A VCSEL semiconductor laser array near-field light-emitting hole detection device and method

Abstract

The invention belongs to the technical field of semiconductor laser testing, and proposes a VCSEL semiconductor laser array near-field light-emitting hole detection device and method, which automatically collect and identify the VCSEL semiconductor laser cavity surface light spot, and realize automatic and accurate triggering mode through a wafer tray controlled by a probe pin and a stepping instruction, which saves a lot of time compared with traditional manual adjustment, has a high degree of automation, and reduces the error caused by manual operation; based on the Mask RCNN network, multiple near-field image features are extracted, and each light-emitting hole in each image is instance-segmented, and different models of VCSEL semiconductor laser data sets are used for training, so as to improve the accuracy of quality judgment of each light-emitting hole in the near field of the VCSEL semiconductor laser.

Description

Technical Field

[0001] The present invention belongs to the technical field of semiconductor laser testing, and in particular relates to a VCSEL semiconductor laser array near-field light-emitting hole detection device and method. Background Art

[0002] With the continuous development of scientific research fields such as laser processing, laser medical treatment, and laser display, VCSEL (vertical cavity surface emitting laser) chips have been used more and more. Compared with traditional edge-emitting lasers, VCSEL chips have more obvious advantages. They have the characteristics of small size, circular output spot, small beam divergence angle, single longitudinal mode output, small threshold current, easy measurement and screening, can use cheap driving power supply, and easy integration into large-area arrays. Therefore, VCSEL chips are widely used in optical communications, optical interconnection, and optical storage. At the same time, as one of the cheap optoelectronic semiconductors with a broad application market and use value, VCSEL chips have always been a hot research object in the optoelectronic semiconductor industry.

[0003] The quality of the near-field light-emitting holes is an important parameter for judging the performance of semiconductor lasers. The main purpose of the near-field test of semiconductor laser chips includes testing whether the number of light-emitting holes in the chip meets the requirements, whether there are bad spots, and whether the brightness value and uniformity are good. Light-emitting holes that are too dark cannot guarantee the needs of use, while light-emitting holes that are too bright are prone to overexposure, resulting in a greatly reduced service life. Therefore, in the research and development stage, near-field light-emitting hole quality detection is crucial. Patent CN105321179A, patent CN101630414B and patent CN107067400A have disclosed that the current method for measuring near-field light-emitting hole images mainly adopts the following methods or steps: near-field light source camera scanning, connected domain analysis, run-length marking method, and grayscale thresholding method. Near-field camera equipment is complex, costly, and difficult to operate. The movement of industrial cameras is extremely accurate when photographing pixel-level light holes. The recursive search involved in the connected domain recognition algorithm requires a large amount of calculation. In comparison, image processing based on run-length technology can significantly reduce the total number of labels, effectively reduce the complexity of parsing label equivalence relations, and speed up algorithm processing. However, in some more complex images, there are still chain problems and rely on dedicated FPGA hardware architecture.

[0004] In traditional VCSEL chip inspection, camera calibration is required before photographing the wafer, and images are mostly binarized, which can easily lead to unclear classification results and poor fault tolerance. In addition, more manual participation is required during the test process, and the degree of automation is low. It is difficult to meet the quality analysis of large-scale VCSEL chip product inspections and the tracing of local information in the wafer. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a VCSEL semiconductor laser array near-field light-emitting hole detection device and method, which adopts a light source camera to automatically identify the vertical cavity surface spot of the VCSEL semiconductor laser, and performs instance segmentation based on the Mask RCNN network, thereby reducing manual adjustment in the process, improving the parallel processing capability of the semiconductor laser near-field test, and realizing the quality classification of the light-emitting holes.

[0006] The technical solution of the present invention is as follows:

[0007] A VCSEL semiconductor laser array near-field light-emitting hole detection device, comprising a wafer tray 1, a light source camera 4, a probe shaft device 6, a processor, a memory, an industrial computer and an adjustment bracket 5; the wafer tray 1 is fixed to a base through a stepping motor; a door-shaped frame is provided on the base; the top of the door-shaped frame is provided with a slideway, and the bottom of the adjustment bracket 5 moves on the slideway; the light source camera 4 is fixed to the side end of the adjustment bracket 5, and is located above the wafer tray 1; the probe shaft device 6 includes a coordinate conversion unit and a rotation unit, which are used to achieve position matching between the VCSEL semiconductor laser and the probe pin 2; one end of the probe shaft device 6 is fixed to the door-shaped frame, which is located above the wafer tray 1, and the probe pin 2 is located above the wafer tray 1. The other end of the shaft device 6 is connected to the probe pin 2 through the probe connecting rod 3, and the probe pin 2 is in contact with the VCSEL semiconductor laser to be tested; the processor, memory, and industrial computer are placed side by side under the wafer tray 1, and the processor and memory transmit data through the bus. After the processor completes data processing, it is connected to the industrial computer through a signal transmission line, and the industrial computer is connected to the adjustment bracket 5 through the signal transmission line; the processor stores a trained instance segmentation model based on the Mask-RCNN network; the memory is used to store sample data and related configuration files of the near-field light-emitting hole; the industrial computer is used to control the height of the light source camera 4, the movement of the adjustment bracket 5 and the rotation of the probe connecting rod 3.

