Deep learning based high-dimensional and hybrid-dimensional orbital angular momentum identification and reconstruction method

Abstract

The application discloses a kind of high-dimensional and mixed dimension orbit angular momentum identification and reconstruction method based on deep learning, pre-acquires light intensity distribution image, for each light intensity image is marked with corresponding orbit angular momentum label, constructs data set, and is divided into training set and test set;Orbit angular momentum classification is defined as dimension classification problem, is divided into single category task, multi-category task and multi-label task;Single category task and multi-category task are trained using ResNet34 model;Multi-label task is encoded and decoded using deep learning model based on coding technology;Test set is input into the deep learning model that has been trained respectively, identifies the orbit angular momentum information of arbitrary high dimension and the sub-dimension analysis of the orbit angular momentum based on mixed dimension, respectively obtains the accuracy of its reconstruction process.The application avoids the dependence on complex hardware equipment, greatly reduces the cost and complexity of system, while improving the recognition speed and accuracy.

Description

Technical Field

[0001] This invention belongs to the field of optical communication, specifically relating to a method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning. Background Technology

[0002] Light, as a fundamental element of nature, plays a crucial role in numerous fields due to its unique properties. Its essence extends beyond the bright waves we see with the naked eye; more profoundly, photons, as particles of light, also possess inherent physical properties and behaviors. In traditional physics, any rotating object has angular momentum; similarly, photons also have their own angular momentum. This angular momentum can be divided into two main types: spin angular momentum (SAM) and orbital angular momentum (OAM). In short, spin angular momentum describes the inherent rotation of the photon, while orbital angular momentum relates to the overall shape and structure of the light. Beams with OAM, or vortex beams, exhibit great potential in optical communication due to their unique spiral shape. In mathematical representation, a vortex beam carrying OAM has a phase factor e. ilθ θ is the azimuth angle, and l is the topological charge number.

[0003] In modern optical communication, vortex beams have attracted widespread attention due to their high-dimensional data encoding capabilities. Different modes of OAM (Optical Aspect-Oriented Marking) can serve as unique information carriers, each corresponding to a specific data signal. This provides optical communication systems with a new and efficient signal encoding and transmission method. Traditional optical communication technologies are often limited by channel capacity, but the diversity of vortex beams and the unlimited mode potential of OAM open up new possibilities for the field. Compared to traditional optical communication methods, this OAM-based approach can significantly increase the information capacity of the channel, thereby providing higher data transmission rates. Furthermore, due to the structural characteristics of vortex beams, they exhibit better robustness against environmental interference during transmission.

[0004] In practical applications (such as information transmission), the precise generation, transmission, and detection of vortex beams with specific OAM modes are also crucial. To achieve efficient OAM encoding and decoding, researchers have developed various advanced techniques, such as spatial light modulation, sophisticated holographic design, and specific interferometry methods. In recent research, machine learning, particularly convolutional neural networks (CNNs), has been proposed for detecting and classifying beams with different OAM modes, thereby further optimizing the performance of optical communication systems. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to propose a method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning. This method analyzes and identifies orbital angular momentum in different dimensions and hybrid dimensions, ensuring accurate and rapid reconstruction of orbital angular momentum in both single and hybrid dimensions.

[0006] Technical solution: The present invention provides a method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning, comprising the following steps:

[0007] (1) Pre-acquire light intensity distribution images, label each light intensity image with its corresponding orbital angular momentum, construct a dataset, and divide it into a training set and a test set;

[0008] (2) The classification of orbital angular momentum is defined as a dimensional classification problem, which is divided into single-class tasks, multi-class tasks and multi-label tasks;

[0009] (3) Single-class and multi-class tasks are trained using the ResNet34 model; multi-label tasks are encoded and decoded using a deep learning model based on encoding techniques.

[0010] (4) Input the test set into the trained deep learning model respectively, identify orbital angular momentum information of arbitrary high dimension and sub-dimensional analysis based on mixed dimension orbital angular momentum, and obtain the accuracy of its reconstruction process respectively.

[0011] Further, the dataset in step (1) is divided into a training set and a test set; the training set is subjected to data augmentation processing by rotation, stretching and twisting and noise injection to simulate the real environment.

[0012] Furthermore, both the single-class task and the multi-class task described in step (2) are based on the judgment of orbital angular momentum in a single dimension. In the case of a single-class task, the training set of the model is only for a specific dimension, and only a specific dimension can be recognized during the recognition process. In contrast, the multi-class task uses the judgment of multiple dimensions in the model training, so different dimensional information can also be recognized during the recognition process. The multi-label task contains multiple dimensional information in one piece of information. By encoding each sub-dimensional information, and then decoding the sub-dimensional information in the mixed dimensional information, the effect of mixed dimensional recognition is achieved.

