A method for controlling column chromatogram image analysis based on BF-EdgeNet

Abstract

The application discloses a column chromatogram image analysis control method based on BF-EdgeNet and relates to the technical fields of computer vision and automation control, and comprises the following steps: acquiring a column chromatogram image and performing pretreatment; performing color band boundary data extraction on the pretreated column chromatogram image based on a preset BF-EdgeNet extraction model; performing structured processing on the extracted color band boundary data to generate boundary feature data, and performing color band segment state judgment according to the boundary feature data; generating corresponding control instructions according to the color band segment state judgment result, and realizing the switching of multiple containers and the collection of chemical components; wherein the extraction model comprises a BN main feature extraction network, an FDA feature fusion module and an ERD boundary decoder which are connected in sequence; and the application can realize high-precision extraction of upper and lower boundary curves of color bands in a column chromatogram image and automatic control of an execution mechanism to complete the switching of multiple receiving containers and the corresponding collection of components.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and automated control technology, and in particular to a column chromatography image analysis and control method based on BF-EdgeNet. Background Technology

[0002] Column chromatography is widely used in organic synthesis, drug separation and purification, etc., but traditional operation requires the experimenter to continuously observe the progress of the chromatographic bands to determine whether to start collecting the components, and manually change the receiving container based on visual judgment. This method is not only labor-intensive, but also prone to missing the replacement opportunity when the boundaries between the bands are unclear, resulting in component mixing and reduced purity. The accuracy of band boundary identification directly affects the purity of the components, but it is limited by human experience, fuzzy diffusion and other factors, making identification susceptible to interference and lacking stability. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of the prior art and provide a column chromatography image analysis control method based on BF-EdgeNet, which can achieve high-precision extraction of the upper and lower boundary curves of the color bands in the column chromatography image and automatic control of the mechanical actuator to complete the switching of multiple receiving containers and the corresponding collection of components.

[0004] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0005] This invention provides a column chromatography image analysis control method based on BF-EdgeNet, comprising:

[0006] Acquire column chromatographic images and preprocess the column chromatographic images;

[0007] Based on the preset BF-EdgeNet extraction model, the color band boundary data of the preprocessed column chromatographic image is extracted;

[0008] The extracted color band boundary data is processed in a structured manner to generate boundary feature data, and the color band segment status is determined based on the boundary feature data.

[0009] Based on the determination result of the color band segment status, corresponding control commands are generated to realize the switching of multiple containers and the collection of chemical components;

[0010] The BF-EdgeNet extraction model includes a BN backbone feature extraction network, an FDA feature fusion module, and an ERD boundary decoder connected in sequence.

[0011] The BN backbone feature extraction network includes a direction-aware initiating convolutional module and multiple layers of feature extraction modules connected sequentially to the output of the initiating convolutional module.

[0012] The FDA feature fusion module is used to fuse feature maps extracted by the feature extraction module at multiple levels to generate a fused feature map;

[0013] The ERD boundary decoder is used to receive the fused features generated by the FDA feature fusion module and extract the upper and lower boundary curves of the color band.

[0014] Optionally, the preprocessing of the column chromatogram image includes:

[0015] Image cropping, retaining only the regions containing color bands in the column chromatography image;

[0016] The cropped column chromatogram image was subjected to noise suppression and smoothing using a Gaussian blur algorithm.

[0017] The column chromatogram image after noise suppression and smoothing is normalized, and the pixel value of each pixel is mapped to the interval [0,1].

[0018] Optionally, the initial convolution module performs 7×1 and 1×7 convolution operations on the input image to extract horizontal and vertical features:

[0019]

[0020]

[0021] In the formula, From the input image Extracted horizontal and vertical features, for Activation function The convolution operations are 7×1 and 1×7 respectively;

[0022] lateral features and longitudinal features Perform channel concatenation, and then fuse the channels using a 3×3 convolution operation:

[0023]

[0024] In the formula, The output features of the initial convolutional module, This is a 3×3 convolution operation. This refers to the splicing operation in the channel dimension.