[0008] The probe pin 2 rotates around the probe shaft device 6 through the probe connecting rod 3. According to different required triggering modes, the probe pin 2 is divided into a single pin, a double pin or a four pins.

[0009] A VCSEL semiconductor laser array near-field light-emitting hole detection method comprises the following steps:

[0010] Step 1: The light-emitting hole of the VCSEL semiconductor laser array faces upward and is placed horizontally in the wafer tray 1. The light source camera 4 emits light vertically to irradiate and scan the VCSEL semiconductor laser array. The light intensity can ensure clear wafer imaging and obtain the spatial position information and angle information of the VCSEL semiconductor laser array to be tested; after the scanning is completed, a MAP map at any angle is obtained;

[0011] Step 2: Determine the deviation between the probe pin 2 and the VCSEL semiconductor laser according to the spatial position information and angle information of the VCSEL semiconductor laser array in step 1; determine whether the probe pin has moved to the first chip position according to the position of the positioning mark in the two-dimensional image of the wafer in step 1. When the positioning mark and the reference coordinates of the probe pin are inconsistent, use the coordinate conversion unit in the probe axis device 6 to calculate the angle and distance between the spatial coordinates of the VCSEL semiconductor laser to be tested and the spatial coordinates of the probe. The rotating unit in the probe axis device 6 drives itself to rotate so that the probe pin can be perpendicular to the wafer surface.

[0012] During the motion control process, a rectangular coordinate system is established based on the MAP map and the probe axis device 6, and the wafer tray 1 is controlled to move through the set stepping command so that the next VCSEL semiconductor laser position matches the position of the probe pin 2. The VCSEL semiconductor laser receives the trigger pulse signal sent by the probe pin 2 and lights up the light-emitting hole;

[0013] Step 3: Obtain the near-field images of semiconductor lasers in different experimental batches and the corresponding operation files, and construct a sample library of near-field light-emitting hole images of various models and specifications; the sample library can show the outline, size, brightness, position, number and pixel points of the light-emitting holes in the near-field images; the images are mainly grayscale images, but not limited to grayscale images; the operation files have corresponding scale information.

[0014] Step 4: Perform instance segmentation and annotation on the original images in the near-field light-emitting hole image sample library constructed in step 3 according to product model specifications and grayscale values ​​to construct a near-field light-emitting hole image dataset;

[0015] The near-field luminous hole image dataset is used to create a training set and a test set in a ratio of 8:2 to train the Mask RCNN classification network model. The Mask RCNN classification network model is saved after training, and subsequent luminous hole image classification is performed based on this model. The training set contains the location and quantity information of the luminous holes in different production batches, and the grayscale threshold can cover most areas.

[0016] Step 5: Send the original image captured by the light source camera 4 to the Mask RCNN network trained in step 4 for instance segmentation;

[0017] Step 6: Perform quality analysis on the luminous holes after instance segmentation, perform error analysis on the roundness in all near-field luminous hole images, and obtain the luminous hole center of mass position offset; perform uniformity judgment on the luminous holes after the above instance segmentation, add up the grayscale sum of all pixels in the actual size of the luminous hole spot, and obtain the brightness variance of the luminous point in each chip.

[0018] In the step 1, the light source camera 4 moves in the vertical direction to emit a vertical light beam; after completing the acquisition of the two-dimensional image information of the wafer surface, the VCSEL semiconductor laser can generate a MAP map at different placement angles.

[0019] The Mask RCNN classification network model includes a backbone feature extraction network including a feature pyramid network, a region proposal network, a region of interest layer, a classification and bounding box branch, and a mask branch; random gradient descent training is used to obtain the optimal solution of the loss function of the Mask RCNN classification network model;

[0020] The backbone feature extraction network adopts the form of ResNet101+FPN, and the output channel is set to n-dimensional; ResNet101 represents a residual neural network with 101 hidden layers and the fully connected layer removed, and FPN is a feature pyramid network; the region proposal network shares the feature extraction network with the backbone feature extraction network, adds a fully connected network layer to obtain the candidate region location information and foreground and background classification, and uses non-maximum suppression to remove overlapping region proposals; the region of interest layer uses ROI Align to obtain candidate region features from the backbone feature extraction network and the region proposal network and outputs 7×7 sampling features to the classification and bounding box branches, and outputs 14×14 sampling features to the Mask branch; the classification and bounding box branches are implemented by the convolutional neural network CNN, that is, the convolutional neural network CNN is used to perform category classification and bounding box regression of the candidate region after ROI Align processing; the Mask branch is implemented by the full convolutional neural network, that is, the full convolutional neural network is used to perform Mask regression on the candidate region after ROI Align processing;

