Gibberellic acid crystal morphology identification method and system

Abstract

The present application provides a kind of gibberellic acid crystal morphology identification method and system, including obtaining the multiple frames polarized microscopic images of the gibberellic acid crystal to be identified under the preset polarization angle sequence, after being stacked as multi-channel input tensor, the fusion feature map is generated by the channel attention mechanism weighting;The feature map is input into the residual network to extract hierarchical features: the network residual unit contains main stem and short circuit connection path, the main stem path adopts the multi-scale heterogeneous convolution kernel group containing bar and square convolution kernel to do feature transformation, and is fused and output by the double-way parallel activation structure of continuous differentiable and segmented linear activation function;The short circuit connection path is a spatial gating path, generates spatial weight mask by micro convolution network, weights the input feature to obtain short circuit connection signal, and the output of residual unit is the element-by-element addition of two-way features;Based on the deep feature vector of network output, the interval loss function of morphological similarity prior is calculated by the classifier to determine the attribution probability, and the crystal morphology category is determined.

Description

Technical Field

[0001] This application belongs to the field of identification, and in particular relates to a method and system for identifying the crystal morphology of gibberellic acid. Background Technology

[0002] The current method for analyzing the morphology of gibberellic acid crystals relies on manual microscopic observation. This method is not only inefficient and time-consuming, but the identification results also depend on the operator's professional knowledge and subjective judgment, failing to meet the testing requirements of modern industrial production. Image processing and machine learning are used to identify crystals by combining manually planned feature extraction operators with classifiers, but their feature extraction capabilities are limited when faced with diverse crystal morphologies, blurred edges, or clustering or overlapping. Convolutional Neural Networks (CNNs) deep learning methods are employed for image recognition, and classic network models such as VGG, GoogLeNet, and ResNet can be used to complete the classification task of crystal microscopic images.

[0003] However, when applying the aforementioned visual model to the task of identifying gibberellic acid crystals, the crystal morphology is diverse, and the differences between different categories can be very subtle. General-purpose models are insufficient in distinguishing high inter-class similarity. Standard convolutional neural networks use square convolutional kernels, which are not sensitive enough to detecting anisotropic geometric features such as differences in the major and minor axes of crystals. The model uses a standard cross-entropy loss function during training, which does not explicitly increase the distance between easily confused categories in the feature space, leading to a blurred classification interface. Furthermore, how to integrate multi-dimensional information provided by polarization microscopy at different angles and make the network focus on the crystal's own region rather than background noise is also crucial for improving recognition accuracy. Therefore, a dedicated deep learning method that can integrate multi-source information, extract crystal features, and optimize the ability to distinguish easily confused categories is urgently needed. Summary of the Invention

[0004] To address the problems of existing technologies failing to integrate multi-source information and lacking the ability to distinguish easily confused categories.

[0005] In the first aspect, the present invention proposes a method for identifying the morphology of gibberellic acid crystals, comprising: Multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence are acquired. The multiple frames of images are stacked along the channel dimension to form a multi-channel input tensor. The multi-channel input tensor is weighted based on the channel attention mechanism to generate a fused feature map. The fused feature map is input into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path employs multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is used, with one path employing a continuously differentiable activation function and the other a piecewise linear activation function. The feature maps output from the two paths are then fused. The short-circuit connection paths are spatially gated paths. A spatial weight mask is generated using a micro-convolutional network, and this mask is multiplied element-wise with the input features of the residual units to generate a weighted short-circuit connection signal. The output of each residual unit is obtained by element-wise summing the output of the backbone path and the weighted short-circuit connection signal. Using the output of the residual network, the probability of crystal morphology is calculated by a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to visually similar class pairs according to a preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The morphological category of the gibberellic acid crystals is determined based on the assigned probability.