[0025] Optionally, each layer of the feature extraction module includes a direction-aware convolution module, a channel-spatial joint attention module, and a downsampling module connected in sequence;

[0026] The orientation-aware convolution module performs 5×1 and 1×5 convolution operations on the input features respectively, and extracts orientation information in parallel:

[0027]

[0028]

[0029] in, For the first Each input feature Convolution operations of 5×1 and 1×5 are respectively. For batch normalization operations, for Activation function From input features Extracted horizontal and vertical features;

[0030] lateral features and longitudinal features Perform channel splicing:

[0031]

[0032] In the formula, The output features of the orientation-aware convolutional module. This refers to the splicing operation at the channel dimension. The number of output channels for convolution in each direction. The height and width of the original input image;

[0033] The channel-space joint attention module provides the output features of the direction-aware convolution module. Perform global average pooling and connect it to a fully connected network. The activation function generates channel-dimensional weight coefficients. ; Output features of the direction-aware convolutional module Compression is performed along the channel dimension, and a fully connected network with a sigmoid activation function is used to obtain the response weights along the spatial dimension. Output features of the channel-space joint attention module for:

[0034]

[0035] In the formula, For pixel coordinates, This is the channel number index value;

[0036] The downsampling module employs a 3×3 convolution with stride=2 downsampling strategy, which halves the spatial resolution of the feature map at each level while increasing the number of channels layer by layer.

[0037]

[0038] In the formula, The output features of the downsampling module, This is a downsampling operation with a 3×3 convolution and a stride of 2. This represents the final output feature of the channel-space joint attention module.

[0039] Optionally, the step of fusing the feature maps extracted by the feature extraction modules at multiple levels to generate a fused feature map includes:

[0040] The feature maps extracted from each level are uniformly processed using 1×1 convolutions to determine the number of channels. :

[0041]

[0042] In the formula, For the first Feature maps extracted at each level This is a 1×1 convolution operation. The feature map after a 1×1 convolution operation ;

[0043] feature map Size upsampling to To achieve spatial dimensional alignment with the same size:

[0044]

[0045] In the formula, This is the feature map extracted from the first layer after a 1×1 convolution operation. For upsampling operation, For size, Feature map after upsampling operation ;

[0046] Feature maps with aligned spatial dimensions and uniform channel count are stitched together along the channel dimension:

[0047]

[0048] In the formula, This refers to the splicing operation at the channel dimension. For the number of levels, To stitch together the feature maps, The height and width of the original input image;

[0049] splicing feature maps Input a deformable attention fusion module, learn attention weights for adaptive spatial shift, and dynamically weight the fusion result:

[0050]

[0051] In the formula, For the final fused feature map, This is a deformable attention fusion operation.

[0052] Optionally, the upper and lower boundary curves of the extracted color band include:

[0053] The fused feature map output by the FDA feature fusion module Perform two convolutional layers to extract salient region features of the color band area:

[0054]

[0055]

[0056] In the formula, For batch normalization operations, for Activation function This is a 3×3 convolution operation; These are the feature maps output from two convolutional layers, respectively.

[0057] Use a 1×1 convolution to transform the feature map Mapped to a single-channel mask image:

[0058]

[0059] In the formula, Use the Sigmoid activation function to make the mask value fall within the interval [0,1]. pixel coordinates The probability that a pixel belongs to a color band region;

[0060] Based on a single-channel mask image, the upper and lower boundary curves of the color band are extracted using a column-by-column scanning method:

[0061]

[0062]

[0063] In the formula, To determine the threshold, The upper boundary curve, This is the lower boundary curve.

[0064] Optionally, after extracting the upper and lower boundary curves of the color band, the method further includes performing SG filtering on the upper and lower boundary curves and smoothing them by fitting:

[0065]

[0066]

[0067] In the formula, The upper boundary curve after SG filtering and lower boundary curve , Here are the SG filter coefficients, and k is the value that controls the size of the filter window, which is 2k+1. The width of the original input image, The index value of the data point within the filter window. The x-axis pixel coordinates;

[0068] upper boundary curve and lower boundary curve Map back to the original input image dimensions:

[0069]

[0070]

[0071] In the formula, This is the space scaling factor. These are the final upper and lower boundary curves.

[0072] Optionally, the BF-EdgeNet extraction model is set up for use after training. A supervised learning method is used to train the BF-EdgeNet extraction model, and the training process includes:

[0073] Create a labeled dataset: acquire a large number of column chromatogram images, and preprocess and manually label the column chromatogram images; the labels include pixel-level mask images of the color band regions, upper boundary curves, and lower boundary curves;

[0074] Define a multi-task loss function: use the binary cross-entropy loss function to calculate the loss of the pixel-level mask image, and use the Smooth L1 loss function to calculate the loss of the upper and lower boundary curves; then, perform a weighted summation of the losses of the pixel-level mask image, the upper boundary curve, and the lower boundary curve to obtain the total loss.

[0075] Parameter optimization: Based on the total loss, the parameters of the BF-EdgeNet extraction model are optimized using the backpropagation algorithm.