[0013] Further, the ResNet34 model in step (3) includes an input layer, multiple residual blocks, a global average pooling layer, and a fully connected output layer; the input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch size; the residual block consists of two 3x3 convolutional layers, an activation function "ReLU", and a fast path structure; to achieve feature extraction at different levels, the second layer conv2_x includes 3 residual blocks, each using 64 convolutional kernels, and the third layer conv3_x includes 4 residual blocks, each using 128 kernels. The convolutional kernels are as follows: the fourth layer (conv4_x) consists of 6 residual blocks, each using 256 convolutional kernels; the fifth layer (conv5_x) consists of 3 residual blocks, each using 512 convolutional kernels. At the beginning of each group of residual blocks, when the feature map size is halved, the number of convolutional kernels is increased, using 1x1 convolutions for matching. The global average pooling layer performs average pooling on each feature map, with an output size of 1x1. The fully connected output layer is a b×N tensor, where N represents the number of OAM state classes, and the probability of each class is output flexibly according to the actual problem.

[0014] Further, the deep learning model based on encoding technology in step (3) includes an input layer, an encoder, an intermediate layer, a decoder, a global average pooling layer, and a fully connected output layer; the encoder consists of multiple convolutional layers and pooling layers, which progressively extract and compress key information in the image; the intermediate layer performs deeper feature extraction on the feature map output by the encoder, ensuring the independence and representational power of sub-dimension features; the decoder consists of multiple upsampling layers and convolutional layers, which progressively decode the compressed information in the encoder and restore the original feature dimensions; the global average pooling layer reduces the size of the feature map output by the decoder, reducing the number of parameters and computational cost.

[0015] Furthermore, the encoder includes a first layer using 64 3x3 convolutional kernels and 2x2 pooling kernels, a second layer using 128 3x3 convolutional kernels and 2x2 pooling kernels, and a third layer using 256 3x3 convolutional kernels and 2x2 pooling kernels.

[0016] Furthermore, the decoder includes a first layer using 256 3x3 convolutional kernels and 2x2 upsampling kernels, a second layer using 128 3x3 convolutional kernels and 2x2 upsampling kernels, and a third layer using 64 3x3 convolutional kernels.

[0017] Furthermore, the intermediate layer contains a series of convolutional layers that further refine the feature representations generated by the encoder using smaller convolutional kernels.

[0018] Furthermore, the input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch training size.

[0019] Furthermore, the fully connected output layer is a b×M tensor, where M represents all possible combinations of mixed-dimensional orbital angular momentum.

[0020] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: The high-dimensional and mixed-dimensional orbital angular momentum reconstruction method based on deep learning provided by this invention integrates ResNet34 and encoder technology, ensuring accurate and rapid reconstruction of single-dimensional and mixed-dimensional orbital angular momentum; through the utilization of deep learning technology, this invention achieves automated and intelligent recognition of high-dimensional and mixed-dimensional orbital angular momentum, reducing reliance on professionals and improving work efficiency; the model structure proposed in this invention is optimized and has strong generalization ability, adapting to different application scenarios and data environments, reducing the need for repeated model adjustments in various scenarios; using encoder technology, this invention can effectively handle mixed-dimensional problems, not only recognizing the overall orbital angular momentum in an image, but also locating and recognizing each sub-dimension, providing strong support for applications in complex scenarios; compared with traditional orbital angular momentum recognition methods, this invention avoids reliance on complex hardware devices, greatly reducing system cost and complexity, while improving recognition speed and accuracy. Attached Figure Description

[0021] Figure 1 A schematic diagram of the structure of a ResNet34 network applied to a single dimension;

[0022] Figure 2 A schematic diagram of a deep learning network structure based on coding techniques applied to mixed-dimensional recognition;

[0023] Figure 3 Create a schematic diagram for the experimental dataset. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] This invention proposes a method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning. First, orbital angular momentum classification is defined as a dimensional classification problem. Second, deep learning techniques are used to divide the problem into three categories: one-class, multi-class, and multi-label tasks. Then, operations are performed on each of these three types of problems. Specifically, the steps include:

[0026] Step 1: Pre-acquire light intensity distribution images, label each light intensity image with its corresponding orbital angular momentum, construct a dataset, and divide it into training and test sets.

[0027] The simulation of the dimensional LG model was performed using PyCharm software, and the orbital angle quantum number phase information and the cylindrical lens modulation phase information were plotted in the image.