[0006] In another aspect, the present invention also proposes a gibberellic acid crystal morphology identification system, comprising the following modules: The weighting module is used to acquire multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence, stack the multiple frames of images along the channel dimension to form a multi-channel input tensor, and weight the multi-channel input tensor based on the channel attention mechanism to generate a fused feature map. An extraction module is used to input the fused feature map into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path employs multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is set, with the two paths using a continuously differentiable activation function and a piecewise linear activation function, respectively, and the feature maps output from the two paths are fused. The short-circuit connection path is a spatially gated path, generating a spatial weight mask through a micro-convolutional network. This mask is then multiplied element-wise with the input features of the residual units to generate a weighted short-circuit connection signal. The output of the residual unit is obtained by element-wise addition of the output of the backbone path and the weighted short-circuit connection signal. The calculation module is used to calculate the probability of crystal morphology classification by using the output of the residual network through a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to the visually similar class pairs according to the preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The determination module is used to determine the morphological category of the gibberellic acid crystals based on the attribution probability.

[0007] This invention enhances the initial feature representation of crystal optical anisotropy by fusing multi-angle polarization microscopic image information and utilizing a channel attention mechanism. In the feature extraction stage, the multi-scale heterogeneous convolutional kernel group in the deep network model, combined with a continuously differentiable activation function and piecewise linear activation, can simultaneously detect the major axis and isoaxial features of the crystal, improving the completeness of feature extraction for diverse morphologies. The dual-path parallel activation structure enhances the model's modeling ability, while the spatial gating mechanism in the short-circuit connection focuses on the region where the crystal is located, suppressing background noise interference. In the classification stage, optimization is performed using a loss function based on morphological similarity priors, increasing the spacing between visually easily confused crystal categories in the feature space, thereby improving the model's discriminative ability. Attached Figure Description

[0008] Figure 1 This is a flowchart of the gibberellic acid crystal morphology identification method in an embodiment; Figure 2 A schematic diagram of input tensors for multichannel polarization microscopy images. Detailed Implementation

[0009] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0010] This invention provides a method for identifying the morphology of gibberellic acid crystals, such as... Figure 1 As shown, the method includes: S1. Acquire multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence. Stack the multiple frames of images along the channel dimension to form a multi-channel input tensor. Weight the multi-channel input tensor based on the channel attention mechanism to generate a fused feature map. By controlling an electrically controlled rotary stage associated with the polarization microscope's optical path, a stepper motor rotates the polarizer or analyzer, setting the polarization angles in a sequence of 0°, 45°, 90°, and 135°. An industrial camera acquires a 1024×1024 grayscale image at each angle. The four images are read using the `imread` function from the OpenCV library. Then, using the `stack` function from the NumPy library or the `cat` function from the PyTorch deep learning framework, the `axis` parameter is set to the channel dimension, stacking the four single-channel images into a four-channel input tensor, as shown below. Figure 2 As shown, the four-channel tensor is input into a Squeeze-and-Excitation module. This module compresses the feature map of each channel into a real number through a global average pooling operation, resulting in a channel representation vector. This vector is then passed through two fully connected layers and a sigmoid activation function to generate weight coefficients for each channel. These weight coefficients are then multiplied element-wise with the corresponding channels of the original input tensor to obtain a weighted fused feature map.

[0011] In an alternative embodiment, acquiring multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence includes: With the polarizer fixed, the analyzer is rotated from 0° to 135° in 45° increments, and one frame of polarization microscopic image is acquired at each of the four angles of 0°, 45°, 90° and 135°.

[0012] Under a microscope equipped with a polarizer, rotating stage, and focusing system, a slide containing a gibberellic acid crystal sample is placed on the stage. To ensure consistent image quality, the microscope's optical path is calibrated using the Köhler illumination method, and a constant light source brightness is set. Before the acquisition sequence begins, the crystal sample is clearly imaged via the focusing module or manual fine-tuning. During acquisition, the polarization direction of the polarizer is fixed, for example, set to 0° horizontally. The analyzer is controlled by a precision stepper motor to set the initial angle to 0°, at which point the first frame is acquired, preferably an 8-bit grayscale image with a resolution of 512×512 pixels. The analyzer is rotated 45° clockwise, and the second frame is acquired at the 45° position, maintaining the focal point throughout the acquisition process. This operation is repeated sequentially, acquiring the third and fourth frames at 90° and 135° positions, respectively. These four frames were used to detect the extinction and transmission characteristics of the crystal under different polarization interference conditions, and they are temporally aligned and spatially registered. Before stacking, each frame of the image is normalized, for example, by linearly scaling the pixel values ​​to the [0,1] range to eliminate minor brightness fluctuations during the acquisition process. The four 512×512 single-channel images are stacked to obtain a multi-channel input tensor of size 512×512×4, where the channel dimensions correspond to the four polarization angles.

