Method and system for detecting optical catastrophe damage of strip-shaped laser chip

Abstract

The invention relates to a method and a system for detecting optical catastrophe damage of a strip-shaped laser chip, the system comprises an acquisition unit, a detection network unit, a detection processing unit, a failure analysis unit and a display unit, firstly, the detection system completes acquisition of an image, and then the image is preprocessed to obtain the optical catastrophe damage of the strip-shaped laser chip. Inputting the preprocessed image into a detection algorithm of a detection network structure, finally outputting a detection result, and analyzing and processing the detected sample. According to the technical scheme, a complete strip-shaped laser chip optical catastrophe damage detection platform is formed by connecting hardware such as a probe station, an optical microscope, an industrial camera and a computer.

Description

Technical Field

[0001] This invention relates to the field of defect detection and identification of bar laser chips, and in particular to a detection method and system based on an asymmetric expansion roll-up algorithm to extract and identify optical catastrophic damage defects in bar laser chips. Background Technology

[0002] Bar lasers, with their small size, high photoelectric conversion efficiency, and excellent beam quality, are widely used in communications, medical, and sensing fields. Therefore, the reliability and stability of bar laser chips have become a research hotspot in recent years. When the output power of a laser increases, defects in the cavity surface film itself or its surface, such as cavity surface contamination and voids in the solder layer, can absorb light energy and cause localized heat accumulation. When the cavity surface reaches the damage threshold, it will melt at high temperature, causing dark line damage and leading to laser chip failure. This damage is called optical catastrophic damage (COD). In the actual failure modes of bar lasers, optical catastrophic damage (COD) is the most significant. Once a bar laser chip experiences optical catastrophic damage, it will lead to a significant decrease in the laser's output power, a drastically shortened lifetime, or even complete failure.

[0003] Therefore, it is essential to detect and analyze optical catastrophic damage in bar laser chips. Currently, the detection of optical catastrophic damage in bar laser chips still mainly relies on manual detection methods. Generally speaking, the bar laser is first powered on (the current is usually lower than its threshold current), and then observed through a microscope and infrared camera to determine whether there are dark line defects inside its active region.

[0004] However, manual inspection methods are limited by the operator's skill level, making it difficult to guarantee their reliability. Furthermore, due to the complex environmental background and extreme class imbalance in the acquired bar laser infrared images, traditional inspection methods struggle to meet the demands of industrial production.

[0005] Chinese patent document CN111934186A describes a method for determining the optical catastrophic type of a semiconductor laser chip, addressing the problems of high cost, long assessment time, and low production efficiency in existing COBD / COMD assessment methods for semiconductor laser chips. The method includes the following steps: setting a loading current, applying the loading current to the semiconductor laser chip, and acquiring the current-power / current-voltage curves of the semiconductor laser chip. If the current-power curve suddenly drops while the current-voltage curve suddenly rises, the chip's optical catastrophic type is determined to be COD; conversely, if both curves suddenly drop, the chip's optical catastrophic type is determined to be COBD. COMD damage requires a high-magnification metallurgical microscope for observation, while COBD requires sample preparation and analysis using electroluminescence technology (EL system) and infrared CCD, among other key equipment. Sample preparation requires highly experienced technicians.

[0006] Chinese patent document CN112051036A describes a method and apparatus for testing the peak power of the cavity optical catastrophic damage (COMD) of a semiconductor laser. This method avoids COMD burnout by measuring the peak power of the laser's COMD online.

[0007] Currently, there are many machine vision methods for detecting defects in laser chips. However, for industrial applications, few methods utilize hypermodels to locate and classify defects. Complex algorithms lead to long localization and classification times, resulting in untimely localization and low efficiency.

[0008] To address this, Chinese patent document CN116452847A provides a simple algorithm for defect detection and classification, enabling timely observation and classification of chip defects and quality issues during chip manufacturing. This method is based on a YOLOv5s neural network (including inputs, Backbone, Neck, and Prediction), collects defect datasets and training datasets, extracts features, performs computational learning, and connects the output branches through convolutional layers to achieve the final laser chip defect location and identification using this supermodel.