[0021] The specific process of the Mask RCNN classification network model is as follows:

[0022] Input a near-field original light-emitting hole image collected by the light source camera 4, and perform data preprocessing and normalization; pass the processed near-field original light-emitting hole image into the backbone feature extraction network to obtain the corresponding feature map; set the region of interest layer for each point in the feature map, obtain multiple region of interest candidate frames, and calculate the relative step distance with the original image for each point in the feature map; input the region of interest candidate frame into the region proposal network for binary classification, classify it as foreground or background, filter out the region of interest of the candidate area that exceeds the pixel size of the original light-emitting hole image, perform ROI Align operation on the remaining regions of interest, correspond the pixels of the near-field original light-emitting hole image and the feature map, and correspond the feature map to the fixed features; the remaining regions of interest are processed by the full convolutional neural network to classify all light-emitting hole categories and generate Masks.

[0023] The stochastic gradient descent optimizes the multi-task loss function L

[0024] Loss = L rpn +L fast-rcnn +L mask

[0025] Among them, L rpn is the candidate box loss, L fast-rcnn is the backbone feature network loss, L mask It is the mask branch loss, and a two-stage learning rate adjustment strategy is introduced in network optimization.

[0026] The specific process of the Mask branch is as follows: the candidate box result with an input size of H*W*256 is changed to 14*14*256 after pooling in the ROI Align layer; then it passes through 4 convolutional layers, the convolution kernel of each convolutional layer is 3*3, the convolution kernel scanning step is set to 1, and the edge filling size of the result image after each convolutional layer is set to 1; a RELU function is set after each convolutional layer; through a transposed convolution, the height and width of the input feature are doubled; finally, through a 1*1 convolution kernel, the number of convolution kernels is equal to the number of classifications NUM, and finally an output of 28*28*NUM is obtained.

[0027] In step six, the near-field light-emitting hole after instance segmentation is converted to an actual size according to the step size of the feature layer relative to the original light-emitting hole image.

[0028] The roundness error analysis uses the least squares circle method to calculate the roundness error.

[0029] The least squares model of roundness is established in the rectangular coordinate system. Let (x i ,y i ), where i = 1, 2, ..., n (n> 3) are the measurement sampling points on the actual circumference of the measured circle. Suppose the equation of the least squares circle to be calculated is (xa) 2 +(yb) 2 =R 2 , where (a, b) is the center of the least squares circle, and R is the radius of the least squares circle. According to the above equation, the radial deviation from the sampling point to the least squares circle can be obtained as

[0030]

[0031] The coordinates of the center of the least squares circle are calculated as:

[0032] The roundness error is expressed as

[0033]

[0034] According to the calculation results of the above steps, as long as f≤t is satisfied, it means that the actual contour of the light-emitting hole does not exceed the tolerance range, and the quality can be considered qualified. The centroid is the point whose horizontal coordinate and vertical coordinate of the above-mentioned even-numbered points are the average horizontal coordinate and the average vertical coordinate of the n points respectively. Assuming that the coordinates of the n points are (x1, y1), (x2, y2),..., the coordinates of the centroid are ((x1+x2+...) / n, (y1+y2+...) / n). By calculating the distance from the center coordinate of the least squares circle to the center coordinate, the detailed offset information of each light-emitting hole can be obtained. By calculating the variance between the grayscale sum of pixels in all light-emitting holes, the uniformity of each light-emitting hole in the chip can be obtained.

[0035] The present invention has the following beneficial effects:

[0036] (1) The present invention uses a rotatable probe and a light source camera to position and test the laser array, which saves a lot of time compared to traditional manual adjustment, has a high degree of automation, and reduces errors caused by manual operation;

[0037] (2) The features of multiple near-field images are extracted based on the Mask RCNN network, and instance segmentation is performed on each light-emitting hole in each image. The data sets of VCSEL semiconductor lasers of different models are used for training, which improves the accuracy of quality judgment of each light-emitting hole in the near field of VCSEL semiconductor lasers. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the near-field image of the VCSEL semiconductor laser array described in the present invention.

[0039] Figure 2 It is a schematic diagram of a grayscale image of a near-field light-emitting hole of a VCSEL semiconductor laser described in the present invention.

[0040] Figure 3 This is a schematic diagram of a MAP map obtained after scanning by the light source camera in the present invention.