[0076] Optionally, the boundary feature data includes the bandwidth of the color band and the bandwidth variation trend;

[0077] Bandwidth of the ribbon for:

[0078]

[0079] Bandwidth variation trend of color band for:

[0080] ;

[0081] In the formula, These are the upper and lower boundary curves of the color band; The x-axis pixel coordinates;

[0082] The step of determining the color band segment status based on the boundary feature data includes:

[0083] When both conditions A1 and A2 are met, a new color band segment is determined to begin.

[0084] Condition A1 is: the bandwidth of the current column. Exceeding the minimum effective band width ;

[0085] Condition A2 is: the bandwidth change trend of the current column. Exceeding the bandwidth growth threshold ;

[0086] If any one of conditions B1, B2, or B3 is met, the current color band segment is determined to have ended.

[0087] Condition B1 is: the bandwidth of the current column. Below the minimum effective band width ;

[0088] Condition B2 is: the bandwidth change trend of the current column. Below the bandwidth reduction threshold ;

[0089] Condition B3 is: No valid upper or lower boundaries were detected in N consecutive columns.

[0090] Optionally, generating corresponding control commands based on the determination result of the ribbon segment status includes:

[0091] Based on the status of the ribbon segments, the detected ribbon segments are automatically numbered in chronological order. , This represents the number of color band segments, and records the start and end columns for each color band segment;

[0092] For color band segment Set the corresponding container After detecting the first Each color band segment After the initial column is set, the lower-level control module generates control commands to move the robotic arm to the container. Chemical components were collected, and the first chemical component was detected. Each color band segment After the termination column, the collection of chemical components is stopped.

[0093] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0094] This invention provides a column chromatography image analysis and control method based on BF-EdgeNet. Combining deep image analysis and intelligent execution control, it enables dynamic switching of band boundary extraction, segment identification, and receiving container switching without manual intervention. This invention employs a BN backbone feature extraction network to extract direction-sensitive multi-scale features, and fuses information from different semantic levels through an FDA feature fusion module. The ERD boundary decoder then accurately outputs the band region mask and upper and lower boundary curves. Finally, Savitzky-Golay filtering is used to achieve boundary smoothing, effectively suppressing noise interference and improving boundary detection accuracy, significantly outperforming traditional centroid tracking methods.

[0095] Based on the extraction of the upper and lower boundary positions of the color bands, the system further calculates boundary feature data such as the bandwidth and bandwidth change trend of the color bands. This provides a quantitative basis for judging color band state changes, segment identification, and intelligent control, enhancing the system's adaptability and reliability. The system can determine the start and end of color band segments in real time based on boundary change trends and drive the multi-container receiving system to perform automatic switching operations, achieving a one-to-one mapping between color band segments and receiving containers. This effectively solves the problem of difficulty in distinguishing components and resulting component mixing in existing column chromatography experiments using single-container collection. Attached Figure Description

[0096] Figure 1 This is a schematic flowchart of the column chromatography image analysis control method based on BF-EdgeNet provided in an embodiment of the present invention;

[0097] Figure 2 This is a schematic diagram illustrating the implementation of the column chromatography image analysis and control method based on BF-EdgeNet provided in this embodiment of the invention.

[0098] Figure 3 This is a schematic diagram of the structure based on BF-EdgeNet provided in an embodiment of the present invention;

[0099] Figure 4 This is a schematic diagram illustrating the working principle of the BN backbone feature extraction network provided in this embodiment of the invention.

[0100] Figure 5 This is a schematic diagram illustrating the working principle of the FDA feature fusion module provided in this embodiment of the invention.

[0101] Figure 6 This is a schematic diagram illustrating the working principle of the image processing process provided in this embodiment of the invention. Detailed Implementation

[0102] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0103] Example 1

[0104] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a column chromatography image analysis control method based on BF-EdgeNet, including the following steps:

[0105] Step S1: Obtain column chromatographic images and preprocess them.

[0106] Specifically, in this embodiment, the column chromatography image is obtained as follows:

[0107] The camera is mounted vertically aligned with the chromatographic column, ensuring the lens's field of view fully covers the effective range of band movement during column chromatography while avoiding background interference. For lighting, a stable, flicker-free diffuse LED light source is used to minimize reflections and shadows from the glass column. During image acquisition, appropriate frame rates (5–15 frames / second) and exposure parameters (2–10 milliseconds) are set to ensure continuous capture of band movement and image clarity. Images acquired by the camera are directly transmitted to the processing program in RGB format and uniformly saved as uncompressed image formats, providing raw visual input for subsequent band boundary identification and automatic control.

