Positioning method based on quantum dot optical encoder

Abstract

The invention discloses a positioning method based on a quantum dot optical encoder, and the method comprises the steps: constructing an HSV three-channel feature extraction network which comprises three feature extraction channels based on an improved U-Net structure, inputting a to-be-measured image into the three feature extraction channels based on the improved U-Net structure, and carrying out the detection of the to-be-measured image through the three feature extraction channels based on the improved U-Net structure, the method comprises the following steps of: obtaining an H-channel feature map, an S-channel feature map and a V-channel feature map correspondingly, performing channel fusion and feature extraction on the H-channel feature map, the S-channel feature map and the V-channel feature map to obtain a first feature map, and performing code channel positioning processing on the first feature map to obtain a positioning result; and performing feature extraction on the positioning result through the HSV three-channel feature extraction network to obtain a second feature map, and performing coding processing and displacement prediction on the second feature map to obtain a measurement result. According to the invention, the measurement precision of the quantum dot optical encoder can be improved.

Description

Technical Field

[0001] This invention relates to the field of high-precision measurement, and in particular to a positioning method based on a quantum dot optical encoder. Background Technology

[0002] Optical encoders, as key components in micro-nano ultra-precision manufacturing, are widely used in high-end precision instruments and equipment such as industrial micro-nano robots, mid-to-high-end CNC machine tools, and precision measuring instruments, determining the measurement and control accuracy of these instruments and equipment. Currently, optical encoders primarily employ "black and white" chrome film technology to create black and white code tracks on a base scale. For the black and white chrome film gratings, a "0-1" binary encoding mode is mainly used for encoding. The drawback is that when representing the same length of displacement, binary encoding requires a large number of code elements (0 or 1) (taking quaternary and binary encoding as examples, the information capacity of quaternary absolute position encoding is 1 / 2, while that of binary is 1 / 2). In actual products, typically 16 binary code elements are used to represent one absolute position code area to achieve encoding for measurement strokes exceeding 0.5 m. This binary encoding mechanism leads to excessively long encoding code areas, severely limiting the measurement stroke of the optical encoder and making it difficult to achieve high-precision measurement with large strokes.

[0003] Due to the limitations of optical encoders, researchers have developed quantum dot optical encoders by fabricating self-emissive quantum dot gratings using micro-nano additive manufacturing technology. Quantum dot optical encoders achieve high-precision measurement of color gamut tracks (i.e., multi-color tracks made from a material called quantum dots) by introducing multi-color gamut code tracks. This recompiling of the absolute position code in quantum dot optical encoders replaces the black-and-white tracks in traditional optical encoders with multi-color gamut code tracks, resulting in larger encoding / decoding capacity, more easily distinguishable absolute position codes, and more accurate absolute position measurement precision. This improves decoding accuracy and measurement stability, better meeting the measurement needs of current high-end manufacturing industries.

[0004] The measurement accuracy of quantum dot optical encoders is highly dependent on the code track positioning results. However, due to the inherent multi-color gamut mixing phenomenon (two colors mixing to generate a new color gamut) of quantum dot optical encoders, their multi-color gamut code tracks will produce nonlinear gradient edges or even nonlinear optical distortion. This problem makes it impossible to accurately detect the edges of the code tracks when locating code tracks in multi-color gamut images, resulting in significant positioning errors. Furthermore, in actual measurement processes, quantum dot optical encoders are also affected by various internal and external factors, including the type and ratio of quantum dots, light intensity, ambient temperature, and vibration. These internal and external factors can affect image quality to some extent and also introduce code track positioning errors, seriously impacting the measurement accuracy and reliability of quantum dot optical encoders. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, this invention provides a positioning method based on a quantum dot optical encoder, which can improve the measurement accuracy of the quantum dot optical encoder.

[0006] One embodiment of the present invention provides a positioning method based on a quantum dot optical encoder, comprising the following steps: An HSV three-channel feature extraction network is constructed, comprising three feature extraction channels based on an improved U-Net structure. The image to be measured is input into the three feature extraction channels based on the improved U-Net structure, respectively, to obtain H-channel feature maps, S-channel feature maps, and V-channel feature maps. Channel fusion and feature extraction are performed on the H-channel feature maps, S-channel feature maps, and V-channel feature maps to obtain a first feature map. Code channel localization processing is then performed on the first feature map to obtain the localization result. The positioning result is subjected to feature extraction by the HSV three-channel feature extraction network to obtain a second feature map. The measurement result is obtained by encoding the second feature map and performing displacement prediction.