[0041] Figure 4 Schematic diagram of the processing network flow of the collected original images.

[0042] Figure 5 This is a flowchart of the Mask branch of the statistical method for near-field luminous hole classification based on Mask RCNN instance segmentation.

[0043] Figure 6 It is a schematic diagram of the rotatable probe station testing device described in the present invention.

[0044] In the figure: 1 is a wafer tray; 2 is a probe pin; 3 is a probe connecting rod; 4 is a light source camera; 5 is an adjustment bracket; and 6 is a probe shaft device. DETAILED DESCRIPTION

[0045] Figure 2 Examples of near-field light-emitting hole images in the semiconductor laser array sample library involved in the present invention are shown. Figure 5 The flow chart of an embodiment of the Mask branch of the statistical method for instance segmentation of near-field luminous hole classification based on the Mask RCNN model is shown. Figure 6 The schematic diagram of VCSEL semiconductor laser array near-field light-emitting hole detection is shown. Figure 2 , Figure 5 and Figure 6 , the implementation steps of the method of this embodiment are described in detail as follows.

[0046] Step 1: The light source camera 4 can be installed perpendicular to the wafer tray 1 and can emit a parallel light beam that propagates freely in the air. After the image is clear, the position coordinates of the VCSEL semiconductor laser array wafer with positioning marks are obtained, and the two-dimensional information of the VCSEL semiconductor laser array wafer is captured, and a MAP map is established after the camera scans.

[0047] If the project requires that the minimum feature size to be detected and the number of pixels representing the minimum feature remain unchanged, then when the lens focal length is fixed, the minimum resolution of the selected camera should be proportional to the field of view size. Once the camera (resolution and sensor size) and lens focal length are selected, the project requirements also limit the maximum field of view of the imaging system and the maximum installation distance between the camera and the observed target. In this case, the closer you are to the target, the more pixels you can use to represent the minimum feature.

[0048] In this step, the above lenses can be used as long as the field of view can cover the test target well. After the lens and camera are selected, the working distance should be adjusted within the maximum field of view and maximum installation distance specified by the project requirements so that the field of view of the camera is

[0049] FOV = L object +l dynamic +l buffer

[0050] Among them, L object is the maximum length of the detected target, l dynamic is the maximum distance the target may move, l buffer is the reserved safety distance, usually equal to one tenth of L object In this way, more pixels can be used to represent the smallest features of the target while ensuring that the target is captured by the camera, so as to obtain a near-field image with higher definition.

[0051] Step 2: In this step, the probe axis device mainly includes a rotation unit and a coordinate conversion unit. According to the position of the positioning mark in the two-dimensional image of the wafer described in step 1, it is judged whether the probe pin has moved to the first chip position. When the positioning mark and the reference coordinates of the probe pin are inconsistent, the coordinate conversion unit is used for adjustment. The rotation unit drives the probe axis device to rotate synchronously so that the probe pin can be perpendicular to the wafer surface.

[0052] The wafer tray 1 is moved by a stepper motor to adjust the relative position between the VCSEL chip wafer and the probe shaft device so that each chip in the VCSEL chip wafer can be aligned with the probe pin.

[0053] The movement of the probe pin adopts a soft triggering method. A rectangular coordinate system is established according to the MAP map described in step 1 and the probe axis device of the operation table. During the motion control process, the wafer tray is automatically moved by the pre-set stepping command so that the next chip position is aligned with the probe pin to apply a trigger signal to the chip.

[0054] The specific scheme of the above coordinate transformation unit can be divided into: assuming that the probe axis device of the operating table is the origin, the direction parallel to the probe axis device and the MAP map is the positive direction of the X-axis, the angle between the positioning mark and the reference direction of the probe pin is α, and the angle between the probe connecting rod and the positive direction of the X-axis is β. Therefore, X1 = r*cos(α+β), and the sum-difference-product formula can be obtained that X1 = X*cosα-Y*sinα, and similarly, Y1 = X*sinα+Y*cosα, which is expressed in matrix form as follows:

[0055]

[0056] Among them, (X, Y) is the current position of the probe pin, (X1, Y1) is the absolute position coordinate of the first chip in the wafer. r is the length of the probe connecting rod. Through the coordinate conversion matrix, the first chip position is found, and according to the stepping instruction, the probe can automatically light up each semiconductor light emitting device in turn.

[0057] Step 3: Obtain the near-field images and corresponding operation files of semiconductor lasers in different experimental batches, and build a near-field image sample library of various models and specifications. During the annotation process, the sample library is appropriately classified according to specific needs. Generally, the grayscale value range between 144 and 220 is considered good. According to the product model specifications and grayscale values, the luminous holes are divided into dark spots, sub-dark spots, good spots, bright spots, and over-bright spots. Non-luminous points are considered bad points, and their contours are annotated. In order to adapt to the data loading interface of Mask RCNN, the annotated labels are uniformly saved as json files, and the corresponding classification is based on the picture.