[0108] To improve the accuracy and robustness of the subsequent extraction model in identifying column chromatography band regions, standardized preprocessing operations are required for the acquired raw images. These operations mainly include image region cropping, noise suppression, edge smoothing, contrast enhancement, and numerical normalization. This aims to improve feature clarity, enhance boundary contrast, reduce background interference, and standardize the image format to meet the input requirements of the neural network model. Specifically, this includes:

[0109] (1.1) Image cropping, retaining only the regions containing color bands in the column chromatography image;

[0110] (1.2) The Gaussian blur algorithm is used to suppress noise and smooth the cropped column chromatogram image; Gaussian blur helps to eliminate high-frequency noise in the image while preserving the low-frequency structural information of the color band edge.

[0111] (1.3) The column chromatogram image after noise suppression and smoothing is normalized, and the pixel value of each pixel is mapped to the interval [0,1]. The calculation formula is as follows:

[0112]

[0113] in, pixel coordinates in the image The pixel value of channel c (R / G / B) ranges from 0 to 255; These are the normalized pixel values, ranging from [0,1]. These correspond to the R, G, and B color channels, respectively.

[0114] Step S2: Extract the color band boundary data from the preprocessed column chromatogram image based on the preset BF-EdgeNet extraction model.

[0115] like Figure 3 and Figure 4 As shown, the BF-EdgeNet extraction model includes a BN backbone feature extraction network, an FDA feature fusion module, and an ERD boundary decoder connected in sequence.

[0116] (2.1) The BN backbone feature extraction network includes a direction-aware initial convolutional module and multiple layers of feature extraction modules connected sequentially to the output of the initial convolutional module.

[0117] The initial convolutional module is denoted as the Stem layer, which uses a direction-sensitive convolutional structure for initial feature extraction. The feature extraction module consists of four layers, designated BandConv Block1 to Block4. Features extracted by the Stem layer are input into BandConv Block1 to Block4 for feature extraction at all four layers. The BN backbone feature extraction network extracts direction-sensitive and multi-scale semantic representation features from the preprocessed column chromatography image.

[0118] (2.1.1) The initial convolution module performs 7×1 and 1×7 convolution operations on the input image to extract horizontal and vertical features:

[0119]

[0120]

[0121] In the formula, From the input image Extracted horizontal and vertical features, for Activation function The convolution operations are 7×1 and 1×7 respectively;

[0122] lateral features and longitudinal features Perform channel concatenation, and then fuse the channels using a 3×3 convolution operation:

[0123]

[0124] In the formula, The output features of the initial convolutional module, This is a 3×3 convolution operation. This refers to the splicing operation in the channel dimension.

[0125] (2.1.2) Each feature extraction module includes a direction-aware convolution module, a channel-space joint attention module, and a downsampling module connected in sequence; specifically, four levels are set in this embodiment.

[0126] The orientation-aware convolution module performs 5×1 and 1×5 convolution operations on the input features respectively, extracting orientation information in parallel:

[0127]

[0128]

[0129] in, For the first Each input feature Convolution operations of 5×1 and 1×5 are respectively. For batch normalization operations, for Activation function From input features Extracted horizontal and vertical features;

[0130] lateral features and longitudinal features Perform channel splicing:

[0131]

[0132] In the formula, The output features of the orientation-aware convolutional module. This refers to the splicing operation at the channel dimension. The number of output channels for convolution in each direction. The height and width of the original input image;

[0133] Channel-spatial joint attention module for output features of orientation-aware convolutional module Perform global average pooling and connect it to a fully connected network. The activation function generates channel-dimensional weight coefficients. Output features of the orientation-aware convolutional module Compression is performed along the channel dimension, and a fully connected network with a sigmoid activation function is used to obtain the response weights along the spatial dimension. Output features of the channel-space joint attention module for:

[0134]

[0135] In the formula, For pixel coordinates, This is the channel number index value;

[0136] The downsampling module employs a 3×3 convolution with stride=2 downsampling strategy, which halves the spatial resolution of the feature map at each level and increases the number of channels layer by layer.

[0137]

[0138] In the formula, The output features of the downsampling module, This is a downsampling operation with a 3×3 convolution and a stride of 2. This represents the final output feature of the channel-space joint attention module.