[0013] In an alternative embodiment, the weighting of the multi-channel input tensor based on the channel attention mechanism includes: Global average pooling is performed on the multi-channel input tensor to obtain the channel representation vector; The channel representation vector is sequentially fed into a fully connected dimensionality reduction layer, a ReLU activation function, a fully connected dimensionality increase layer, and a Sigmoid activation function to generate the weight coefficients for each channel. The weighting coefficients are multiplied element-wise with the corresponding channels of the original multi-channel input tensor to complete the weighting.

[0014] This weighting process is implemented through a channel attention network module, whose input is a feature map of size H×W×C. This module consists of two parts: Squeeze and Excitation.

[0015] The Squeeze part uses a global average pooling layer to process the input feature map. Compression is performed in the spatial dimension, generating a scalar for each channel, resulting in a 1×1×C channel representation vector z. The calculation formula is as follows: .

[0016] The excitation part consists of two fully connected layers and an activation function. Vector z is passed through a fully connected dimensionality reduction layer and a ReLU activation function, outputting a dimension of C / r, where r is the dimensionality reduction ratio. A fully connected dimensionality increase layer restores the dimension to C. Channel weights are generated using a Sigmoid activation function. The formula for this process is as follows: ,in , , It is the ReLU activation function. This is the Sigmoid activation function.

[0017] Output a channel weight vector s of size 1×1×C, with element values ​​between 0 and 1, representing the importance of each channel.

[0018] S2, the fused feature map is input into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path uses multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is set up, with the two paths employing a continuously differentiable activation function and a piecewise linear activation function, respectively. The feature maps output from the two paths are then fused. The short-circuit connection path is a spatially gated path. A spatial weight mask is generated through a micro-convolutional network, and the mask is multiplied element-wise with the input features of the residual unit to generate a weighted short-circuit connection signal. The output of the residual unit is obtained by element-wise addition of the output of the backbone path and the weighted short-circuit connection signal. In the main path, the multi-scale heterogeneous convolutional kernel group is implemented by setting up three convolutional branches in parallel. The first branch uses a two-dimensional convolutional layer with a kernel size of 1×7, the second branch uses a two-dimensional convolutional layer with a kernel size of 7×1, and the third branch uses a two-dimensional convolutional layer with a kernel size of 3×3. The bar convolutional kernel is used to detect the anisotropic features of the long axis of the crystal, and the square convolutional kernel is used to detect the isotropic features of the crystal. After all branches have processed the input features, the PyTorch cat function is used to concatenate the output feature map along the channel dimension, and then a convolutional layer with a kernel size of 1×1 is used for feature fusion and dimensionality reduction. A dual-path parallel activation structure receives the fused feature maps, with one path using the Mish activation function and the other using the ReLU activation function. The output feature maps from the two activation paths are then fused element-wise by addition. In the short-circuit connection path, the spatial gating signal is generated by a miniature convolutional network. This network receives the input features from the residual unit and sequentially passes them through a 3×3 convolutional layer, a ReLU activation function, a 1×1 convolutional layer, and a Sigmoid activation function to generate a spatial weight mask with the same spatial dimension as the input feature map and values ​​ranging from 0 to 1. This mask is combined with the original input features of the residual unit using element-wise multiplication to obtain the weighted short-circuit connection signal. The output of the main path is then element-wise added to the weighted short-circuit connection signal to complete the computation of the residual unit. The dual-path parallel activation structure, in the negative range, uses the small negative responses preserved by Mish to compensate for the neuron death caused by the direct truncation of ReLU, ensuring stable backpropagation of gradients in deep networks. In the positive range and near zero, the superposition of both enhances the response intensity to positive salient features and smooths the hard inflection points of ReLU through the smoothing properties of Mish. This enables the model to have stronger generalization ability and more stable convergence speed, helping to distinguish the small and easily confused morphological details of gibberellic acid crystals. Those skilled in the art should know that if the dimensions are inconsistent in the calculation, they can be converted to be consistent through upsampling, downsampling, pooling, etc., which is common knowledge and will not be elaborated upon in this invention.