[0009] Taiyuan University of Technology has disclosed a laser chip defect detection method based on feature fusion and attention mechanisms in Chinese patent document CN117764950A. This method addresses the technical problem that the defect area of ​​laser chips is small and the automation and intelligence level of existing defect recognition technologies are not high enough. The method can effectively enhance robustness to multi-scale features and improve the accuracy and speed of automatic chip defect detection. It employs a Faster R-CNN model, which includes a feature extraction network, a Region Proposal Network (RPN), a Region of Interest Pooling (RolPooling) layer, and a Classification and Regression layer. A ResNet-101 model is embedded within the Faster R-CNN model; that is, the feature extraction network uses the ResNet-101 model, and a feature pyramid structure FP-VA is connected after the ResNet-101 model.The VA (Feature Pyramid Network) comprises a Feature Pyramid Network (FPN), a Multi-Scale Feature Fusion Structure (FPUD), and an Ultra-Lightweight Subspace Attention Mechanism (ULSAM). The FP-VA structure connects the Region Proposal Networks (RPN) and the Region of Interest Pooling (ROLPooling) layer, respectively. The ROLPooling layer connects the Classification and Regression layers. An image of a microarray chip is input into a ResNet-101 model. The ResNet-101 model extracts important feature information from the image, resulting in n sets of feature maps. These n sets of feature maps are then input into a Feature Pyramid Network (FP-VA) structure. First, they pass through a Feature Pyramid Network (FPN) to obtain n-1 sets of feature maps. For each set of feature maps, a Multi-Scale Feature Fusion Network (FPUD) scales the feature maps from other sets to the same size. Then, the n-1 sets of feature maps are stacked and passed through a 1x1 convolutional layer to obtain a single output. The final output is n-1 sets of multi-scale fused feature maps. These n-1 sets of multi-scale fused feature maps are then processed by an Ultra-Lightweight Subspace Attention (ULSAM) mechanism to achieve multi-scale, multi-frequency feature representation. After processing by the FP-VA structure, the n-1 sets of feature maps are processed by Region Proposal Networks (RPN) and an Interest Region Pooling layer (RolPooling), respectively, before being output to the Classification and Regression layer. The Proposal Networks (RPN) generate candidate regions for each feature map. The Region Proposal Networks (RPN) pooling layer collects the coordinate information of the candidate regions generated by the RPN and extracts information from the n-1 sets of feature maps input from the Feature Pyramid Structure (FP-VA) to obtain n-1 sets of proposal feature maps. The Classification and Regression layer uses the n-1 sets of proposal feature maps to calculate the specific category, obtaining the final precise location of the defect in the candidate region. Boundary regression is then used to correct the precise location of the object, thus providing a basis for subsequent chip defect analysis.

[0010] However, in the existing technology, there is no technical solution that can determine the type of optical disaster without using power-voltage curves, and can be based on machine vision for deep learning architecture and form a complete strip laser chip optical disaster damage detection platform with hardware such as microscopes and cameras. Summary of the Invention

[0011] In view of the above-mentioned problems in the prior art, the present invention aims to provide a technical solution based on an improved encoder-decoder architecture, which forms a complete optical catastrophic damage detection platform for bar laser chips by connecting with hardware such as probe stations, optical microscopes, industrial cameras, and computers.

[0012] Specifically, according to one aspect of the present invention, a detection system for optical catastrophic damage of a bar laser chip is provided, characterized in that it includes an acquisition unit, a detection network unit, a detection processing unit, a failure analysis unit, and a display unit. The acquisition unit is used to acquire image data and includes a probe station for fixing the chip sample, a microscope for observing optical catastrophic damage of the chip sample, and a CCD for acquiring image information. The image data acquired by the acquisition unit is transmitted to a trained detection network for defect segmentation and detection, and the detection data is transmitted to the detection processing unit. The detection processing unit processes the segmented detection data and transmits the results to the failure analysis unit, analyzes and processes the acquired sample, and transmits the final analysis and processing results to the display unit for display.

[0013] According to the invention, image data is acquired using devices such as a probe station, microscope, and CCD, and then transmitted to a trained detection network via a data cable for defect segmentation and detection. Finally, the segmentation and detection results are displayed on a screen, thereby making the detection process visible.

[0014] According to the invention, it is mainly aimed at bar laser chips, and the main defect types are optical catastrophic cavity damage (COBD) and optical catastrophic mirror damage (COMD), especially COBD.

