CN114495106A - A deep learning MOCR method applied to DFB laser chips - Google Patents

Abstract

The invention discloses a deep learning MOCR method applied to a DFB laser chip, which belongs to the field of image processing. Including: step 1: collect the image of the DFB chip online; step 2: preprocess the image; step 3: correct the direction of the characters on the chip through the character area detection network; step 4: use the character recognition network to correct the characters on the chip Recognize the content; Step 5: Determine whether the recognized character content is correct, if it is correct, store it in the associated database; if it is incorrect, remove the chip or perform manual rechecking. The invention can efficiently and accurately identify the character information on the DFB laser chip without changing the original production process flow. Compared with the traditional recognition method, the recognition accuracy is greatly improved, and the recognition accuracy can reach 98.8%.

Description

A deep learning MOCR method applied to DFB laser chips

technical field

The invention relates to the field of optoelectronic semiconductor microscopic visual inspection, in particular to a deep learning MOCR method applied to a DFB laser chip.

Background technique

Distributed feedback laser diodes (DFBs) are widely used in optical fiber communication systems due to their high side-mode rejection ratio and ultra-narrow spectral width. With the continuous development of communication capacity and communication bandwidth, laser chips are also becoming smaller and smaller, so as to facilitate integration and packaging into optical communication devices; with the rapid development of 5G communication, FTTR and data centers, the communication rate has also become Higher and higher. As the most critical core chip in the optical fiber communication system, the reliability and stability of the performance and quality of the DFB laser will become more and more important.

The continuous development of intelligent manufacturing and artificial intelligence has given powerful capabilities to the informatization, intelligence and traceability of manufacturing. Due to its importance in technology and price, DFB laser chips need to be able to trace the quality and control the process in the entire manufacturing process of optical devices. The key parameters of DFB laser chips such as optical power, extinction ratio, slope efficiency, central wavelength, crossover point and side mode suppression ratio need to be associated and corresponding to each chip one by one. As the character information on the DFB laser chip is the unique identification of the chip itself, its accurate and efficient identification will become particularly important.

However, the characters on the DFB laser chip are quite different and different from the traditional characters, mainly because the characters on the chip are particularly small, about less than 100um (the character area accounts for about 10% of the chip area), which requires microscopic imaging technology to image the target clearly; secondly, there is no contextual semantics between characters, which makes it impossible to use semantic recognition methods; thirdly, because the chip is small, its characters are not very standard fonts, which also brings certain difficulties to recognition. Coupled with the industrial environment and the interference of chips due to dirt, defects and dust, etc., it increases the difficulty of identification. All these problems and characteristics lead to the uniqueness and difficulty of character recognition of DFB laser chips.

Using the traditional character recognition method, the recognition accuracy rate is only about 60%, which leads to the need for special engineers in real-time human assistance to judge unrecognized characters at each recognition station. This is unsatisfactory in the real-time manufacturing line of optical communication devices. Seriously lead to waste of manpower and efficiency, and due to people's easy fatigue and emotions, there are also small loopholes and risks in quality control.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the deficiencies of the prior art, and to provide a deep learning automatic character area positioning and character recognition method applied to microscopic characters on a DFB laser chip in an optical device packaging process, which can be used without changing the original production. Efficiently and accurately identify the character information on the DFB laser chip under the condition of the technological process. Compared with the traditional recognition method, the recognition accuracy rate is greatly improved, and the recognition accuracy rate can reach 98.8%.

The purpose of this invention is to realize through the following technical solutions:

A deep learning MOCR method applied to a DFB laser chip, comprising the following steps:

Step 1: Collect the image of the DFB laser chip online;

Step 2: Preprocess the image;

Step 3: Correct the direction of the characters on the chip through the character area detection network;

Step 3.1: Perform downsampling on the input image four times in turn, sum the result of the fourth downsampling and the result of the third downsampling to obtain feature 1, and sum the result of feature 1 and the result of the second downsampling to obtain the feature Second, sum the result of feature two and the first downsampling to obtain feature three;

Step 3.2: Perform feature fusion on the result of the fourth downsampling, feature 1, feature 2, and feature 3 to obtain fused features;

Step 3.3: Perform convolution and deconvolution on the fusion feature to obtain a probability map and a threshold map;

Step 3.4: Differentiable binarization is performed on the probability map and the threshold map to obtain an approximate binarized image;

Step 3.5: Output the recognition frame according to the connected domain in the approximate binarized image, and convert the recognition frame to horizontal according to the length ratio of the recognition frame;

Step 3.6: Determine whether the text is reversed through the text direction classifier; if the text direction is reversed, rotate the text 180 degrees, and then input the character recognition network for recognition; if not, directly input the character recognition network for recognition;

Step 4: Recognize the character content on the chip through the character recognition network;

Step 5: Determine whether the recognized character content is correct, if correct, store it in the associated database; if not, remove the chip or perform manual rechecking.