[0019] In an optional embodiment, the network structure is specifically composed of a backbone path and a spatially gated short-circuit connection path in parallel, and the outputs of the two paths are fused by adding them element by element.

[0020] Input feature map The convolutional layers are processed in parallel using three sets of convolutional kernels with different shapes: a 3×3 square kernel, a 1×5 stripe kernel, and a 5×1 stripe kernel. The output feature maps of the three sets of convolutions are concatenated along the channel dimension, and then fused across channels using a 1×1 convolutional layer to obtain the transformed feature map. .Will The function is copied and fed into a dual-path parallel activation structure. One path uses the Mish activation function, and the other uses the ReLU activation function. The output feature maps from both paths are summed element-wise to obtain the output of the main path. .

[0021] Input feature map Enter a mini convolutional network to generate a spatial weight mask. This miniature network consists of a 1×1 convolutional layer for channel compression, a 7×7 depthwise separable convolutional layer for spatial feature extraction, and another 1×1 convolutional layer to reduce the number of channels to 1, outputting the result through a sigmoid activation function. The resulting spatial weight mask... With the original input Element-wise multiplication yields a weighted short-circuit connection signal. Output feature map with dimensions H×W×C .

[0022] In an alternative embodiment, feature transformation is performed within the backbone path using a multi-scale heterogeneous convolutional kernel group, including: The multi-scale heterogeneous convolution kernel group used in the main path consists of a 3×3 square convolution kernel, a 1×5 strip convolution kernel, and a 5×1 strip convolution kernel; Three sets of convolutional kernels are applied in parallel to the input feature map, and the three output feature maps are concatenated along the channel dimension. A 1×1 convolutional layer is then used to adjust the number of channels to match the input feature map. Figure 1 This is the output of the convolution transformation.

[0023] In an alternative embodiment, a dual-path parallel activation structure is configured, with the two paths employing a continuously differentiable activation function and a piecewise linear activation function, respectively. The feature maps output from the two paths are then fused, including: The feature map after convolution is copied into two copies. One copy is input into the Mish activation function as the continuously differentiable activation function, and the other copy is input into the ReLU activation function as the piecewise linear activation function. The two output feature maps obtained after processing by the two activation functions are added element by element to obtain the fused activation output.

[0024] In an alternative embodiment, the short-circuit connection path is a spatially gated path, which generates a spatial weight mask through a micro-convolutional network, including: The input features of the residual unit are sequentially passed through a 1×1 convolutional layer for channel compression, a 7×7 depth-separable convolutional layer for spatial feature extraction, and another 1×1 convolutional layer to adjust the number of channels to 1. The processed feature map is passed through a Sigmoid function to generate a spatial weight mask with values ​​ranging from 0 to 1.

[0025] S3. Using the output of the residual network, the probability of crystal morphology is calculated by a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to the visually similar class pairs according to the preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The feature map output from the last layer of the residual network is transformed into a one-dimensional deep feature vector through a global average pooling layer; this vector is then input into a fully connected classification layer. During training, the L2 norm normalization function is used to normalize both the feature vector and the weight vector of the fully connected layer. A C×C dimensional morphological similarity prior matrix M is pre-constructed, where C is the total number of classes, and the matrix elements are... This represents the morphological similarity between categories y and j; higher similarity results in larger values. For a training sample belonging to category y, the dot product of the normalized features and the normalized weight vector is calculated to obtain the cosine similarity across all categories. The corresponding angle is then solved using the inverse cosine function (arccos). Based on the similarity between category y and other categories j in the similarity matrix M. ,for Apply an angular interval ,in The larger, The larger it is, the greater the new angle will be after the interval is increased. The target logit value is converted back using the cosine function (cos). The modified target logit value, along with the original logit values ​​for all other classes, is then fed into the SoftmaxWithCrossEntropy loss function to calculate the gradient and update the network parameters.