[0015] According to this invention, the network used is a neural network based on an encoder-decoder architecture. The encoder is responsible for mapping input data (such as an image) into a fixed-length intermediate representation vector, extracting features from the input data, and compressing it into a low-dimensional vector space. The decoder then uses the intermediate vector generated by the encoder to progressively generate output data. When generating output, the decoder predicts the next output element based on the intermediate vector and the already generated partial output. When the input data is an image, the encoder encodes the image, extracts its features, and the decoder converts these features into text describing the image content. Common encoder-decoder architectures include: Recurrent Neural Network (RNN) based: such as encoder-decoder based on Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs).

[0016] Convolutional Neural Network (CNN) based: In image-related tasks, CNNs are often used as encoders to extract spatial features of images, while decoders are usually composed of deconvolutional layers, etc., to convert the encoded feature maps into images.

[0017] Transformer: Based entirely on the attention mechanism, it can perform parallel computation and efficiently process long sequence data.

[0018] In this invention, a convolutional neural network (CNN) or a GPT-based Transformer variant network is preferred.

[0019] In this invention, the method used is still EL (electroluminescence) technology, and the CCD is preferably an infrared CCD.

[0020] In this invention, more preferably, the encoder structure mainly consists of a multi-scale feature fusion module, which is primarily used for extracting defect features. Similarly, the decoder structure also uses a multi-scale feature fusion module for feature reconstruction.

[0021] Behind the decoder, a deep supervision module is used to directly upsample the output of each layer of the decoder to the original image size and participate in the training of the network model to calculate the loss.

[0022] In this invention, a further preferred embodiment of the multi-scale feature fusion module structure (e.g.) Figure 2 As shown in the figure, each layer consists of asymmetric dilated convolutional layers with different dilation coefficients. The difference between asymmetric dilated convolutional layers and conventional convolutional layers is that the former is 1. 1, 3 3 or 5 The asymmetric convolutional layer of this invention has a kernel size of 3, consisting of 5 components. 5, 5 3, 3 7 or 7 The system consists of three parts. Multi-scale feature maps are extracted through asymmetric dilation convolutional layers with different dilation coefficients, then feature maps are fused by concatenation, and then passed through attention modules before being added to the original input image for fusion.

[0023] According to another aspect of the present invention, a method for detecting optical catastrophic damage in a bar laser chip is provided, utilizing the aforementioned detection system for optical catastrophic damage in a bar laser chip. The method is characterized by first acquiring an image through the detection system, then performing image preprocessing, i.e., cropping and enhancing the image, and finally inputting the preprocessed image into a detection network structure (such as...). Figure 1 In the detection algorithm of the detection network structure shown, the detection results are finally output, and the tested sample is analyzed and processed.

[0024] In this invention, preferably, the detection network is a neural network based on an encoder-decoder architecture, including an input-output layer, a feature layer, a pooling layer, and an asymmetric dilated convolutional layer.

[0025] In this invention, preferably, the detection network includes a multi-scale feature fusion module, which takes a feature map as input and performs 3... 5, 5 3, 3 7 or 7 3. Asymmetric dilated convolutional layers extract multi-scale feature maps, which are then fused using a multi-scale feature concatenation method. 1. Convolutional computation, then passing through attention modules separately, and finally fusion with the original input feature map. In the encoder structure, 1 1. Convolution calculation, channel augmentation, in the decoder structure, 1 1. Convolution calculation, channel reduction.

[0026] According to the present invention, a neural network model can be used to achieve efficient defect localization and detection. The algorithm is simple and effective, and can be used to efficiently extract and identify optical catastrophic damage defects, especially COBD damage, of bar laser chips based on the asymmetric expansion roll machine algorithm. Attached Figure Description

[0027] Figure 1 This is a schematic diagram illustrating the structure of the defect detection network used in the method and system for detecting optical catastrophic damage to a strip laser chip according to a specific embodiment of the present invention.

[0028] Figure 2 To show Figure 1 A schematic diagram of the structure of the multi-scale feature fusion module used in the defect detection network.

[0029] Figure 3 The flowchart illustrates a method for detecting optical catastrophic damage to a bar laser chip according to a specific embodiment of the present invention.

[0030] Figure 4 This is a block diagram illustrating a detection system for optical catastrophic damage to a bar laser chip according to a specific embodiment of the present invention.

[0031] Figure 5 To show Figure 1 The diagram shows the detection results of the defect detection network. In the diagram, (a) and (c) are schematic diagrams of the images acquired from the data acquisition, and (c) and (d) are schematic diagrams of the results after detection. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. Those skilled in the art will understand that this description is exemplary and the present invention is not limited to these specific embodiments.