[0058] In this step, the images captured in the near field and the corresponding operation files are sorted into corresponding folders according to the experimental batches, and then the sorting results are simply screened, including eliminating samples with missing image information or missing operation file information.

[0059] The near-field light-emitting hole image and the corresponding operation file in this step are generated during the two-dimensional image acquisition operation of the semiconductor laser array by the light source camera, and the image is mainly a grayscale image, but not limited to a grayscale image. In addition, the operation file has corresponding scale information.

[0060] For example, the industrial camera has a pixel width of 2448 and a pixel height of 2048, and the image pixel width is proportionally transformed to 2448, and the image pixel height is proportionally transformed to 2048. During the transformation, the X coordinate of the cross centerline is 1224, the Y coordinate of the cross centerline is 1024, the width and height are linearly proportionally transformed, and the proportional value is recorded, which is not limited in the embodiment of the present invention.

[0061] Step 4: Perform instance segmentation and annotation on the original images in the above-mentioned near-field light-emitting hole image sample library, and construct a near-field light-emitting hole image instance segmentation training dataset and a test dataset.

[0062] In this step, in the near-field luminous hole image sample library constructed in step three, an appropriate amount of samples are selected to construct labeled samples. Preferably, group sampling is performed in each different production batch in the near-field luminous hole image sample library. It will be randomly divided into training set and test set in a ratio of 8:2. The training set contains the location and quantity information of the luminous holes in different production batches, and the grayscale threshold can cover most areas. The main body of the sample here is the near-field luminous hole image and its corresponding operation file.

[0063] In this step, the instance segmentation and annotation process is to annotate the grayscale value of the circular hole in the near-field light-emitting hole image by thresholding. Class labels are established for bad points, dark points, sub-dark points, good points, bright points, and over-bright points. After the image is pre-trained, the loss function is optimized using stochastic gradient descent. The multi-task loss function L is optimized.

[0064] Loss = L rpn +L fast-rcnn +L mask

[0065] Among them, L rpn is the candidate box loss, L fast-rcnn is the backbone feature network loss, L mask It is a Mask branch loss containing FCN, and a two-stage learning rate adjustment strategy is introduced in network optimization.

[0066] Step 5: Based on the Mask RCNN instance segmentation model, the multi-branch neural network set in the model includes a ResNet101 feature extraction network with a feature pyramid network (FPN), a region proposal network (RPN), a region of interest layer (ROI), a classification and border branch, and a mask branch; among them, the feature pyramid network is used to improve the overall segmentation accuracy of the instance segmentation network. The specific operation process is as follows: input a near-field original image collected by a light source camera, perform data preprocessing and normalization operations, etc.; pass the processed light hole image into the pre-trained ResNet neural network to obtain the corresponding feature map. Set ROI for each point in the feature map to obtain multiple ROI candidate frames; send these multiple ROI candidate frames to RPN for binary classification (foreground or background), filter out some candidate ROIs; perform ROI Align operation on the remaining ROIs, that is, first correspond the pixels of the original image and the feature map, and then correspond the feature map to the fixed features; classify all light hole categories and generate Masks for these ROIs, that is, perform FCN operation in each ROI.

[0067] The detailed operation is as follows: the original image is annotated with candidate boxes, the candidate boxes of each light-emitting hole are mapped to the feature layer, and the step length of the feature layer relative to the original image is obtained. ROI Align is used to map the target candidate box to the feature layer, and the output value of the target candidate box is obtained using bilinear interpolation. When multiple sampling points are used, the output of each sub-region depends on the values ​​of all sampling points.

[0068] For example, if the sampling ratio is set to 2, the output of all feature points in the feature map is extracted. The sampling points in each light hole candidate frame area are the square of the sampling ratio, which is equal to 4. Then the output value K of the target candidate frame can be calculated by the following formula:

[0069] K=(1-α)(1-β)K1+α(1-β)K2+β(1-α)K3+αβK4

[0070] Among them, K1, K2, K3, and K4 represent the feature values ​​of the four sampling points in the target candidate box output value, and α and β represent the horizontal and vertical distances from the center point of the target candidate box output value to the border composed of the four feature points, respectively.