[0026] In an alternative embodiment, the loss function, in the normalized hyperspherical feature space, applies a large angular interval to visually similar class pairs based on a preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space, including: Construct a C×C morphological similarity prior matrix M, where C is the total number of morphological categories, and the matrix elements are... Crystallography experts assign values ​​to morphology i and morphology j based on the degree of visual confusion, with scores ranging from 0 to 1. When calculating the loss for category y, for any non-target category j, the angular margin penalty term is obtained by adding the base margin and the penalty increment, where the penalty increment is the similarity score between category y and category j. The product of the adjusted hyperparameters.

[0027] Assume the total number of gibberellic acid crystal morphology categories is C=6, including needle-like, columnar, plate-like, granular, flaky, and clustered forms. Construct a 6×6 symmetric similarity prior matrix M, where... The values ​​are independently scored by three crystallography experts, ranging from 0 to 1, where 1 indicates extreme similarity. The average value is then used to determine the final value. For example, M(acicular, columnar) = 0.85, M(acicular, granular) = 0.10, and the diagonal element M(i,i) = 0. During training, for an input sample, both the deep feature vector x output by the network and the weight matrix W of the classifier are L2 normalized and mapped onto the unit hypersphere. The center of x with respect to each class, i.e., the weight vector, is calculated. The cosine value of the included angle between When calculating the cross-entropy loss, for a sample belonging to class y, the angular distance between the sample and the non-target class j is... Set as: .in, It is a basic interval, with a preferred value of 0.35 radians, to ensure that there is a basic degree of differentiation between all categories; This is an adjustable hyperparameter used to control the influence strength of the similarity prior, preferably 0.15 radians. For easily confused needle-like and columnar pairs, the applied angular interval is 0.4775 radians; while for easily distinguishable needle-like and granular pairs, the applied angular interval is only 0.365 radians. After being scaled by a fixed scaling hyperparameter s, preferably s=32, the loss function is defined as a modification of the Softmax input, forcing the model to leave a larger decision boundary for visually more similar class pairs, thereby improving the recognition accuracy of easily confused shapes.

[0028] S4. Determine the morphological category of the gibberellic acid crystals based on the assigned probability.

[0029] During the inference phase, the image of the crystal to be identified is input into the trained network to obtain the raw score vector, i.e., logits, output by the fully connected classifier. The Softmax function is then called to convert this vector into a probability distribution representing the classification of each category. The argmax function is used to find the index corresponding to the element with the largest value in this probability distribution vector; this index represents the crystal morphology category predicted by the model.

[0030] This invention also provides a gibberellic acid crystal morphology identification system, the system comprising the following modules: The weighting module is used to acquire multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence, stack the multiple frames of images along the channel dimension to form a multi-channel input tensor, and weight the multi-channel input tensor based on the channel attention mechanism to generate a fused feature map. An extraction module is used to input the fused feature map into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path employs multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is set, with the two paths using a continuously differentiable activation function and a piecewise linear activation function, respectively, and the feature maps output from the two paths are fused. The short-circuit connection path is a spatially gated path, generating a spatial weight mask through a micro-convolutional network. This mask is then multiplied element-wise with the input features of the residual units to generate a weighted short-circuit connection signal. The output of the residual unit is obtained by element-wise addition of the output of the backbone path and the weighted short-circuit connection signal. The calculation module is used to calculate the probability of crystal morphology classification by using the output of the residual network through a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to the visually similar class pairs according to the preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The determination module is used to determine the morphological category of the gibberellic acid crystals based on the attribution probability.

[0031] In an alternative embodiment, acquiring multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence includes: With the polarizer fixed, the analyzer is rotated from 0° to 135° in 45° increments, and one frame of polarization microscopic image is acquired at each of the four angles of 0°, 45°, 90° and 135°.

[0032] In an alternative embodiment, the weighting of the multi-channel input tensor based on the channel attention mechanism includes: Global average pooling is performed on the multi-channel input tensor to obtain the channel representation vector; The channel representation vector is sequentially fed into a fully connected dimensionality reduction layer, a ReLU activation function, a fully connected dimensionality increase layer, and a Sigmoid activation function to generate the weight coefficients for each channel. The weighting coefficients are multiplied element-wise with the corresponding channels of the original multi-channel input tensor to complete the weighting.