[0028] The collected light intensity distribution images were preprocessed, with each image labeled with its corresponding orbital angular momentum. Using these preprocessed images, 80% of the data was used as the training set, and the remaining 20% ​​as the validation set. The training set data underwent data augmentation processes such as rotation, stretching, and distortion, and was also injected with appropriate levels of noise to simulate real-world conditions. The test set was dynamically selected based on the model's performance on the actual training dataset to ensure accuracy and robustness in the testing.

[0029] Step 2: Define orbital angular momentum classification as a dimensional classification problem, which is divided into single-class tasks, multi-class tasks, and multi-label tasks.

[0030] The reconstruction of orbital angular momentum in high-dimensional and mixed-dimensional scenarios is categorized into three types: single-class tasks, multi-class tasks, and multi-label tasks. Both single-class and multi-class tasks are based on judging orbital angular momentum in a single dimension. The difference lies in the model's approach: in single-class tasks, the model's training set focuses on only one specific dimension, and the recognition process can only identify that specific dimension. In contrast, multi-class tasks utilize judgments across multiple dimensions during model training, thus recognizing different dimensional information. Multi-label tasks involve identifying information containing multiple dimensions. Each sub-dimensional information is encoded, and then the sub-dimensional information within the mixed-dimensional information is decoded to achieve mixed-dimensional recognition.

[0031] Step 3: Building a deep learning model: Single-class and multi-class tasks use a simple ResNet34 model, while multi-label tasks involve mixed-dimensional classification and employ encoder technology for encoding and decoding.

[0032] For single-class tasks (single-dimensional recognition), the ResNet34 model is used for training. The purpose of this task is to determine whether an input belongs to a specific class. ResNet (Residual Network) can solve the gradient vanishing and gradient exploding problems in deep neural networks. Its core idea introduces the concept of "residual learning". Each residual block learns a residual function and superimposes it with its input. ResNet34, as a version of ResNet, contains 34 layers. Its "bottleneck" structure helps reduce the amount of computation. Data is input into ResNet34 in random order and iteratively trained, up to 5 times, to obtain the optimized model.

[0033] For the single-dimensional recognition model ResNet34, the ResNet34 deep convolutional neural network structure constructed in this invention is as follows: Figure 1 As shown, the entire network mainly consists of an input layer, multiple residual blocks, a global average pooling layer, and a fully connected output layer. The input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch size. The residual block contains two 3x3 convolutional layers (Conv), an activation function (ReLU), and a shortcut structure. To achieve feature extraction at different depths, ResNet34 is configured as follows: the second layer conv2_x consists of 3 residual blocks, each using 64 convolutional kernels; the third layer conv3_x consists of 4 residual blocks, each using 128 convolutional kernels; the fourth layer conv4_x consists of 6 residual blocks, each using 256 convolutional kernels; and the fifth layer conv5_x consists of 3 residual blocks, each using 512 convolutional kernels. At the beginning of each residual block, when the feature map size is halved, feature matching is performed using 1x1 convolutions, while simultaneously increasing the number of convolutional kernels. Global Average Pooling layers are used to reduce the feature map size to 1x1, thereby reducing the number of parameters and computational cost. The fully connected output layer is a b×N tensor, where N represents the number of classes in the OAM state. This feature makes the model highly flexible, allowing for fine-tuning according to the specific needs of the problem and meeting the requirements of different class recognition.

[0034] For multi-class tasks (single-dimensional recognition), we still use the ResNet34 model for training. The goal is to determine which category the input belongs to. By learning the features more deeply, the model can better recognize different image features. After 20 runs of the multi-label task, the model can fully learn the features in the dataset and ensure generalization ability to avoid overfitting.

[0035] This invention first extracts image features using a ResNet34 deep convolutional neural network, fully utilizing the residual structure to avoid the problems of vanishing and exploding gradients, thereby achieving deep feature learning. Then, a global average pooling layer is used to reduce the number of model parameters and computational cost. Finally, a fully connected layer is used for classification, achieving accurate identification of single-dimensional orbital angular momentum.

[0036] For multi-label tasks (mixed-dimensional recognition), an input may be assigned to multiple labels. Autoencoders are chosen to address this issue. Autoencoders learn important features of the data through encoding and decoding processes, rather than performing explicit classification. In multi-label tasks, autoencoders are used to extract multiple features from an image, which can then be used for multi-label classification. Specifically, for mixed-dimensional classification tasks, feature extraction (encoding) of each sub-dimension is performed first, followed by decoding to obtain the recognition result for each sub-dimension.

