extraction operator is responsible for extracting gradients from the feature map, and the two convolutional layers are responsible for maintaining a consistent number of channels. The output feature map is added to two densely connected convolutional neural network (Dense Block) blocks, and then the image is reconstructed to obtain the fused image. The encoder decomposes each pair of input images into multiple pairs of images with resolutions of 256×256, 128×128, 64×64, and 32×32 and inputs them into the fusion network.

[0015] Example 4: The decoder is implemented based on the Res and Dense architecture, which can fuse global contextual feature information with shallow feature information, compensate for the fading of high-level semantic information, and suppress background noise of shallow features. Res and Dense can repeatedly reuse features, improving the quality of the final fused image.

[0016] The technical solution of this invention is embodied in three parts; the invention will be further described below with reference to the accompanying drawings.

[0017] I. Construction of the Improved RFN-Nest Model The RFN-Nest network consists of three parts: an encoder, a fusion layer, and a decoder. To integrate global contextual features with shallow feature information, suppress background noise in shallow features, and more accurately locate salient targets, we adopt the Nest network structure, i.e., the encoder and decoder structure, which will not be elaborated upon here. The RFN module (Fusion) is based on the Res structure. While it can reuse features and improve the network's convergence speed and learnability, it does not consider the efficiency of gradient information extraction from the image.

[0018] To obtain high-quality VCSEL fused images, we propose a GradCross-Nest fusion network based on RFN-Nest, such as... Figure 1 As shown in the diagram. The Grad Cross_Fusion network is implemented using a Dense Block structure, and it is divided into two paths to extract pixel intensity information from visible light and infrared light images respectively, as shown in the diagram. Figure 2 As shown. The Grad Cross Path module consists of a Sobel gradient extraction operator and two convolutional layers, as follows. Figure 3 As shown, the Sobel gradient extraction operator is responsible for extracting gradients from the visible and infrared feature maps. Two convolutional layers maintain a consistent number of channels, and the output feature maps are added to the two Dense Block branches respectively. The resulting image is then reconstructed to obtain the fused image. This design not only enables gradient information exchange between the two paths, preserving texture information from both visible and infrared images, but the pre-fused gradient feature maps also enrich the detail information of subsequent feature maps, helping the network converge faster and achieving better fusion performance.

[0019] II. Model Training Process A dataset of 10,100 VCSEL visible and infrared images with a resolution of 256×256 was selected as the training set. The network was implemented using PyTorch 1.8 and trained on a PC with an Intel CPU (i9-10900X) and an NVIDIA GPU (RTX 3090). The Adam optimizer was used to minimize the training loss. The training epochs and batch sizes were both set to 4, and the learning rate was constant at 1e-4.

[0020] III. Classification of VCSEL Defects This method mainly targets surface defects and dark defects at the aperture of VCSEL chips. The main defect types are scratches, contamination, epitaxial defects, and dark defects at the aperture.

Claims

1. A VCSEL quality screening method based on image fusion algorithm, which proposes a Grad Cross-Nest fusion network based on the RFN-Nest model; characterized in that: The Grad Cross-Nest fusion network includes an encoder, a fusion network, and a decoder; Pairs of VCSEL infrared and visible light images are input into the encoder, which decomposes each pair of input images into multiple pairs of images with different resolutions and inputs them into the fusion network. The fusion network employs a gradient cross-fusion strategy. This network includes two tightly connected convolutional neural network modules and a gradient cross-path module. The two tightly connected convolutional neural networks extract pixel intensity information from visible light and infrared light images, respectively. The gradient cross-path module simultaneously extracts gradient cross-information from both visible light and infrared light images. The resulting gradient cross-information contains gradient information from both visible light and infrared light images. This information is then added to the feature maps of the visible light and infrared light images extracted from pixel intensity information, respectively, to initially form a multi-channel pre-fused feature map. The decoder upsamples the multi-channel pre-fused feature maps obtained by the fusion network to reconstruct the image, resulting in a fused image for quality screening.

2. The VCSEL quality screening method based on image fusion algorithm as described in claim 1, characterized in that, After the Grad Cross-Nest fusion network is built, it needs to be trained. The training process uses a VCSEL visible light and infrared light image dataset with a resolution of 256×256 as the training set. The network is implemented using PyTorch 1.8 and trained on a PC with an Intel CPU (i9-10900X) and an NVIDIA GPU (RTX 3090). The Adam optimizer is used to minimize the training loss.

3. The VCSEL quality screening method based on image fusion algorithm as described in claim 2, characterized in that, The training cycle and the number of samples taken in one training session are both set to 4, and the learning rate is constant and set to 1e-4.

4. The VCSEL quality screening method based on image fusion algorithm as described in claim 3, characterized in that, A training set of 10,100 pairs of VCSEL visible and infrared defect images with a resolution of 256×256 was created.

5. A VCSEL quality screening method based on an image fusion algorithm as described in any one of claims 1-4, characterized in that, The gradient cross path module consists of a Sobel gradient extraction operator and two convolutional layers. The Sobel gradient extraction operator is responsible for extracting the gradient from the feature map, and the two convolutional layers are responsible for keeping the number of channels consistent. The output feature map is added to two tightly connected convolutional neural networks, and then the image is reconstructed to obtain the fused image.

6. A VCSEL quality screening method based on an image fusion algorithm as described in any one of claims 1-4, characterized in that, The encoder decomposes a pair of input images into multiple pairs of images with resolutions of 256×256, 128×128, 64×64, and 32×32 and inputs them into the fusion network.

7. A VCSEL quality screening method based on an image fusion algorithm as described in any one of claims 1-4, characterized in that, This method targets surface defects and dark defects at the aperture of VCSEL chips. The main defect types are scratches, contamination, epitaxial defects, and dark defects at the aperture.

8. A VCSEL quality screening method based on an image fusion algorithm as described in any one of claims 1-4, characterized in that, The decoder is implemented based on the Res and Dense architectures.

Images