[0007] Furthermore, the step of performing channel fusion and feature extraction on the H-channel feature map, S-channel feature map, and V-channel feature map to obtain the first feature map specifically includes: The H-channel feature map is fused with the S-channel feature map and the V-channel feature map respectively, and the two fused results are convolved through a preset 3×3 convolutional layer to obtain the convolution result; The first feature map is obtained by performing feature extraction processing on the convolution result using a preset Euler module; wherein the preset Euler module includes a wavelet convolution layer, a 1×1 convolution layer and a phase extraction layer.

[0008] Preferably, the construction step of the feature extraction channel based on the improved U-Net structure specifically includes: A wavelet convolutional layer and a phase extraction layer are added before the input of the original U-Net structure. The output of the wavelet convolutional layer is connected to the input of the phase extraction layer, and the output of the phase extraction layer is connected to the input of the original U-Net structure. A depthwise separable convolutional layer is added to the downsampling layer of the original U-Net structure, and the Euler module is added to the connection path of all upsampling and downsampling layers.

[0009] Furthermore, the step of obtaining the positioning result by performing code channel positioning processing on the first feature map specifically includes: The first feature map is input into the feature extraction channel based on the improved U-Net structure, and the result output by the feature extraction channel is processed by a preset code channel localization model to obtain the localization result.

[0010] Furthermore, the step of extracting features from the localization result using the HSV three-channel feature extraction network to obtain a second feature map specifically includes: The localization results are input into the three feature extraction channels based on the improved U-Net structure, and after channel fusion and convolution processing of the feature extraction results, feature extraction is performed through the preset Euler module to obtain the second feature map.

[0011] Furthermore, the step of obtaining the measurement result based on the second feature map through encoding processing and displacement prediction specifically includes: The second feature map is input into a preset encoding processing model and a preset displacement prediction model respectively to obtain the encoding result and displacement prediction value; wherein, the preset encoding processing model is composed of GRU and Swin Transformer encoding layers, and the preset displacement prediction model is composed of an MLP network; The measurement result is calculated by looking up the table based on the encoding result and the displacement prediction value.

[0012] Preferably, after inputting the first feature map into the feature extraction channel based on the improved U-Net structure, the method further includes: The edge detection result is obtained by performing edge detection processing on the output of the feature extraction channel using a preset edge detection model; Based on the positioning results and the edge detection results, the loss function of the preset code track positioning model and the loss function of the edge detection model are constructed respectively. Based on the loss function of the preset code channel positioning model and the loss function of the preset edge detection model, a weighted positioning loss function is constructed through weighted calculation, and the preset code channel positioning model is optimized through the weighted positioning loss function.

[0013] Preferably, after obtaining the measurement results, the method further includes: Based on the encoding results and displacement prediction values, the loss function of the preset encoding processing model and the loss function of the displacement prediction model are constructed respectively. Based on the loss function of the preset encoding processing model and the loss function of the preset displacement prediction model, a weighted measurement loss function is constructed through weighted calculation, and the preset encoding processing model and the preset displacement prediction model are jointly optimized through the weighted measurement loss function.

[0014] Further preferably, after obtaining the weighted positioning loss function and the weighted measurement loss function, the method further includes: Based on the measurement results, a KL divergence loss function is constructed; Based on the KL divergence loss function, the weighted positioning loss function, and the weighted measurement loss function, a weighted optimization loss function is constructed through weighted calculation. The preset code track positioning model, the preset coding processing model, and the preset displacement prediction model are then jointly optimized using the weighted optimization loss function.