[0033] In an alternative embodiment, feature transformation is performed within the backbone path using a multi-scale heterogeneous convolutional kernel group, including: The multi-scale heterogeneous convolution kernel group used in the main path consists of a 3×3 square convolution kernel, a 1×5 strip convolution kernel, and a 5×1 strip convolution kernel; Three sets of convolutional kernels are applied in parallel to the input feature map, and the three output feature maps are concatenated along the channel dimension. A 1×1 convolutional layer is then used to adjust the number of channels to match the input feature map. Figure 1 This is the output of the convolution transformation.

[0034] In an alternative embodiment, a dual-path parallel activation structure is configured, with the two paths employing a continuously differentiable activation function and a piecewise linear activation function, respectively. The feature maps output from the two paths are then fused, including: The feature map after convolution is copied into two copies. One copy is input into the Mish activation function as the continuously differentiable activation function, and the other copy is input into the ReLU activation function as the piecewise linear activation function. The two output feature maps obtained after processing by the two activation functions are added element by element to obtain the fused activation output.

[0035] In an alternative embodiment, the short-circuit connection path is a spatially gated path, which generates a spatial weight mask through a micro-convolutional network, including: The input features of the residual unit are sequentially passed through a 1×1 convolutional layer for channel compression, a 7×7 depth-separable convolutional layer for spatial feature extraction, and another 1×1 convolutional layer to adjust the number of channels to 1. The processed feature map is passed through a Sigmoid function to generate a spatial weight mask with values ​​ranging from 0 to 1.

[0036] In an alternative embodiment, the loss function, in the normalized hyperspherical feature space, applies a large angular interval to visually similar class pairs based on a preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space, including: Construct a C×C morphological similarity prior matrix M, where C is the total number of morphological categories, and the matrix elements are... Crystallography experts assign values ​​to morphology i and morphology j based on the degree of visual confusion, with scores ranging from 0 to 1. When calculating the loss for category y, for any non-target category j, the angular margin penalty term is obtained by adding the base margin and the penalty increment, where the penalty increment is the similarity score between category y and category j. The product of the adjusted hyperparameters.

[0037] In this specification, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise limited, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. In this document, "a," "an," "the," "the," and "its" may also include plural forms unless the context clearly indicates otherwise. "Multiple" refers to at least two, such as 2, 3, 5, or 8, etc. "And / or" includes any and all combinations of the associated listed items.

[0038] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

[0039] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying the morphology of gibberellic acid crystals, characterized in that, Includes the following steps: Multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence are acquired. The multiple frames of images are stacked along the channel dimension to form a multi-channel input tensor. The multi-channel input tensor is weighted based on the channel attention mechanism to generate a fused feature map. The fused feature map is input into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path employs multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is used, with one path employing a continuously differentiable activation function and the other a piecewise linear activation function. The feature maps output from the two paths are then fused. The short-circuit connection paths are spatially gated paths. A spatial weight mask is generated using a micro-convolutional network, and this mask is multiplied element-wise with the input features of the residual units to generate a weighted short-circuit connection signal. The output of each residual unit is obtained by element-wise summing the output of the backbone path and the weighted short-circuit connection signal. Using the output of the residual network, the probability of crystal morphology is calculated by a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to visually similar class pairs according to a preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The morphological category of the gibberellic acid crystals is determined based on the assigned probability.

2. The method according to claim 1, characterized in that, The process of acquiring multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence includes: With the polarizer fixed, the analyzer is rotated from 0° to 135° in 45° increments, and one frame of polarization microscopic image is acquired at each of the four angles of 0°, 45°, 90° and 135°.

3. The method according to claim 1, characterized in that, The weighting of the multi-channel input tensor based on the channel attention mechanism includes: Global average pooling is performed on the multi-channel input tensor to obtain the channel representation vector; The channel representation vector is sequentially fed into a fully connected dimensionality reduction layer, a ReLU activation function, a fully connected dimensionality increase layer, and a Sigmoid activation function to generate the weight coefficients for each channel. The weighting coefficients are multiplied element-wise with the corresponding channels of the original multi-channel input tensor to complete the weighting.