[0033] Figure 1 The structure of the defect detection network used in detecting optical catastrophic damage to the bar laser chip of the present invention is shown.

[0034] The structure of the optical catastrophic damage and defect segmentation detection network is as follows: Figure 1 As shown, this network is mainly designed based on an encoder-decoder architecture. The encoder structure primarily consists of a multi-scale feature fusion module, which is mainly used for extracting defect features. Similarly, the decoder structure uses a multi-scale feature fusion module for feature reconstruction. Following the decoder, a deep supervision module (described in detail later) is employed, which directly upsamples the output of each layer of the decoder to the original image size and participates in the training of the network model for loss calculation. Specifically, in this embodiment, the detection network structure mainly consists of the following parts: encoder, skip connections, decoder, and deep supervision module. The encoder is mainly used for extracting defect features; the decoder is used for reconstructing image features; skip connections are used to connect the encoder and decoder at the same layer to compensate for information loss caused during image upsampling; the deep supervision module upsamples the output of each layer of the decoder (…). Figure 1 After fusing S1 to S5 in the network, loss calculation is performed to improve the overall detection accuracy of the network. Figure 5 Show Figure 1 A schematic diagram of the defect detection network is shown. (a) and (c) are schematic diagrams of the images acquired from the data acquisition, and (c) and (d) are schematic diagrams of the results after detection. In this specific embodiment, the image detected using the structure of the optical catastrophic damage defect segmentation detection network of this invention must meet the following condition: the current applied to the bar laser chip should be less than its threshold current. After applying a small current to the sample laser chip using a DC power supply, it is magnified using a microscope, and then its emission image data is acquired using a camera mounted above the microscope. After completing the image acquisition, image annotation software is used for image annotation. In this embodiment, different pixel information in the image is divided into two categories: optical catastrophic damage area and background area (i.e., the black background in the figure). Through detection, when a defective image is input ( Figure 1 After the input layer is fed into the detection network, the output image ( Figure 1 The output layer marks the location, shape, and size of defects as colored areas different from the black background (as shown in Figures b and d), while normal laser samples do not show colored areas.

[0035] Figure 2The structure of the multi-scale feature fusion module used in the defect detection network is shown.

[0036] In this specific implementation, both the encoder and decoder employ multi-scale feature fusion modules in their specific structures. Since the detected sample is a bar-shaped laser, its active region structure is also bar-shaped, and the optical catastrophe loss occurring in the active region is also bar-shaped. Using asymmetric convolutional kernels can more effectively enhance the bar-shaped target features, thereby increasing its receptive field.

[0037] The structure of the multi-scale feature fusion module is as follows: Figure 2 As shown, each layer consists of asymmetric dilated convolutional layers with different dilation coefficients. The asymmetric dilated convolutional layers differ from conventional convolutional layers in that they have a 1:1 dilation coefficient. 1, 3 3 or 5 The asymmetric convolutional layer of this invention has a kernel size of 3, consisting of 5 components. 5, 5 3, 3 7 or 7 The process consists of three parts. Multi-scale feature maps are extracted using asymmetric dilation convolutional layers with different dilation coefficients. These maps are then fused through concatenation, passed through attention modules, and finally added to the original input image for fusion. The specific calculation process can be represented by the following formula:

[0038] in, Represents the input feature map, This represents a splicing operation. This represents a convolution kernel of 1. 1. A convolutional layer with an expansion coefficient of 1. This is the feature map after being fused through five dilated convolutions according to this formula. Furthermore:

[0039] in, This represents the output of the multi-scale feature fusion module. Represents the input feature map, This represents the attention module.

[0040] The following provides a detailed explanation of the deep monitoring module.