[0037] For mixed-dimensional recognition, such as Figure 2 As shown, this invention employs a deep convolutional neural network based on the encoder-decoder paradigm, comprising an input layer, encoder, intermediate layers, decoder, global average pooling layer, and fully connected output layer.

[0038] The input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch size. The encoder structure is similar to the traditional ResNet34, but it has been specifically optimized for mixed-dimensional recognition.

[0039] The encoder consists of multiple convolutional and pooling layers, designed to progressively extract and compress key information from the image. Specifically, the first layer uses 64 3x3 convolutional kernels and 2x2 pooling kernels; the second layer uses 128 3x3 convolutional kernels and 2x2 pooling kernels; and the third layer uses 256 3x3 convolutional kernels and 2x2 pooling kernels. These layers aim to extract key features from the image and perform spatial compression, preparing for higher-level feature extraction. The encoder extracts features at different scales and levels of abstraction through multiple convolutional operations, performs deep learning on these sub-dimensional features, and encodes them into a series of feature maps.

[0040] Furthermore, a basic intermediate layer architecture is added after the encoder structure. This intermediate layer contains a series of convolutional layers that further refine the feature representations generated by the encoder using smaller convolutional kernels (e.g., 1x1 or 3x3). It performs deeper feature extraction on the feature map output by the encoder, ensuring the independence and representational power of sub-dimensional features.

[0041] The decoder structure is similar to the encoder structure, but the operations are reversed. The decoder consists of multiple upsampling and convolutional layers, aiming to progressively decode the compressed information from the encoder and recover the original feature dimensions. A specific configuration includes a first layer using 256 3x3 convolutional kernels and 2x2 upsampling kernels, a second layer using 128 3x3 convolutional kernels and 2x2 upsampling kernels, and a third layer using 64 3x3 convolutional kernels. The decoder receives feature maps from intermediate layers and uses a series of upsampling and deconvolutional operations to reconstruct and identify specific features of sub-dimensions. To achieve more accurate sub-dimension feature recognition, the decoder also fuses low-level features from the encoder.

[0042] Global average pooling is used to reduce the size of the feature map output by the decoder to 1x1, thereby reducing the number of parameters and computational cost.

[0043] Finally, the fully connected output layer is a b×M tensor, where M represents the number of categories in all possible sub-dimensions of the mixed-dimensional image, describing the probability that the input image belongs to that particular combination. The goal of this layer is to output a probability value for each combination, which allows the model to be fine-tuned according to the needs of the actual problem and adapt to the requirements of different sub-dimension recognition.

[0044] An encoder extracts and compresses key information from the input image, and a decoder then recovers the spatial dimensions of this information. This structure effectively learns features in multiple dimensions of the image and provides a corresponding description for each dimension. The fully connected output layer then outputs probability values ​​for each combination of mixed dimensions based on the decoded features, thereby achieving accurate identification of mixed-dimensional orbital angular momentum.

[0045] Step 4: Input the test set into the trained deep learning model to identify orbital angular momentum information of arbitrary high dimensions and sub-dimensional analysis based on mixed-dimensional orbital angular momentum, and obtain the accuracy of the reconstruction process for each.

[0046] like Figure 3 As shown in the simplified diagram of the experimental optical path system used in this invention, it mainly includes three processes: first, a laser emits a light source; then, the beam passes through a spatial light modulator (SLM) carrying orbital angular momentum (OAM) information; finally, a charge-coupled device (CCD) is used to take a picture, and the obtained image is used to create the dataset required for the model. The network input is... Figure 1The image dataset obtained after passing through a charge-coupled device (CCD) is preprocessed and enhanced before being fed into the corresponding deep models according to the problem task. The following describes in detail the actual performance of the OAM state recognition method based on a ResNet34 deep convolutional neural network, using two examples.

[0047] Example 1: Classification of OAM states in a single dimension.

[0048] In this embodiment, OAM state information with dimensions of 2, 3, 4, 5, 6, and 7 is used. Each dimension contains 500 image data points, with 80% used as the training set and 20% as the validation set. By using the convolutional neural network method described in this invention, the model achieves a loss value of 10 on the validation set. -3 The magnitude is on the order of magnitude. Furthermore, to further verify the model's generalization ability, another 100 images were selected from each OAM state as a test set. The results show that the model's classification accuracy on this test set is nearly 100%.

[0049] Example 2: OAM state classification for mixed dimensions.