[0139] In this embodiment, the multi-scale feature maps output by the four-level feature extraction modules are denoted as follows: These multi-scale feature maps together form the underlying feature foundation of the network, preserving the edge shape and detailed texture of the color bands, as well as capturing the overall direction and structural semantics of the color bands, providing multi-resolution and highly expressive representation support for subsequent feature fusion and accurate boundary recognition.

[0140] (2) The FDA feature fusion module is used to fuse the feature maps extracted by the feature extraction modules at multiple levels to generate a fused feature map.

[0141] like Figure 5 As shown, this specifically includes: The feature maps extracted from each level are uniformly processed using 1×1 convolutions to determine the number of channels. This is used to reduce computational complexity while ensuring consistency in subsequent splicing:

[0142]

[0143] In the formula, For the first Feature maps extracted at each level This is a 1×1 convolution operation. The feature map after a 1×1 convolution operation Feature map For multi-scale feature maps ;

[0144] feature map Size upsampling to To achieve spatial dimensional alignment with the same size:

[0145]

[0146] In the formula, This is the feature map extracted from the first layer after a 1×1 convolution operation. For upsampling operation, For size, Feature map after upsampling operation ;

[0147] Feature maps with aligned spatial dimensions and uniform channel count are stitched together along the channel dimension:

[0148]

[0149] In the formula, This refers to the splicing operation at the channel dimension. For the number of levels, To stitch together the feature maps, The height and width of the original input image;

[0150] splicing feature maps Input a deformable attention fusion module to learn attention weights with adaptive spatial offset (introducing spatial offset parameters). By using learnable sampling locations, the responsiveness to color band boundaries and blurred regions is enhanced, and the fusion results are dynamically weighted.

[0151]

[0152] In the formula, For the final fused feature map, This is a deformable attention fusion operation.

[0153] In this embodiment, feature maps are extracted at four levels. The corresponding number of channels is The number of channels is unified by 1×1 convolution. Achieve this through channel-dimensional splicing aisle.

[0154] (3) The ERD boundary decoder is used to receive the fusion features generated by the FDA feature fusion module and extract the upper and lower boundary curves of the color band.

[0155] like Figure 6As shown, after completing multi-scale feature fusion, an ERD boundary decoder was further designed to accurately locate the upper and lower boundaries of the bands in the column chromatography image. This decoder receives the fused feature map and outputs the probability map of the band region and the positions of the upper and lower boundaries. This module combines mask prediction and boundary regression mechanisms, enabling both region-level discrimination of the band shape and output of continuous upper and lower boundary curves.

[0156] Specifically, the extraction of the upper and lower boundary curves of the color band includes:

[0157] The fused feature map output by the FDA feature fusion module Perform two convolutional layers to extract salient region features of the color band area:

[0158]

[0159]

[0160] In the formula, For batch normalization operations, for Activation function This is a 3×3 convolution operation; These are the feature maps output from two convolutional layers, respectively.

[0161] Use a 1×1 convolution to transform the feature map Mapped to a single-channel mask image:

[0162]

[0163] In the formula, Use the Sigmoid activation function to make the mask value fall within the interval [0,1]. pixel coordinates The probability that a pixel belongs to a color band region;

[0164] Based on a single-channel mask image, the upper and lower boundary curves of the color band are extracted using a column-by-column scanning method:

[0165]

[0166]

[0167] In the formula, To determine the threshold, The upper boundary curve, This is the lower boundary curve.

[0168] (4) To further improve the accuracy and continuity of upper and lower boundary extraction and avoid boundary jitter and jumps caused by changes in illumination, color banding, or image noise, this embodiment introduces Savitzky-Golay filtering (SG filtering) on ​​the boundary curve after boundary extraction to achieve smooth fitting of the upper and lower boundaries. Savitzky-Golay filtering is a local polynomial smoothing method that fits a polynomial to the data points within a sliding window based on the least squares method and replaces the original data points with the fitting result, thus preserving the local trend and differential features of the boundary curve. The boundaries after SG filtering of the upper and lower boundaries are represented as follows:

[0169]

[0170]

[0171] In the formula, The upper boundary curve after SG filtering and lower boundary curve , Here are the SG filter coefficients, and k is the value that controls the size of the filter window, which is 2k+1. The width of the original input image, The index value of the data point within the filter window. The x-axis pixel coordinates;

[0172] upper boundary curve and lower boundary curve Map back to the original input image dimensions:

[0173]

[0174] In the formula, This is the space scaling factor. These are the final upper and lower boundary curves.