[0041] To fully utilize the multi-scale features generated by deep and shallow layers in the network, and to minimize the loss of detailed information without increasing the load on feature modules, this invention proposes a deep supervision module, as shown in the gray dashed box on the right side of Figure 1. Its specific structure is as follows: it uses a convolutional kernel with a size of 3... The convolution with size 3 connects the outputs (S1, S2, S3, S4, and S5) of each layer in the decoder part of the network. Then, these five output feature maps are scaled up to the size of the input image using bilinear interpolation and named Sup1, Sup2, Sup3, Sup4, and Sup5 respectively. These feature maps are then concatenated and fed into a convolution kernel with a size of 1. A convolutional layer of size 1 is used to generate the feature segmentation result map Sup0. Finally, its dimensions are adjusted through convolution to obtain the final output. Through these operations, the final data feature map will simultaneously contain shallow spatial information and deep semantic information. The specific calculation is shown in the following formula:

[0042] in, These represent the feature layers output by S1, S2, S3, S4, and S5, respectively. This indicates that bilinear interpolation is used to calculate the first... Each size is Feature maps upsampled to the input image size Same size This indicates that the convolution kernel is 3. A convolutional layer of 3. This represents the fused output feature map. This indicates a splicing operation.

[0043] The model training process is described in detail below.

[0044] During training, the selected loss function is a fusion of the binary cross-entropy loss function (BCE) and the Dice loss function. The BCE loss function is used to calculate the class loss between different pixels, while the Dice loss function measures the similarity between the predicted value and the true label. The specific calculation process is as follows.

[0045]

[0046] in, Represents the binary cross-entropy loss function. This represents the Dice loss function. These are the weights for different loss functions.

[0047] During model training, to improve the model's detection accuracy, the outputs Sup0 to Sup5 of each layer are treated as pixel-level classification devices, and a predicted feature map is generated for each, which together participate in the loss calculation process. The calculation process is as follows:

[0048] in, This represents the total loss during the training process. The output layer Sup(m) represents the output layer, where m takes values ​​from [0, 5]. This invention primarily targets bar laser chips, where the main optical catastrophic damage defect types are: Optical Catastrophic Cavity Damage (COBD) and Optical Catastrophic Mirror Damage (COMD).

[0049] Figure 3 This is a flowchart of a method for detecting optical catastrophic damage to a bar laser chip.

[0050] First, image data is acquired using equipment such as a probe station, microscope, and CCD. Then, the data is transmitted to the trained detection network via a data cable for defect segmentation and detection. Finally, the segmentation and detection results are displayed on the screen.

[0051] The detection process is as follows: First, the image is acquired using the aforementioned equipment. Then, the image undergoes preprocessing, including cropping and enhancement. Finally, the preprocessed image is input into... Figure 1 The detection algorithm shown outputs the detection results and analyzes the tested sample.

[0052] Figure 4 This is a block diagram illustrating a system for detecting optical catastrophic damage to a bar laser chip. Figure 4 The detection system shown can be specifically used for Figure 3 This detection method can also be used for other detection methods. In COMD defect detection, it is preferred to use... Figure 3 The detection method is described. Specifically, in this embodiment, the detection system for optical catastrophic damage of the strip laser chip includes, from left to right, an acquisition unit, a detection network unit, a detection processing unit, a failure analysis unit, and a display unit. The acquisition unit is used to acquire image data and includes a probe station for fixing the chip sample, a microscope for observing optical catastrophic damage to the chip sample, and a CCD for acquiring image information. The image data acquired by the acquisition unit is transmitted to the trained detection network for defect segmentation and detection, and the detection data is transmitted to the detection processing unit. The detection processing unit processes the segmented detection data and transmits the results to the failure analysis unit for analysis and processing of the acquired sample, and transmits the final analysis and processing results to the display unit for display.

[0053] In this invention, image data is acquired using devices such as a probe station, microscope, and CCD. This data is then transmitted via data cable to a trained detection network unit for defect segmentation and detection. Finally, the segmentation and detection results are displayed on the screen of a display unit, thus making the detection process visible. It primarily targets bar laser chips, with the main defect types being optically catastrophic cavity damage (COBD) and optically catastrophic mirror damage (COMD), especially COBD.

[0054] The detection network unit uses a neural network based on an encoder-decoder architecture. When the input data is an image, the encoder encodes the image and extracts its features, and the decoder converts these features into text describing the image content.

[0055] In this invention, the CCD is preferably an infrared CCD. The microscope is preferably a metallurgical microscope.

[0056] In this invention, more preferably, the encoder structure mainly consists of a multi-scale feature fusion module, which is primarily used for extracting defect features. Similarly, the decoder structure also uses a multi-scale feature fusion module for feature reconstruction.

[0057] Behind the decoder, a deep supervision module is used to directly upsample the output of each layer of the decoder to the original image size and participate in the training of the network model to calculate the loss.