[0015] Preferably, constructing the KL divergence loss function based on the measurement results specifically includes: The image to be measured is measured using a preset laser interferometer to obtain the laser measurement result; Based on the measurement results and the laser measurement results, the KL divergence loss function is constructed; wherein, the specific formula of the KL divergence loss function is:

[0016] in, Let KL divergence loss function be the KL divergence loss function. and These are the probability density functions of the measured value and the laser measurement result, respectively.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: By employing the HSV three-channel feature extraction network to process the image under test, cross-color gamut optical crosstalk can be suppressed, improving the accuracy of code channel positioning and encoding / decoding, thereby enhancing the measurement accuracy of the quantum dot optical encoder. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a positioning method based on a quantum dot optical encoder, provided as an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of a first feature map extraction process provided in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of the structure of a feature extraction channel provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of a positioning result acquisition process provided in an embodiment of the present invention.

[0022] Figure 5 This is a schematic diagram illustrating the acquisition of measurement results according to an embodiment of the present invention.

[0023] Figure 6 This is a schematic diagram of the workflow of a sub-pixel level end-to-end measurement model provided in an embodiment of the present invention. Detailed Implementation

[0024] The accompanying drawings are for illustrative purposes only and should not be construed as limiting the scope of this patent. It will be understood by those skilled in the art that certain well-known structures and their descriptions may be omitted in the accompanying drawings.

[0025] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0026] Reference Figure 1 The following is a flowchart illustrating a positioning method based on a quantum dot optical encoder according to an embodiment of the present invention, comprising the following steps: S1: Construct an HSV three-channel feature extraction network. The HSV three-channel feature extraction network includes three feature extraction channels based on an improved U-Net structure. The image to be measured is input into the three feature extraction channels based on the improved U-Net structure, respectively, to obtain H-channel feature maps, S-channel feature maps, and V-channel feature maps. Channel fusion and feature extraction are performed on the H-channel feature maps, S-channel feature maps, and V-channel feature maps to obtain a first feature map. Code channel localization processing is performed on the first feature map to obtain the localization result. S2: The positioning result is subjected to feature extraction through the HSV three-channel feature extraction network to obtain a second feature map, and the measurement result is obtained by encoding the second feature map and performing displacement prediction.

[0027] For step S1, specifically, the process of channel fusion and feature extraction on the H-channel feature map, S-channel feature map, and V-channel feature map to obtain the first feature map includes: The H-channel feature map is fused with the S-channel feature map and the V-channel feature map respectively, and the two fused results are convolved through a preset 3×3 convolutional layer to obtain the convolution result; The first feature map is obtained by performing feature extraction processing on the convolution result using a preset Euler module; wherein the preset Euler module includes a wavelet convolution layer, a 1×1 convolution layer and a phase extraction layer.

[0028] In a preferred embodiment, refer to Figure 2 This is a schematic diagram of a first feature map extraction process provided in an embodiment of the present invention.

[0029] Depend on Figure 2As can be seen, after the image to be measured is processed through the three feature extraction channels (H, S, and V) and corresponding feature maps are generated, it needs to be further fused using a channel fusion layer. Specifically, the features output from the H and S channels are fused, and the features output from the H and V channels are also fused. In this way, high-saturation features are dynamically enhanced in the transition region while effectively suppressing low-saturation noise, thereby reducing the impact of color mixing. The fused result is processed through a 3×3 convolutional layer, and finally, the feature extraction is further refined through an Euler module to obtain the first feature map.

[0030] In the above extraction process, this preferred embodiment first employs H / S / V three-channel feature extraction technology to achieve feature decoupling, thereby suppressing cross-color gamut optical crosstalk and ensuring the accuracy of code channel positioning. Secondly, an Euler feature extraction module is introduced to enhance the network's ability to extract horizontal and vertical bar and channel features, further improving positioning accuracy. Simultaneously, wavelet convolution is added to improve the model's sensitivity to frequency domain features, effectively suppressing image noise interference, thus further enhancing positioning accuracy.

[0031] Preferably, the construction step of the feature extraction channel based on the improved U-Net structure specifically includes: A wavelet convolutional layer and a phase extraction layer are added before the input of the original U-Net structure. The output of the wavelet convolutional layer is connected to the input of the phase extraction layer, and the output of the phase extraction layer is connected to the input of the original U-Net structure. A depthwise separable convolutional layer is added to the downsampling layer of the original U-Net structure, and the Euler module is added to the connection path of all upsampling and downsampling layers.