Further, the data set used in the training of the character region detection network is formed by manually labeling the original image and performing data enhancement.

Further, the character recognition network is a fully convolutional neural network CRNN based on character recognition, which is used to recognize the characters contained in the character frame.

Further, the data set used in the character recognition network training is formed by manually labeling the original image and performing data enhancement.

Further, the equipment used to collect the image of the DFB laser chip online in the step 1 includes a light source, a CMOS camera and a telecentric lens, and the surface character image of the DFB laser chip is collected using an infrared ranging-based autofocus technology.

Beneficial effects of the present invention: The purpose of the present invention is to provide a deep learning automatic character area positioning and character recognition method applied to the microscopic characters on the DFB laser chip during the packaging process of the optical device, which can be used without changing the original production process flow. It can efficiently and accurately identify the character information on the DFB laser chip under different circumstances. Compared with the traditional recognition method, the recognition accuracy rate is greatly improved, and the recognition accuracy rate can reach 98.8%.

Description of drawings

Fig. 1 is a flow chart of image character area location and recognition.

Figure 2 is a diagram of the character area detection network architecture.

Figure 3 is a character recognition network architecture diagram.

Detailed ways

It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

In this embodiment, as shown in FIG. 1 , a deep learning MOCR method applied to a DFB laser chip includes the following steps:

A deep learning MOCR method applied to a DFB laser chip, comprising the following steps:

Step 1: Collect the image of the DFB laser chip online;

Step 2: Preprocess the image;

Step 3: Correct the direction of the characters on the chip through the character area detection network;

Step 3.1: Perform downsampling on the input image four times in turn, sum the result of the fourth downsampling and the result of the third downsampling to obtain feature 1, and sum the result of feature 1 and the result of the second downsampling to obtain the feature Second, sum the result of feature two and the first downsampling to obtain feature three;

Step 3.2: Perform feature fusion on the result of the fourth downsampling, feature 1, feature 2, and feature 3 to obtain fused features;

Step 3.3: Perform convolution and deconvolution on the fusion feature to obtain a probability map and a threshold map;

Step 3.4: Differentiable binarization is performed on the probability map and the threshold map to obtain an approximate binarized image;

Step 3.5: Output the recognition frame according to the connected domain in the approximate binarized image, and convert the recognition frame to horizontal according to the length ratio of the recognition frame;

Step 3.6: Determine whether the text is reversed through the text direction classifier; if the text direction is reversed, rotate the text 180 degrees, and then input the character recognition network for recognition; if not, directly input the character recognition network for recognition;

Step 4: Recognize the character content on the chip through the character recognition network;

Step 5: Determine whether the recognized character content is correct, if correct, store it in the associated database; if not, remove the chip or perform manual rechecking.

Among them, the character area detection network is the DBNet deep learning network, and the DBNet deep learning network is optimized in combination with the character features of the DFB laser chip; the optimization is based on the latest DBNet model, which is responsible for detecting the character area in the image. The box is converted to a horizontal orientation. A text orientation classifier is then used to determine if the text is reversed. If it is, the text needs to be rotated 180 degrees to be recognized again. Furthermore, the final character recognition method uses a fully convolutional character-based recognition model to process the detected character boxes and identify the characters contained within them.

Among them, the data set used in the training of the DBNet deep learning network is formed by manually labeling the original images and performing data enhancement.

Among them, the character recognition network is a fully convolutional neural network CRNN based on character recognition, which is used to recognize the characters contained in the character frame.

The dataset used to train the character recognition network is formed by manually annotating the original images and performing data augmentation.

The MOCR method in the present invention is divided into two stages: text detection and text recognition. The detection method is based on the latest DBNet deep learning model, and on this basis, the model is optimized in combination with the character features of the DFB laser chip. The optimized model is responsible for detecting the character area frame in the image and turning the text frame into a horizontal direction, as shown in Figure 2 for the character area localization network model. A text orientation classifier is then used to determine whether the text is reversed, where "1/2", "1/4" and "1/32" in the figure represent the scale compared to the input image. "pred" consists of two stride 2 deconvolution operators and a 3x3 convolution operator.

If the orientation is reversed, the text needs to be rotated 180 degrees to be recognized again. Furthermore, the final character recognition method uses a fully convolutional neural network character-based recognition model to process the detected character boxes and identify the characters contained within them. The identification network is shown in Figure 3.