[0050] This embodiment, based on OAM state information of dimensions 2, 3, 4, 5, 6, and 7, increases information capacity by pairwise mixing of the six dimensions to form new mixed-dimensional information. During model training, 4000 images were used for each sub-dimension, with 80% used for training and 20% for validation. In the testing phase, 1000 images were used for each new mixed dimension. During decoding, the model first decodes each sub-dimension to obtain its features, and then performs a decoding process to identify the OAM state. On this test set, the model achieved an average accuracy of 98.3%.

[0051] The above embodiments are merely illustrative of the present invention and are not intended to limit the scope of the present invention. All equivalent implementations of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning, characterized in that, Includes the following steps: (1) Acquire light intensity distribution images in advance, mark each light intensity image with the corresponding orbital angular momentum, construct a dataset, and divide it into training set and test set; (2) The classification of orbital angular momentum is defined as a dimensional classification problem, which is divided into single-class tasks, multi-class tasks and multi-label tasks; (3) Single-class and multi-class tasks are trained using the ResNet34 model; multi-label tasks are encoded and decoded using a deep learning model based on encoding techniques. (4) Input the test set into the trained deep learning model respectively, identify orbital angular momentum information of arbitrary high dimension and sub-dimensional analysis based on mixed dimension orbital angular momentum, and obtain the accuracy of its reconstruction process respectively. The deep learning model based on encoding technology in step (3) includes an input layer, encoder, intermediate layer, decoder, global average pooling layer and fully connected output layer connected in sequence; the encoder is composed of multiple convolutional layers and pooling layers, which gradually extract and compress key information in the image; The intermediate layer performs deeper feature extraction on the feature map output by the encoder to ensure the independence and representational power of sub-dimension features; the decoder, composed of multiple upsampling layers and convolutional layers, gradually decodes the feature map received from the intermediate layer to restore the original feature dimensions. The global average pooling layer reduces the size of the feature map output by the decoder, thereby reducing the number of parameters and computational cost.

2. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, In step (1), the dataset is divided into a training set and a test set. The training set is augmented by rotation, stretching and twisting, and noise is injected to simulate the real environment.

3. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, Step (2) describes both single-class and multi-class tasks based on the judgment of orbital angular momentum in a single dimension. In the case of single-class tasks, the training set of the model is only for a certain dimension, and only a certain dimension can be recognized during the recognition process. In contrast, multi-class tasks use judgments of multiple dimensions during model training, so different dimensional information can also be recognized during the recognition process. The multi-label task contains multiple dimensional information in one piece of information. By encoding each sub-dimensional information and then decoding the sub-dimensional information in the mixed dimensional information, the effect of mixed dimensional recognition is achieved.

4. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The ResNet34 model in step (3) includes an input layer, multiple residual blocks, a global average pooling layer, and a fully connected output layer; the input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch size; the residual block consists of two 3x3 convolutional layers, an activation function "ReLU", and a fast path structure; to achieve feature extraction at different levels, the second layer conv2_x includes 3 residual blocks, each using 64 convolutional kernels, and the third layer conv3_x includes 4 residual blocks, each using 128 convolutional kernels. The fourth layer, conv4_x, includes 6 residual blocks, each using 256 convolutional kernels. The fifth layer, conv5_x, includes 3 residual blocks, each using 512 convolutional kernels. At the beginning of each group of residual blocks, when the feature map size is halved, the number of convolutional kernels is increased, and 1x1 convolutions are used for matching. The global average pooling layer performs average pooling on each feature map, with an output size of 1x1. The fully connected output layer is a b×N tensor, where N represents the number of classes in the OAM state, and the probability of each class is output flexibly according to the actual problem.

5. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The encoder includes a first layer using 64 3x3 convolutional kernels and 2x2 pooling kernels, a second layer using 128 3x3 convolutional kernels and 2x2 pooling kernels, and a third layer using 256 3x3 convolutional kernels and 2x2 pooling kernels.

6. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The decoder includes a first layer using 256 3x3 convolutional kernels and 2x2 upsampling kernels, a second layer using 128 3x3 convolutional kernels and 2x2 upsampling kernels, and a third layer using 64 3x3 convolutional kernels.

7. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The intermediate layer contains a series of convolutional layers that further refine the feature representations generated by the encoder using 1x1 or 3x3 convolutional kernels.

8. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The input layer is a single-channel grayscale image of size b×224×224×1, where b is the batch training size.

9. The method for identifying and reconstructing high-dimensional and hybrid-dimensional orbital angular momentum based on deep learning according to claim 1, characterized in that, The fully connected output layer is a b×M tensor, where M represents the number of categories in all sub-dimensions of the mixed-dimensional image.

Images