[0175] (5) The BF-EdgeNet extraction model is set up and used after training. The BF-EdgeNet extraction model is trained using a supervised learning method. The training process includes:

[0176] Create a labeled dataset: acquire a large number of column chromatogram images (RGB three-channel images), and preprocess and manually label the column chromatogram images; the labels include pixel-level mask images of the color band regions, upper boundary curves, and lower boundary curves;

[0177] Define a multi-task loss function: use the binary cross-entropy loss function to calculate the loss of the pixel-level mask image, and use the Smooth L1 loss function to calculate the losses of the upper and lower boundary curves; then, perform a weighted sum of the losses of the pixel-level mask image, the upper boundary curve, and the lower boundary curve to obtain the total loss. ;

[0178]

[0179] In the formula, The losses are the pixel-level mask image, the upper boundary curve, and the lower boundary curve, respectively. , , All are weighted coefficients.

[0180] Parameter optimization: Based on the total loss, the backpropagation algorithm is used to optimize the parameters of the BF-EdgeNet model. SGD and other optimization algorithms are used to update the network weights and gradually optimize the model parameters to improve the overall recognition performance.

[0181] Through this training process, the BF-EdgeNet extraction model of this invention not only has a strong ability to segment color band regions, but also can simultaneously achieve accurate positioning of the upper and lower boundary positions, providing a solid data foundation for the subsequent automatic control and segmented collection of column chromatography images.

[0182] Step S3: Perform structured processing on the extracted color band boundary data to generate boundary feature data, and determine the state of the color band segments based on the boundary feature data.

[0183] Specifically, in this embodiment, the boundary feature data includes the bandwidth of the color band and the bandwidth variation trend; the bandwidth reflects whether the color band locally diffuses, tails, or thins, and can be used for quality control and boundary anomaly identification. The bandwidth variation trend reflects the appearance / disappearance, sudden widening, or narrowing of the color band.

[0184] Bandwidth of the ribbon for:

[0185]

[0186] Bandwidth variation trend of color band for:

[0187] ;

[0188] In the formula, These are the upper and lower boundary curves of the color band; The x-axis pixel coordinates;

[0189] Determining the state of color band segments based on boundary feature data includes:

[0190] When both conditions A1 and A2 are met, a new color band segment is determined to begin.

[0191] Condition A1 is: the bandwidth of the current column. Exceeding the minimum effective band width ;

[0192] Condition A2 is: the bandwidth change trend of the current column. Exceeding the bandwidth growth threshold ;

[0193] If any one of conditions B1, B2, or B3 is met, the current color band segment is determined to have ended.

[0194] Condition B1 is: the bandwidth of the current column. Below the minimum effective band width ;

[0195] Condition B2 is: the bandwidth change trend of the current column. Below the bandwidth reduction threshold ;

[0196] Condition B3 is: No valid upper or lower boundaries were detected in N consecutive columns.

[0197] Step S4: Based on the results of the color band segment status judgment, generate corresponding control commands to realize the switching of multiple containers and the collection of chemical components;

[0198] Specifically, based on the result of the color ribbon segment status determination, the corresponding control commands generated include:

[0199] Based on the status of the ribbon segments, the detected ribbon segments are automatically numbered in chronological order. , This represents the number of color band segments, and records the start and end columns for each color band segment;

[0200] For color band segment Set the corresponding container After detecting the first Each color band segment After the initial column is set, the lower-level control module generates control commands to move the robotic arm to the container. Chemical components were collected, and the first chemical component was detected. Each color band segment After the termination column, the collection of chemical components is stopped.

[0201] The automatic identification and control system for column chromatography of this invention can extract the upper and lower boundary information of the color band with high precision, intelligently judge the change state of the color band segment, and link multiple container systems to achieve automatic segmented collection, which significantly improves the efficiency, accuracy and automation level of column chromatography experiments, and provides a brand-new technical solution for the construction of modern chemical analysis and intelligent laboratories.

[0202] Example 2

[0203] This invention provides a column chromatography image analysis and control system based on BF-EdgeNet, comprising:

[0204] The image processing module is configured to acquire column chromatographic images and preprocess the column chromatographic images;

[0205] The band extraction module is configured to extract band boundary data from the preprocessed column chromatogram image based on a preset BF-EdgeNet extraction model.

[0206] The state determination module is configured to perform structured processing on the extracted color band boundary data to generate boundary feature data, and to determine the state of the color band segment based on the boundary feature data.

[0207] The analysis and control module is configured to generate corresponding control commands based on the judgment result of the color band segment status, so as to realize the switching of multiple containers and the collection of chemical components;

[0208] The BF-EdgeNet extraction model includes a BN backbone feature extraction network, an FDA feature fusion module, and an ERD boundary decoder connected in sequence.