[0032] In a preferred embodiment, refer to Figure 3 This is a schematic diagram of the structure of a feature extraction channel provided in an embodiment of the present invention.

[0033] Depend on Figure 3 It is known that all feature extraction channels are based on an improved U-Net structure, with a wavelet convolutional layer and a phase extraction layer added before the input of the original U-Net structure. The wavelet convolutional layer combines wavelet transform with a convolutional neural network, effectively suppressing the interference of image noise on the feature extraction results, thereby improving measurement accuracy. The wavelet convolution operation does not affect the number of channels in the input image. Meanwhile, the phase extraction layer, constructed based on Euler's formula, can extract bar (horizontal and vertical) and channel features, clearly capturing the boundary information of multi-color gamut images, making it very suitable for feature extraction in multi-color gamut mixed regions. In addition, the feature fusion layer can integrate features from different directions, thereby providing a more comprehensive understanding of the overall structure of multi-color gamut images.

[0034] To improve the inference speed of the model, this preferred embodiment also introduces depthwise separable convolutions into the downsampling layers of the original U-Net to reduce model parameters and enhance computational efficiency. Simultaneously, an Euler module containing wavelet convolution layers, 1×1 convolution layers, and a phase extraction layer is added to the skip connections between the upsampling and downsampling layers, collectively improving the ability to extract fine features from multi-color-gamut images.

[0035] Therefore, the improved U-Net can effectively decouple the multi-color domain image features of the quantum dot optical encoder, avoiding the limitations of single-channel feature extraction networks when processing color mixing regions.

[0036] Furthermore, the step of obtaining the positioning result by performing code channel positioning processing on the first feature map specifically includes: The first feature map is input into the feature extraction channel based on the improved U-Net structure, and the result output by the feature extraction channel is processed by a preset code channel localization model to obtain the localization result.

[0037] Preferably, after inputting the first feature map into the feature extraction channel based on the improved U-Net structure, the method further includes: The edge detection result is obtained by performing edge detection processing on the output of the feature extraction channel using a preset edge detection model; Based on the positioning results and the edge detection results, the loss function of the preset code track positioning model and the loss function of the edge detection model are constructed respectively. Based on the loss function of the preset code channel positioning model and the loss function of the preset edge detection model, a weighted positioning loss function is constructed through weighted calculation, and the preset code channel positioning model is optimized through the weighted positioning loss function.

[0038] In a preferred embodiment, refer to Figure 4 This is a schematic diagram of a positioning result acquisition process provided in an embodiment of the present invention.

[0039] Depend on Figure 4 It can be seen that the first feature map, after being processed by the feature extraction channels based on the improved U-Net structure, is input to two different branches: the first branch is responsible for generating the code channel localization result, and the second branch outputs the edge detection result. In this preferred embodiment, the size of the input image is (1024, 1024, 3). Through the H / S / V three-channel U-Net, the two output sizes obtained are a localization result of (1024, 1024, 4) and an edge detection result of (1024, 1024, 1), where 4 represents the four colors (red, green, blue, and black) contained in the original image.

[0040] After obtaining the positioning result and edge detection result, this preferred embodiment further constructs a weighted positioning loss function based on the positioning result and edge detection result. The specific loss function is as follows:

[0041]

[0042]

[0043] in, All are preset weighted weights. and These are the loss functions of the preset code track localization model and the preset edge detection model, respectively. It is the sample size. It is the number of categories. These are real labels, if the sample Category The value is 1 if it is 1, otherwise it is 0. It is the probability predicted by the model, representing the sample Category The probability of. It is an edge intensity map of the real image. It is the predicted edge intensity map.

[0044] Training the preset code track localization model using the weighted localization loss function can improve the accuracy of multi-color-domain code track edge detection and further enhance the accuracy of code track localization.

[0045] For step S2, specifically, the step of extracting features from the localization result using the HSV three-channel feature extraction network to obtain a second feature map specifically includes: The localization results are input into the three feature extraction channels based on the improved U-Net structure, and after channel fusion and convolution processing of the feature extraction results, feature extraction is performed through the preset Euler module to obtain the second feature map.