4. The method according to claim 3, characterized in that, The main path employs multi-scale heterogeneous convolutional kernel groups for feature transformation, including: The multi-scale heterogeneous convolution kernel group used in the main path consists of a 3×3 square convolution kernel, a 1×5 strip convolution kernel, and a 5×1 strip convolution kernel; Three sets of convolutional kernels are applied in parallel to the input feature map, and the three output feature maps are concatenated along the channel dimension. The number of channels is then adjusted to match the input feature map using a 1×1 convolutional layer, which serves as the output of the convolutional transformation.

5. The method according to claim 1, characterized in that, A dual-path parallel activation structure is configured, with one path employing a continuously differentiable activation function and the other a piecewise linear activation function. The feature maps output from the two paths are then fused, including: The feature map after convolution is copied into two copies. One copy is input into the Mish activation function as the continuously differentiable activation function, and the other copy is input into the ReLU activation function as the piecewise linear activation function. The two output feature maps obtained after processing by the two activation functions are added element by element to obtain the fused activation output.

6. The method according to claim 1, characterized in that, The short-circuit connection path is a spatially gated path, which generates a spatial weight mask through a micro-convolutional network, including: The input features of the residual unit are sequentially passed through a 1×1 convolutional layer for channel compression, a 7×7 depth-separable convolutional layer for spatial feature extraction, and another 1×1 convolutional layer to adjust the number of channels to 1. The processed feature map is passed through a Sigmoid function to generate a spatial weight mask with values ​​ranging from 0 to 1.

7. The method according to any one of claims 1-6, characterized in that, The loss function, in the normalized hyperspherical feature space, applies a large angular interval to class pairs with high visual similarity based on a preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space, including: Construct a C×C morphological similarity prior matrix M, where C is the total number of morphological categories, and the matrix elements are... Crystallography experts assign values ​​to morphology i and morphology j based on the degree of visual confusion, with scores ranging from 0 to 1. When calculating the loss for category y, for any non-target category j, the angular margin penalty term is obtained by adding the base margin and the penalty increment, where the penalty increment is the similarity score between category y and category j. The product of the adjusted hyperparameters.

8. A gibberellic acid crystal morphology identification system, characterized in that, Includes the following modules: The weighting module is used to acquire multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence, stack the multiple frames of images along the channel dimension to form a multi-channel input tensor, and weight the multi-channel input tensor based on the channel attention mechanism to generate a fused feature map. An extraction module is used to input the fused feature map into a residual network for hierarchical feature extraction. The residual units in the network consist of a backbone path and short-circuit connection paths. The backbone path employs multi-scale heterogeneous convolutional kernels for feature transformation. A dual-path parallel activation structure is set, with the two paths using a continuously differentiable activation function and a piecewise linear activation function, respectively, and the feature maps output from the two paths are fused. The short-circuit connection path is a spatially gated path, generating a spatial weight mask through a micro-convolutional network. This mask is then multiplied element-wise with the input features of the residual units to generate a weighted short-circuit connection signal. The output of the residual unit is obtained by element-wise addition of the output of the backbone path and the weighted short-circuit connection signal. The calculation module is used to calculate the probability of crystal morphology classification by using the output of the residual network through a classifier. The classifier is trained using an interval loss function based on morphological similarity prior. In the normalized hyperspherical feature space, the loss function applies a large angular interval to the visually similar class pairs according to the preset morphological similarity prior matrix, thereby increasing the inter-class distance of easily confused morphological features in the angular space. The determination module is used to determine the morphological category of the gibberellic acid crystals based on the attribution probability.

9. The system according to claim 8, characterized in that, The process of acquiring multiple frames of polarized microscopic images of the gibberellic acid crystal to be identified under a preset polarization angle sequence includes: With the polarizer fixed, the analyzer is rotated from 0° to 135° in 45° increments, and one frame of polarization microscopic image is acquired at each of the four angles of 0°, 45°, 90° and 135°.

10. The system according to claim 8, characterized in that, The weighting of the multi-channel input tensor based on the channel attention mechanism includes: Global average pooling is performed on the multi-channel input tensor to obtain the channel representation vector; The channel representation vector is sequentially fed into a fully connected dimensionality reduction layer, a ReLU activation function, a fully connected dimensionality increase layer, and a Sigmoid activation function to generate the weight coefficients for each channel. The weighting coefficients are multiplied element-wise with the corresponding channels of the original multi-channel input tensor to complete the weighting.

Images