[0058] In this invention, more preferably, the structure of the multi-scale feature fusion module is composed of asymmetric dilated convolutional layers with different dilation coefficients, and the kernel size of the asymmetric dilated convolutional layers is 3. 5, 5 3, 3 7 or 7 The system consists of three parts. Multi-scale feature maps are extracted through asymmetric dilation convolutional layers with different dilation coefficients, then feature maps are fused by concatenation, and then passed through attention modules before being added to the original input image for fusion.

[0059] The present invention has been described in detail above with reference to specific embodiments. However, those skilled in the art will understand that the description is exemplary and various modifications and changes can be made. As long as they do not depart from the spirit and purpose of the present invention, all such modifications and changes should fall within the protection scope of the present invention, which is defined by the appended claims.

Claims

1. A detection system for optical catastrophic damage to a strip laser chip, characterized in that, The system includes an acquisition unit, a detection network unit, a detection processing unit, a failure analysis unit, and a display unit. The acquisition unit is used to acquire image data and includes a probe station for fixing the chip sample, a microscope for observing optical catastrophic damage to the chip sample, and a CCD for acquiring image information. The image data acquired by the acquisition unit is transmitted to the trained detection network for defect segmentation and detection, and then transmitted to the detection processing unit. The detection processing unit processes the segmented detection data and transmits the results to the failure analysis unit for analysis and processing of the acquired sample. Finally, the analysis and processing results are transmitted to the display unit for display.

2. The detection system for optical catastrophic damage to a strip laser chip as described in claim 1, characterized in that, Image data is acquired using a probe station, microscope, and CCD, and then transmitted via a data cable to a trained detection network unit for defect segmentation and detection. Finally, the segmentation and detection results are displayed on a screen, thus making the detection process visible.

3. The detection system for optical catastrophic damage to a strip laser chip as described in claim 1, characterized in that, The defect types of optical catastrophic damage are: optical catastrophic cavity damage (COBD) and optical catastrophic mirror damage (COMD).

4. The detection system for optical catastrophic damage to a strip laser chip as described in claim 1, characterized in that, The network used is a neural network based on an encoder-decoder architecture. In the encoder structure, it mainly consists of a multi-scale feature fusion module for extracting defect features. In the decoder structure, a multi-scale feature fusion module is also used for feature reconstruction.

5. The detection system for optical catastrophic damage to a strip laser chip as described in claim 4, characterized in that, Behind the decoder, a deep supervision module is used to directly upsample the output of each layer of the decoder to the original image size and participate in the training of the network model to calculate the loss.

6. The detection system for optical catastrophic damage to a strip laser chip as described in claim 4 or 5, characterized in that, The multi-scale feature fusion module is composed of asymmetric dilated convolutional layers with different dilation coefficients, and the kernel size of the asymmetric dilated convolutional layers is 3. 5, 5 3, 3 7 or 7 It consists of 3 components.

7. The detection system for optical catastrophic damage to a bar laser chip as described in claim 6, characterized in that, Multi-scale feature maps are extracted by asymmetric dilation convolutional layers with different dilation coefficients, then feature fusion is performed by concatenation, and then each feature is passed through an attention module before being added to the original input image for fusion.

8. A method for detecting optical catastrophic damage to a strip laser chip, utilizing the detection system described in any one of claims 1-7, characterized in that, First, the detection system acquires the image. Then, the image undergoes preprocessing. Next, the preprocessed image is input into the detection algorithm of the detection network structure. Finally, the detection result is output, and the sample is analyzed and processed.

9. The method for detecting optical catastrophic damage to a strip laser chip as described in claim 8, characterized in that, The detection network is a neural network based on an encoder-decoder architecture, including input-output layers, feature layers, pooling layers, and asymmetric dilated convolutional layers.

10. The method for detecting optical catastrophic damage to a strip laser chip as described in claim 8, characterized in that, The detection network includes a multi-scale feature fusion module, which takes a feature map as input and performs fusion on a 3D model. 5, 5 3, 3 7 or 7 3. Asymmetric dilated convolutional layers extract multi-scale feature maps, which are then fused using a multi-scale feature concatenation method.

1. Convolutional computation, then passing through attention modules separately, and finally fusion with the original input feature map. In the encoder structure, 1 1. Convolution calculation, channel augmentation, in the decoder structure, 1 1. Convolution calculation, channel reduction.

Images