[0046] Furthermore, the step of obtaining the measurement result based on the second feature map through encoding processing and displacement prediction specifically includes: The second feature map is input into a preset encoding processing model and a preset displacement prediction model respectively to obtain the encoding result and displacement prediction value; wherein, the preset encoding processing model is composed of GRU and Swin Transformer encoding layers, and the preset displacement prediction model is composed of an MLP network; The measurement result is calculated by looking up the table based on the encoding result and the displacement prediction value.

[0047] In a preferred embodiment, refer to Figure 5 This is a schematic diagram illustrating the acquisition of measurement results according to an embodiment of the present invention.

[0048] Both encoding and decoding results have a combined impact on measurement accuracy, but previous localization methods often neglected encoding errors. Therefore, during the encoding process, H / S / V three-channel feature extraction technology is employed, and wavelet convolution and Euler feature extraction modules are introduced to enhance the network's ability to extract key information features.

[0049] Subsequently, by Figure 5 It can be seen that after the positioning result is processed by the preset HSV three-channel feature extraction network, the second feature map can be obtained. Subsequently, the second feature map is input into the two branches respectively.

[0050] The first branch is the preset encoding processing model, which consists of a gated recurrent unit (GRU) and a Swing Transformer encoding layer. The lightweight GRU captures contextual information from the second feature map, and the Swing Transformer encoding layer ultimately generates a highly robust encoding result. Specifically, the size of the localization result is (1024, 1024, 4). After feature extraction, the localization result generates a second feature map. The second feature map is transformed into a one-dimensional vector of length n by the encoding processing model, where n is the number of codes. This one-dimensional vector is the encoding result.

[0051] The other branch is the preset displacement prediction model, which is composed of an MLP network. After the second feature map is input, the displacement prediction value can be output.

[0052] Finally, by looking up the obtained encoding results in a table and adding the displacement value, the final absolute position value of the quantum dot optical encoder can be calculated, which is the measurement result.

[0053] Preferably, after obtaining the measurement results, the method further includes: Based on the encoding results and displacement prediction values, the loss function of the preset encoding processing model and the loss function of the displacement prediction model are constructed respectively. Based on the loss function of the preset encoding processing model and the loss function of the preset displacement prediction model, a weighted measurement loss function is constructed through weighted calculation, and the preset encoding processing model and the preset displacement prediction model are jointly optimized through the weighted measurement loss function.

[0054] In a preferred embodiment, the weighted measurement loss function can be constructed based on the encoding result and the displacement prediction value. The specific formula for ) is:

[0055]

[0056]

[0057] in, All are preset weighted weights. and These are the loss functions of the preset encoding processing model and the displacement prediction model, respectively. It is the sample size. It is the number of codes. These are the actual encoded labels corresponding to multi-color gamut images. Then predict the encoding. These are the actual displacement values ​​corresponding to the multi-color gamut image. This is the predicted displacement value.

[0058] By optimizing the loss function of the preset encoding processing model and the displacement prediction model using the weighted measurement loss function, the accuracy of the encoding results and displacement prediction values ​​can be improved.

[0059] Further preferably, after obtaining the weighted positioning loss function and the weighted measurement loss function, the method further includes: Based on the measurement results, a KL divergence loss function is constructed; Based on the KL divergence loss function, the weighted positioning loss function, and the weighted measurement loss function, a weighted optimization loss function is constructed through weighted calculation. The preset code track positioning model, the preset coding processing model, and the preset displacement prediction model are then jointly optimized using the weighted optimization loss function.

[0060] Preferably, constructing the KL divergence loss function based on the measurement results specifically includes: The image to be measured is measured using a preset laser interferometer to obtain the laser measurement result; Based on the measurement results and the laser measurement results, the KL divergence loss function is constructed; wherein, the specific formula of the KL divergence loss function is:

[0061] in, Let KL divergence loss function be the KL divergence loss function. and The probability density functions of the measured value and the laser measurement result are respectively. In a preferred embodiment, refer to Figure 6 This is a schematic diagram of the workflow of a sub-pixel level end-to-end measurement model provided in an embodiment of the present invention.