[0071] The backbone feature extraction network adopts the form of ResNet101+FPN, and the output channel is set to n dimensions (for example, 256 dimensions); wherein ResNet101 represents a residual neural network with 101 hidden layers, and the fully connected layer is removed in this application, and FPN is a Feature Pyramid Network feature pyramid network, which introduces strong semantic information obtained by downsampling high-level features on the basis of the features of each stage of the ResNet101 feature extraction network, so that the features of each stage of the ResNet101 network retain the corresponding resolution and have strong semantic information;

[0072] The region proposal network (RPN) shares a feature extraction network with the backbone network, adds a fully connected network to infer candidate region location information and foreground and background classification, and uses non-maximum suppression (NMS) to remove overlapping region proposals;

[0073] The region of interest (ROI) uses ROI Align to obtain candidate region features from the backbone feature extraction network and RPN and outputs 7×7 sampling features to the classification and bounding box branches, and outputs 14×14 sampling features to the Mask branch;

[0074] The classification and border branches are implemented by a convolutional neural network (CNN), that is, the convolutional neural network (CNN) is used to perform category classification and border regression of the luminous hole on the candidate area after ROI Align processing;

[0075] The Mask branch is implemented by a fully convolutional neural network FCN, that is, the fully convolutional neural network FCN is used to perform Mask regression on the candidate area after ROIAlign processing. Specifically, it is first pooled by ROL Align to become 14*14*256 in size, and then passes through 4 convolutional layers. The kernel of each convolutional layer is 3*3, the step size is 1, and the surrounding padding value is set to 1. A RELU function is set after each convolutional layer. Next, a transposed convolution is used to double the height and width of the input feature. Finally, a 1*1 convolution kernel is used, and the number of convolution kernels is equal to the number of light-emitting hole classifications NUM set, and finally an output of 28*28*NUM is obtained.

[0076] In this step, the Mask RCNN model training uses the near-field luminous hole dataset constructed from the sample library, and uses an end-to-end approach to complete the luminous hole instance segmentation of the near-field image samples using the trained Mask RCNN instance segmentation model.

[0077] In this step, the target of the Mask branch input during training is the candidate region provided by RPN (Region Proposal Network, used to generate candidate regions). Each type of luminous hole predicted by the network is the Mask target information of each category corresponding to each candidate region, and the predicted Mask target size is 28x28. The category of the Proposal (target region) can be known through the positive and negative sample matching process of the Fast RCNN branch.

[0078] A candidate rectangular box is obtained through RPN, and the corresponding feature information is obtained after RoI Align. The size is 14x14. The candidate boxes input to the Mask branch are all positive samples sampled in the Fast RCNN stage. The Mask branch predicts the Mask information of each category to obtain the logits (output of network prediction) of various light-emitting hole classifications. After the logits pass through the sigmoid activation function, all values ​​are mapped to between 0 and 1.

[0079] Extract the 28x28 predicted Mask of the corresponding category in logits. Then crop and scale the GT (ground truth, classification accuracy) corresponding to the original image to 28x28 size according to Proposal, and get the GTMask in the figure (corresponding to the target area is 1, and the background area is 0). Finally, calculate the cross entropy loss (BinaryCrossEntropyLoss) between the Mask of the predicted category in logits and the GT Mask, and find the optimal solution.

[0080] Step 6: Based on the instance segmentation result of the near-field light-emitting hole image obtained in step 5, multiply it by the step ratio to restore the instance segmentation result to the original image size. Perform error analysis on the roundness of the restored near-field light-emitting hole image, and calculate the centroid position offset based on the chip design parameters, thereby realizing the detection of the light-emitting hole quality and accurate analysis of local information.

[0081] The segmentation result restored to the original image size contains the same pixels as the original image. In one embodiment, the number of light-emitting holes is 200, and the grayscale value of all pixels in each light-emitting hole is recorded as v i , the mean of the grayscale value and is recorded as in,

[0082] In this step, the roundness tolerance refers to the variation weight t allowed for the shape of the actual circular hole on the same light-emitting hole spot. The roundness tolerance zone refers to the area between two concentric circles with a radius difference of the tolerance value t on the same light-emitting hole spot. The roundness error is the distance between two concentric circles that contain the actual contour of the same light-emitting hole spot and have the smallest radius difference. The roundness error value is recorded as f. In this embodiment, the least squares model of roundness in a rectangular coordinate system is adopted for evaluation.

[0083] The specific processing process of roundness error analysis is as follows:

[0084] Step 6a: Establish the least squares model of roundness in the rectangular coordinate system. Let (x i ,y i ), where i = 1, 2, ..., n (n> 3) are the measurement sampling points on the actual circumference of the measured circle. Suppose the equation of the least squares circle to be calculated is (xa) 2 +(yb) 2 =R 2 , where (a, b) is the center of the least squares circle, and R is the radius of the least squares circle. According to the above equation, the radial deviation from the sampling point to the least squares circle can be obtained as

[0085]

[0086] Step 6b: In the actual measurement and evaluation of roundness error, the measurement sampling points are required to meet the following constraints: the measurement sampling points are evenly spaced on the measured circle; the number of measurement sampling points is an even number. Under this constraint, the least squares circle center coordinates can be expressed as:

[0087] Step 6c: Assume that the points with the largest and smallest distances to the center of the least squares circle are (x m ,y m ) and (x l ,y l ), the roundness error can be obtained by the following formula

[0088]

[0089] According to the calculation results of the above steps, if f≤t is satisfied, it means that the actual contour of the light-emitting hole does not exceed the tolerance band, and the quality can be considered qualified. On the contrary, the actual contour of the light-emitting hole does not exceed the tolerance band. In this embodiment, the centroid coordinates are obtained based on the points whose horizontal coordinates and vertical coordinates are the average values ​​of the horizontal coordinates and vertical coordinates of n points of an even number of points, respectively. Assuming that the coordinates of the n points are (x1, y1), (x2, y2),..., the coordinates of the centroid are ((x1+x2+…) / n, (y1+y2+…) / n). By calculating the distance from the center coordinates of the least squares circle to the centroid coordinates, the detailed offset information of each light-emitting hole can be obtained.

[0090] The specific processing process of uniformity judgment is as follows:

[0091] Using the Compensated Algorithm, the grayscale value and data of a single light-emitting hole are scanned twice. The variance value can be obtained by the following formula

[0092]

[0093] like When it is an exact value, the second term is zero, otherwise, the rounding error of the first term can be corrected. The calculation result can be made more accurate by using the CompensatedAlgorithm compensation algorithm.

Claims

1. A VCSEL semiconductor laser array near-field light-emitting hole detection device, characterized in that: The VCSEL semiconductor laser array near-field light-emitting hole detection device comprises a wafer tray (1), a light source camera (4), a probe shaft device (6), a processor, a memory, an industrial computer and an adjustment bracket (5); the wafer tray (1) is fixed to a base via a stepping motor; a door-shaped frame is provided on the base; the top of the door-shaped frame is provided as a slideway, and the bottom of the adjustment bracket (5) moves on the slideway; the light source camera (4) is fixed to the side end of the adjustment bracket (5) and is located above the wafer tray (1); the probe shaft device (6) comprises a coordinate conversion unit and a rotation unit, and is used to achieve position matching between the VCSEL semiconductor laser and the probe pin (2); one end of the probe shaft device (6) is fixed to the door-shaped frame, and is located above the wafer tray (1). The other end of the probe shaft device (6) is connected to the probe pin (2) through the probe connecting rod (3), and the probe pin (2) is in contact with the VCSEL semiconductor laser to be tested; the processor, the memory, and the industrial computer are placed side by side under the wafer tray (1), the processor and the memory transmit data through a bus, and after the processor completes data processing, it is connected to the industrial computer through a signal transmission line, and the industrial computer is connected to the adjustment bracket (5) through the signal transmission line; the processor stores a trained instance segmentation model based on the Mask-RCNN network; the memory is used to store sample data and related configuration files of the near-field light-emitting hole; the industrial computer is used to control the height of the light source camera (4), the movement of the adjustment bracket (5) and the rotation of the probe connecting rod (3).

2. The VCSEL semiconductor laser array near-field light-emitting hole detection device according to claim 1, characterized in that: The probe pin (2) rotates around the probe shaft device (6) via the probe connecting rod (3). Depending on the desired triggering mode, the probe pin (2) is divided into a single pin, a double pin or a quad pin.

3. A VCSEL semiconductor laser array near-field light-emitting hole detection method, characterized in that: The method is implemented by using a VCSEL semiconductor laser array near-field light-emitting hole detection device as described in any one of claims 1-2, and comprises the following steps: Step 1: The VCSEL semiconductor laser array is placed horizontally in a wafer tray (1) with its light-emitting hole facing upward, and the light source camera (4) emits light vertically to irradiate and scan the VCSEL semiconductor laser array to obtain spatial position information and angle information of the VCSEL semiconductor laser array to be tested; after the scanning is completed, a MAP map at any angle is obtained; Step 2: judging the deviation between the probe pin (2) and the VCSEL semiconductor laser according to the spatial position information and angle information of the VCSEL semiconductor laser array in step 1; using the coordinate conversion unit in the probe shaft device (6), calculating the angle and distance between the spatial coordinates of the VCSEL semiconductor laser to be measured and the spatial coordinates of the probe, and the rotating unit in the probe shaft device (6) drives itself to rotate; During the motion control process, a rectangular coordinate system is established based on the MAP map and the probe axis device (6), and the wafer tray (1) is controlled to move through the set stepping command so that the position of the next VCSEL semiconductor laser matches the position of the probe pin (2). The VCSEL semiconductor laser receives a trigger pulse signal from the probe pin (2) and lights up the light-emitting hole; Step 3: Obtain the near-field images and corresponding operation files of semiconductor lasers in different experimental batches, and construct a sample library of near-field light-emitting hole images of various models and specifications; the sample library can show the contour, size, brightness, position, number and pixel points of the light-emitting holes in the near-field images; Step 4: Perform instance segmentation and annotation on the original images in the near-field light-emitting hole image sample library constructed in step 3 according to product model specifications and grayscale values ​​to construct a near-field light-emitting hole image dataset; Create a training set and a test set with a ratio of 8:2 for the near-field luminous hole image dataset, and train the Mask RCNN classification network model; save the trained Mask RCNN classification network model, and perform subsequent luminous hole image classification based on this model; Step 5: Send the original image captured by the light source camera (4) to the Mask RCNN network trained in step 4 for instance segmentation; Step 6: Perform quality analysis on the luminous holes after instance segmentation, perform error analysis on the roundness in all near-field luminous hole images, and obtain the luminous hole center of mass position offset; perform uniformity judgment on the luminous holes after the above instance segmentation, add up the grayscale sum of all pixels in the actual size of the luminous hole spot, and obtain the brightness variance of the luminous point in each chip.

4. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 3, characterized in that: In the step 1, the light source camera (4) moves in the vertical direction to emit a vertical light beam; after completing the acquisition of two-dimensional image information of the wafer surface, the VCSEL semiconductor laser can generate a MAP map at different placement angles.

5. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 3, characterized in that: The Mask RCNN classification network model includes a backbone feature extraction network including a feature pyramid network, a region proposal network, a region of interest layer, a classification and bounding box branch, and a mask branch; random gradient descent training is used to obtain the optimal solution of the loss function of the Mask RCNN classification network model; The backbone feature extraction network adopts the form of ResNet101+FPN, and the output channel is set to n-dimensional; ResNet101 represents a residual neural network with 101 hidden layers and the fully connected layer removed, and FPN is a feature pyramid network; the region proposal network shares the feature extraction network with the backbone feature extraction network, adds a fully connected network layer to obtain the candidate region location information and foreground and background classification, and uses non-maximum suppression to remove overlapping region proposals; the region of interest layer uses ROI Align to obtain candidate region features from the backbone feature extraction network and the region proposal network and output 7×7 sampling features to the classification and border branches, and outputs 14×14 sampling features to the Mask branch; The classification and bounding box branches are implemented by the convolutional neural network CNN, that is, the convolutional neural network CNN is used to perform the category classification and bounding box regression of the luminous hole on the candidate area after ROI Align processing; the mask branch is implemented by the full convolutional neural network, that is, the full convolutional neural network is used to perform the mask regression on the candidate area after ROI Align processing; The specific process of the Mask RCNN classification network model is as follows: Input a near-field original light-emitting hole image captured by a light source camera (4), perform data preprocessing and normalization; pass the processed near-field original light-emitting hole image into the backbone feature extraction network to obtain a corresponding feature map; set a region of interest layer for each point in the feature map, obtain multiple region of interest candidate frames, and calculate the relative step distance with the original image for each point in the feature map; The candidate boxes of the regions of interest are input into the region proposal network for binary classification, and they are classified as foreground or background. The regions of interest of the candidate regions that exceed the pixel size of the original luminous hole image are filtered out, and the remaining regions of interest are subjected to ROI Align operation to correspond the pixels of the near-field original luminous hole image and the feature map, and the feature map is corresponded to the fixed features; the remaining regions of interest are processed by the fully convolutional neural network to classify all luminous hole categories and generate masks.

6. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 5, characterized in that: The stochastic gradient descent optimizes the multi-task loss function L Loss=L rpn +L fast-rcnn +L mask Among them, L rpn is the candidate box loss, L fast-rcnn is the backbone feature network loss, L mask It is the mask branch loss, and a two-stage learning rate adjustment strategy is introduced in network optimization.

7. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 5, characterized in that: The specific process of the Mask branch is as follows: the candidate box result with an input size of H*W*256 is changed to 14*14*256 after pooling in the ROI Align layer; then it passes through 4 convolutional layers, the convolution kernel of each convolutional layer is 3*3, the convolution kernel scanning step is set to 1, and the edge filling size of the result image after each convolutional layer is set to 1; a RELU function is set after each convolutional layer; through a transposed convolution, the height and width of the input feature are doubled; finally, through a 1*1 convolution kernel, the number of convolution kernels is equal to the number of classifications NUM, and finally an output of 28*28*NUM is obtained.

8. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 5, characterized in that: In step six, the near-field light-emitting hole after instance segmentation is converted to an actual size according to the step size of the feature layer relative to the original light-emitting hole image.

9. The VCSEL semiconductor laser array near-field light-emitting hole detection method according to claim 3, characterized in that: The roundness error analysis uses the least squares circle method to calculate the roundness error.

Images