In this embodiment, the equipment used for online acquisition of the image of the DFB laser chip includes a light source, a CMOS camera and a telecentric lens, and the surface character image of the DFB laser chip is acquired using an infrared ranging-based autofocus technology. The COMS camera adopts the area array COMS image sensor (MER-502-79U3C) produced by Daheng Company (resolution: 2448*2048, frame rate: 79fps, image size: 3.45um; optical interface: C interface, data interface: USB 3.0); using Canrui telecentric lens (XF-T6X65D) (magnification: 6.0X, object field: Φ1.3mm, depth of field: 0.06, telecentricity: <0.02°).

The size of the characters on the surface of the DFB laser chip is 70um*80um. Such characters are represented as microcharacters with a maximum dimension of less than 100 microns, as they are almost invisible to the naked eye. At the same time, we have adopted autofocus technology based on infrared ranging, which ensures the real-time and high efficiency of the production line, and minimizes the impact of defocused images caused by environmental and batch errors.

In this embodiment, since the original image is an image of 2448*2048 pixels, in which the character area only occupies a small part, a two-stage deep learning network model training method (detection and recognition) is adopted; The images are labeled to obtain the images and corresponding labels required by the detection and recognition network.

First, mark the character area in the original image with a rectangular frame, and record the four coordinate points clockwise from the upper left corner. After that, the coordinate position of the rectangular box in the original image is obtained, that is, the label required by the detection network. The rectangular area of the original image is then cropped and the text content is recorded to obtain the pictures and labels needed to recognize the network. Finally, we replace the text content of the labels used for recognition with 0 and 180 degrees to get the labels required by the orientation classifier.

The above is a description of the specific method for MOCR to realize character recognition of DFB laser chip. The method has been successfully applied to the packaging and inspection production line of DFB optical devices. A large number of practical use data show that the method has good robustness and recognition accuracy. At present, the recognition accuracy rate is about 98.8%. The first realization and successful application of MOCR has played a key role in the intelligent manufacturing and quality traceability of DFB laser chips and devices in the field of optical communication, and is also of great significance for further accelerating the development of the Industrial Internet.

It should be noted that, for the sake of simple description, the foregoing method embodiments are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the described action sequence. Because in accordance with the present application, certain steps may be performed in other orders or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and units involved are not necessarily required by the present application.

In the above-mentioned embodiments, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference may be made to related descriptions of other embodiments.

Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and the program is During execution, it may include the processes of the embodiments of the above-mentioned methods. Wherein, the storage medium may be a magnetic disk, an optical disk, a ROM, a RAM, and the like.

The above disclosures are only preferred embodiments of the present invention, and of course, the scope of the rights of the present invention cannot be limited by this. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims (5)

1. a deep learning MOCR method applied to DFB laser chip, is characterized in that, comprises the following steps: Step 1: Collect the image of the DFB laser chip online; Step 2: Preprocess the image; Step 3: Correct the direction of the characters on the chip through the character area detection network; Step 3.1: Perform downsampling on the input image four times in turn, sum the result of the fourth downsampling and the result of the third downsampling to obtain feature 1, and sum the result of feature 1 and the result of the second downsampling to obtain the feature Second, sum the result of feature two and the first downsampling to obtain feature three; Step 3.2: Perform feature fusion on the result of the fourth downsampling, feature 1, feature 2, and feature 3 to obtain fused features; Step 3.3: Perform convolution and deconvolution on the fusion feature to obtain a probability map and a threshold map; Step 3.4: Differentiable binarization is performed on the probability map and the threshold map to obtain an approximate binarized image; Step 3.5: Output the recognition frame according to the connected domain in the approximate binarized image, and convert the recognition frame to horizontal according to the length ratio of the recognition frame; Step 3.6: Determine whether the text is reversed through the text direction classifier; if the text direction is reversed, rotate the text 180 degrees, and then input the character recognition network for recognition; if not, directly input the character recognition network for recognition; Step 4: Recognize the character content on the chip through the character recognition network; Step 5: Determine whether the recognized character content is correct, if correct, store it in the associated database; if not, remove the chip or perform manual rechecking. 2 . The deep learning MOCR method applied to a DFB laser chip according to claim 1 , wherein the character area detection network is a DBNet deep learning network. 3 . 3. a kind of deep learning MOCR method that is applied to DFB laser chip according to claim 1, is characterized in that, described character recognition network is the full convolutional neural network CRNN based on character recognition, is used to recognize that character frame contains character of. 4. a kind of deep learning MOCR method applied to DFB laser chip according to claim 1, is characterized in that, in described step 1, the equipment that the image of DFB laser chip is collected online comprises light source, CMOS camera and telecentricity The lens adopts the automatic focusing technology based on infrared ranging to collect the surface character image of the DFB laser chip. 5. a kind of deep learning MOCR method applied to DFB laser chip according to claim 2, is characterized in that, the data set used during described DBNet deep learning network training is to manually mark and carry out data enhancement by original image formed later.

Images