[0062] In the actual measurement process of quantum dot optical encoders, dividing the measurement process into multiple steps leads to complexity and inconvenience, and each step introduces different errors, affecting measurement accuracy. Therefore, an end-to-end measurement model is constructed, consisting of a sub-pixel-level localization model and a sub-pixel-level encoding / decoding model. The sub-pixel-level localization model comprises a preset HSV three-channel feature extraction network, a feature extraction channel based on an improved U-Net structure, a preset code track localization model, and a preset edge detection model, with its working principle shown in S1. The sub-pixel-level encoding / decoding model consists of a sub-pixel-level encoding model and a decoder. The sub-pixel-level encoding model comprises the preset HSV three-channel feature extraction network, a preset encoding processing model, and a preset displacement prediction model, with its working principle shown in S2.

[0063] By weighted and integrated the loss functions of the two models, joint training of the two models can be achieved. That is, the loss function of the sub-pixel-level end-to-end measurement model ( The loss function is obtained by combining the sub-pixel level localization model and the sub-pixel level encoding / decoding model, that is: .

[0064] In this joint training process, this preferred embodiment also introduces high-precision ground truth data from a laser interferometer as a calibration benchmark based on the measured data. This is because although the measured values ​​obtained through the end-to-end measurement model are already relatively accurate, errors still exist, thus requiring the introduction of the KL divergence loss function.

[0065] By calculating the probability density functions of the model measurements, tag measurements, and laser interferometer measurements, and constructing a KL divergence loss function to optimize the measurement model, the measurement accuracy can be further improved. The KL divergence calculation is shown below.

[0066]

[0067]

[0068] As shown in the formula above, This represents the probability density function, while This is the KL divergence, where and These represent the probability density functions of the model measurements and the laser interferometer measurements, respectively.

[0069] However, since both measurements are discrete, optimizing the measurement network may result in insufficient generalization ability, necessitating the conversion of the probability density function of the measurements to continuous. Due to the stability of laser interferometer measurements, a Gaussian kernel function can be used... and smoothing parameters A relatively good continuous probability density function can already be obtained. However, the measured values ​​of the model are unstable, so the smoothing methods mentioned above are far from sufficient. Therefore, Bayesian inference is used to incorporate prior knowledge, and the posterior distribution is updated using the prior distribution and observed data to obtain a more robust density estimate. The specific formula is as follows.

[0070]

[0071]

[0072]

[0073] As shown above, The likelihood function represents the model's measured values. This means taking the probability distribution of the label measurements as prior knowledge, obtaining the posterior distribution through Bayesian inference, and finally obtaining the KL divergence. .

[0074] The above formula can further reduce measurement errors. Integrating all loss function results, the final loss function is as follows:

[0075] in, The preset weighting is used.

[0076] The measurement model is optimized using the aforementioned loss function to obtain the final measurement model parameters. During the model's inference phase, based on the jointly trained final measurement model parameters, the multi-color gamut image is input into the model to obtain the absolute position values ​​corresponding to the multi-color gamut image. Through this process, an end-to-end measurement workflow between the multi-color gamut image and the measurement data can be completed, ultimately yielding accurate measurement results.

[0077] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A positioning method based on a quantum dot optical encoder, characterized in that, Includes the following steps: An HSV three-channel feature extraction network is constructed, comprising three feature extraction channels based on an improved U-Net structure. The image to be measured is input into the three feature extraction channels based on the improved U-Net structure, respectively, to obtain H-channel feature maps, S-channel feature maps, and V-channel feature maps. Channel fusion and feature extraction are performed on the H-channel feature maps, S-channel feature maps, and V-channel feature maps to obtain a first feature map. Code channel localization processing is then performed on the first feature map to obtain the localization result. The positioning result is subjected to feature extraction by the HSV three-channel feature extraction network to obtain a second feature map. The measurement result is obtained by encoding the second feature map and performing displacement prediction.

2. The positioning method based on a quantum dot optical encoder as described in claim 1, characterized in that, The step of performing channel fusion and feature extraction on the H-channel feature map, S-channel feature map, and V-channel feature map to obtain the first feature map specifically includes: The H-channel feature map is fused with the S-channel feature map and the V-channel feature map respectively, and the two fused results are convolved through a preset 3×3 convolutional layer to obtain the convolution result; The first feature map is obtained by performing feature extraction processing on the convolution result using a preset Euler module; wherein the preset Euler module includes a wavelet convolution layer, a 1×1 convolution layer and a phase extraction layer.