[0209] The BN backbone feature extraction network includes a direction-aware initiating convolutional module and multiple layers of feature extraction modules connected sequentially to the output of the initiating convolutional module.

[0210] The FDA feature fusion module is used to fuse feature maps extracted by the feature extraction module at multiple levels to generate a fused feature map;

[0211] The ERD boundary decoder is used to receive the fused features generated by the FDA feature fusion module and extract the upper and lower boundary curves of the color band.

[0212] Example 3

[0213] Based on the column chromatography image analysis control method provided in Embodiment 1, this embodiment of the invention provides an electronic device, including a processor and a storage medium;

[0214] The storage medium is used to store instructions;

[0215] The processor is configured to operate according to the instructions to perform the steps according to the method described above.

[0216] Example 4

[0217] Based on the column chromatography image analysis control method provided in Embodiment 1, this embodiment of the invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above method.

[0218] Example 5

[0219] Based on the column chromatography image analysis control method provided in Embodiment 1, this embodiment of the invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the above-described method.

[0220] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0222] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0223] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0224] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A column chromatography image analysis control method based on BF-EdgeNet, characterized in that, include: Acquire column chromatographic images and preprocess the column chromatographic images; Based on the preset BF-EdgeNet extraction model, the color band boundary data of the preprocessed column chromatographic image is extracted; The extracted color band boundary data is processed in a structured manner to generate boundary feature data, and the color band segment status is determined based on the boundary feature data. Based on the determination result of the color band segment status, corresponding control commands are generated to realize the switching of multiple containers and the collection of chemical components; The BF-EdgeNet extraction model includes a BN backbone feature extraction network, an FDA feature fusion module, and an ERD boundary decoder connected in sequence. The BN backbone feature extraction network includes a direction-aware initiating convolutional module and multiple layers of feature extraction modules connected sequentially to the output of the initiating convolutional module. The FDA feature fusion module is used to fuse feature maps extracted by the feature extraction module at multiple levels to generate a fused feature map; The ERD boundary decoder is used to receive the fused features generated by the FDA feature fusion module and extract the upper and lower boundary curves of the color band. Each layer of the feature extraction module includes a direction-aware convolution module, a channel-space joint attention module, and a downsampling module connected in sequence. The orientation-aware convolution module performs 5×1 and 1×5 convolution operations on the input features respectively, and extracts orientation information in parallel: ； ； in, For the first Each input feature Convolution operations of 5×1 and 1×5 are respectively. For batch normalization operations, for Activation function From input features Extracted horizontal and vertical features; lateral features and longitudinal features Perform channel splicing: ； In the formula, The output features of the orientation-aware convolutional module. This refers to the splicing operation at the channel dimension. The number of output channels for convolution in each direction. The height and width of the original input image; The channel-space joint attention module provides the output features of the direction-aware convolution module. Perform global average pooling and connect it to a fully connected network. The activation function generates channel-dimensional weight coefficients. ; Output features of the direction-aware convolutional module Compression is performed along the channel dimension, and a fully connected network with a sigmoid activation function is used to obtain the response weights along the spatial dimension. Output features of the channel-space joint attention module for: ； In the formula, For pixel coordinates, This is the channel number index value; The downsampling module employs a 3×3 convolution with stride=2 downsampling strategy, which halves the spatial resolution of the feature map at each level while increasing the number of channels layer by layer. ； In the formula, The output features of the downsampling module, This is a downsampling operation with a 3×3 convolution and a stride of 2. This represents the final output feature of the channel-space joint attention module; The step of fusing the feature maps extracted by the feature extraction modules at multiple levels to generate a fused feature map includes: The feature maps extracted from each level are uniformly processed using 1×1 convolutions to determine the number of channels. : ； In the formula, For the first Feature maps extracted at each level This is a 1×1 convolution operation. The feature map after a 1×1 convolution operation ; feature map Size upsampling to To achieve spatial dimensional alignment with the same size: ； In the formula, This is the feature map extracted from the first layer after a 1×1 convolution operation. For upsampling operation, For size, Feature map after upsampling operation ; Feature maps with aligned spatial dimensions and uniform channel count are stitched together along the channel dimension: ； In the formula, This refers to the splicing operation at the channel dimension. For the number of levels, To stitch together the feature maps, The height and width of the original input image; splicing feature maps Input a deformable attention fusion module, learn attention weights for adaptive spatial shift, and dynamically weight the fusion result: ； In the formula, For the final fused feature map, This is a deformable attention fusion operation.

2. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The preprocessing of the column chromatography image includes: Image cropping, retaining only the regions containing color bands in the column chromatography image; The cropped column chromatogram image was subjected to noise suppression and smoothing using a Gaussian blur algorithm. The column chromatogram image after noise suppression and smoothing is normalized, and the pixel value of each pixel is mapped to the interval [0,1].

3. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The initial convolution module performs 7×1 and 1×7 convolution operations on the input image to extract horizontal and vertical features: ； ； In the formula, From the input image Extracted horizontal and vertical features, for Activation function The convolution operations are 7×1 and 1×7 respectively; lateral features and longitudinal features Perform channel concatenation, and then fuse the channels using a 3×3 convolution operation: ； In the formula, The output features of the initial convolutional module, This is a 3×3 convolution operation. This refers to the splicing operation in the channel dimension.

4. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The upper and lower boundary curves of the extracted color band include: The fused feature map output by the FDA feature fusion module Perform two convolutional layers to extract salient region features of the color band area: ； ； In the formula, For batch normalization operations, for Activation function This is a 3×3 convolution operation; These are the feature maps output from two convolutional layers, respectively. Use a 1×1 convolution to transform the feature map Mapped to a single-channel mask image: ； In the formula, Use the Sigmoid activation function to make the mask value fall within the interval [0,1]. pixel coordinates The probability that a pixel belongs to a color band region; Based on a single-channel mask image, the upper and lower boundary curves of the color band are extracted using a column-by-column scanning method: ； In the formula, To determine the threshold, The upper boundary curve, This is the lower boundary curve.

5. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, After extracting the upper and lower boundary curves of the color band, the method further includes performing SG filtering on the upper and lower boundary curves and smoothing them by fitting: ； ； In the formula, The upper boundary curve after SG filtering and lower boundary curve , Here are the SG filter coefficients, and k is the value that controls the size of the filter window, which is 2k+1. The width of the original input image, The index value of the data point within the filter window. The x-axis pixel coordinates; upper boundary curve and lower boundary curve Map back to the original input image dimensions: ； In the formula, This is the space scaling factor. These are the final upper and lower boundary curves.

6. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The BF-EdgeNet extraction model is set up for use after training. Supervised learning is used to train the BF-EdgeNet extraction model, and the training process includes: Create a labeled dataset: acquire a large number of column chromatogram images, and preprocess and manually label the column chromatogram images; the labels include pixel-level mask images of the color band regions, upper boundary curves, and lower boundary curves; Define a multi-task loss function: use the binary cross-entropy loss function to calculate the loss of the pixel-level mask image, and use the Smooth L1 loss function to calculate the loss of the upper and lower boundary curves; then, perform a weighted summation of the losses of the pixel-level mask image, the upper boundary curve, and the lower boundary curve to obtain the total loss. Parameter optimization: Based on the total loss, the parameters of the BF-EdgeNet extraction model are optimized using the backpropagation algorithm.

7. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The boundary feature data includes the bandwidth of the color band and the bandwidth variation trend; Bandwidth of the ribbon for: ； Bandwidth variation trend of color band for: ； In the formula, These are the upper and lower boundary curves of the color band; The x-axis pixel coordinates; The step of determining the color band segment status based on the boundary feature data includes: When both conditions A1 and A2 are met, a new color band segment is determined to begin. Condition A1 is: the bandwidth of the current column. Exceeding the minimum effective band width ; Condition A2 is: the bandwidth change trend of the current column. Exceeding the bandwidth growth threshold ; If any one of conditions B1, B2, or B3 is met, the current color band segment is determined to have ended. Condition B1 is: the bandwidth of the current column. Below the minimum effective band width ; Condition B2 is: the bandwidth change trend of the current column. Below the bandwidth reduction threshold ; Condition B3 is: No valid upper or lower boundaries were detected in N consecutive columns.

8. The column chromatography image analysis and control method based on BF-EdgeNet according to claim 1, characterized in that, The step of generating corresponding control commands based on the determination result of the ribbon segment status includes: Based on the status of the ribbon segments, the detected ribbon segments are automatically numbered in chronological order. , This represents the number of color band segments, and records the start and end columns for each color band segment; For color band segment Set the corresponding container After detecting the first Each color band segment After the initial column is set, the lower-level control module generates control commands to move the robotic arm to the container. Chemical components were collected, and the first chemical component was detected. Each color band segment After the termination column, the collection of chemical components is stopped.

Images