3. The positioning method based on a quantum dot optical encoder as described in claim 2, characterized in that, The specific steps for constructing the feature extraction channels based on the improved U-Net structure include: A wavelet convolutional layer and a phase extraction layer are added before the input of the original U-Net structure. The output of the wavelet convolutional layer is connected to the input of the phase extraction layer, and the output of the phase extraction layer is connected to the input of the original U-Net structure. A depthwise separable convolutional layer is added to the downsampling layer of the original U-Net structure, and the Euler module is added to the connection path of all upsampling and downsampling layers.

4. The positioning method based on a quantum dot optical encoder as described in claim 2, characterized in that, The step of performing code channel localization processing on the first feature map to obtain the localization result specifically includes: The first feature map is input into the feature extraction channel based on the improved U-Net structure, and the result output by the feature extraction channel is processed by a preset code channel localization model to obtain the localization result.

5. The positioning method based on a quantum dot optical encoder as described in claim 4, characterized in that, The step of extracting features from the localization result using the HSV three-channel feature extraction network to obtain a second feature map specifically includes: The localization results are input into the three feature extraction channels based on the improved U-Net structure, and after channel fusion and convolution processing of the feature extraction results, feature extraction is performed through the preset Euler module to obtain the second feature map.

6. The positioning method based on a quantum dot optical encoder as described in claim 4, characterized in that, The step of obtaining the measurement result based on the second feature map through encoding processing and displacement prediction specifically includes: The second feature map is input into a preset encoding processing model and a preset displacement prediction model respectively to obtain the encoding result and displacement prediction value; wherein, the preset encoding processing model is composed of GRU and Swin Transformer encoding layers, and the preset displacement prediction model is composed of an MLP network; The measurement result is calculated by looking up the table based on the encoding result and the displacement prediction value.

7. The positioning method based on a quantum dot optical encoder as described in claim 6, characterized in that, After inputting the first feature map into the feature extraction channel based on the improved U-Net structure, the method further includes: The edge detection result is obtained by performing edge detection processing on the output of the feature extraction channel using a preset edge detection model; Based on the positioning results and the edge detection results, the loss function of the preset code track positioning model and the loss function of the edge detection model are constructed respectively. Based on the loss function of the preset code channel positioning model and the loss function of the preset edge detection model, a weighted positioning loss function is constructed through weighted calculation, and the preset code channel positioning model is optimized through the weighted positioning loss function.

8. The positioning method based on a quantum dot optical encoder as described in claim 7, characterized in that, After obtaining the measurement results, the following is also included: Based on the encoding results and displacement prediction values, the loss function of the preset encoding processing model and the loss function of the displacement prediction model are constructed respectively. Based on the loss function of the preset encoding processing model and the loss function of the preset displacement prediction model, a weighted measurement loss function is constructed through weighted calculation, and the preset encoding processing model and the preset displacement prediction model are jointly optimized through the weighted measurement loss function.

9. The positioning method based on a quantum dot optical encoder as described in claim 8, characterized in that, After obtaining the weighted localization loss function and the weighted measurement loss function, the method further includes: Based on the measurement results, a KL divergence loss function is constructed; Based on the KL divergence loss function, the weighted positioning loss function, and the weighted measurement loss function, a weighted optimization loss function is constructed through weighted calculation. The preset code track positioning model, the preset coding processing model, and the preset displacement prediction model are then jointly optimized using the weighted optimization loss function.

10. The positioning method based on a quantum dot optical encoder as described in claim 9, characterized in that, The step of constructing the KL divergence loss function based on the measurement results specifically includes: The image to be measured is measured using a preset laser interferometer to obtain the laser measurement result; Based on the measurement results and the laser measurement results, the KL divergence loss function is constructed; wherein, the specific formula of the KL divergence loss function is: in, Let KL divergence loss function be the KL divergence loss function. and These are the probability density functions of the measured value and the laser measurement result, respectively.

Images