34 results about "Sigmoid activation function" patented technology

The invention discloses a GRU-based vehicle-mounted CAN intrusion detection method and system, a storage medium and a computer system. According to the method, an automatic encoder (AE) is introduced to deepen the understanding of a model on input sequence characteristics, a sliding window is used for selecting batch CAN data to be preprocessed to obtain 13-dimensional time sequence data, and a scalar value within the range of [0, 1] is obtained through processing of the encoder, a GRU, a decoder, a full connection layer and a sigmoid activation function and used for classification of abnormal data. The Conv1D is used as a hidden layer, and compared with two-dimensional convolution, the one-dimensional convolution parameter quantity is smaller, and the calculation is simpler and more convenient. An attack message and a normal message can be completely distinguished, the precision and the accuracy rate reach 100%, and the precision in Fuzz detection is 0.9983; compared with the prior art, the method has high accuracy and reliability in the aspect of intrusion behavior detection, can effectively identify most intrusion events, and can keep a relatively low overall error rate, so that good balance between safety and availability is realized.

This application provides a joint emotion recognition method and system based on action unit-driven attention, relating to the fields of computer vision and affective computing. The method includes: inputting a feature map into an action unit (AU)-driven attention branch, processing it through a sigmoid activation function to obtain an activation vector, and inputting this vector into a learnable mapping layer; projecting the AU information back into the spatial geometric space to generate a single-channel spatial attention map, which is then multiplied element-wise with the feature map to obtain a weighted feature map; determining continuous values ​​for discrete emotion category probabilities, valence, and arousal; constructing a total loss function; and training the model using the acquired AffectNet and Aff-wild2 datasets to determine the target model for emotion recognition. This application achieves end-to-end joint optimization of AU detection and emotion recognition, significantly enhancing the interpretability and generalization ability of the model while improving emotion recognition accuracy, enabling it to better adapt to the emotion recognition needs in complex scenarios.

The application relates to the technical field of deep learning and recommendation system, and particularly discloses a hybrid expert model sparse inference method and system for generative recommendation, which comprises the following steps: obtaining original input data, and obtaining an input vector through an embedding layer; obtaining hybrid expert weights through normalization and maximum value selection operation on the input vector; calculating attention-enhanced features according to the hybrid expert weights based on a hierarchical attention mechanism; inputting the input vector and the attention-enhanced features into a hybrid expert model to calculate a recommendation result; wherein, according to the attention-enhanced features, expert weights are calculated based on a parallelized gating mechanism activated by a Sigmoid activation function; the activated expert layer is selected according to the expert weights, and the input vector is used to calculate the recommendation result. The application can achieve a recommendation accuracy comparable to or even higher than that of an advanced dense model under low time overhead and low calculation complexity.

This invention discloses a medical image segmentation method based on enhanced attention. The method acquires a medical image, preprocesses it, and inputs it into an encoder to output a multi-scale feature map of the medical image. The encoder includes N cascaded multi-dimensional attention modules. Each multi-dimensional attention module extracts multi-scale features through dual-channel convolution and residual connections, and performs feature recalibration using an attention enhancement module. The output multi-scale feature map is then fused by a multi-scale dilated fusion attention module to obtain the feature map output by the encoder. The feature map output by the encoder is input to the decoder, where it is combined with high-resolution feature maps of the corresponding levels passed from the encoder via skip connections. Through progressive upsampling and feature fusion, a reconstructed feature map is obtained. The reconstructed feature map output by the decoder is used to generate a medical image segmentation mask through 1×1 convolution and a sigmoid activation function to complete the segmentation of diseased tissues. This invention significantly improves the boundary segmentation accuracy for knee joints, breast tumors, and skin lesions.

The invention discloses an airway segmentation network based on multi-scale directional attention and local graph convolution fusion, and relates to the field of medical image processing, a U-Net type three-dimensional encoder-decoder structure is adopted, space detail information is reserved and gradient propagation is promoted through jump connection, an encoder is used for extracting multi-level semantic features, and the decoder is used for extracting multi-level semantic features; a decoder gradually recovers spatial resolution through jump connection, a segmentation result is obtained through a convolution layer and a Sigmoid activation function, an airway segmentation network integrates a multi-scale direction attention mechanism and a local graph convolution attention mechanism, enhancement is carried out in combination with relative position coding, the identification capability of small airway branches is improved, and boundary false detection is reduced. According to the method, the recognition rate of the tail end fine bronchus is remarkably improved, breakage and leak detection are reduced, the connectivity and topological integrity of the airway tree are enhanced, the model is more sensitive to feature response of small airways, fuzzy boundaries and low-contrast areas, the continuity of the output airway model is higher, and the method can be directly used for bronchoscope robot navigation and path planning.

This invention discloses a method and system for analyzing and outputting TCM acupuncture cases based on machine learning, belonging to the field of medical image processing technology. The method acquires facial image sequences from two pathological cycles of the subject, extracts spatiotemporal feature maps, and obtains local receptive field feature matrix pairs through sliding windowing. It calculates the cross-cycle joint entropy gradient and its matrix, and generates a two-dimensional deformation vector using optical flow. Using the coordinates of the first cycle as a reference, the deformation vector is set to zero when the joint entropy gradient does not exceed the rigidity threshold; otherwise, the joint entropy gradient matrix is ​​used as a weight matrix through a sigmoid activation function, and a Hadamard product is performed with the deformation vector to obtain a denoised deformation vector. The denoised vectors are superimposed to obtain corrected pixel coordinates, which are mapped to the target localization coordinate matrix, and the acupuncture case analysis results are output. This scheme achieves precise spatial decoupling of local tissue evolution, eliminates the contamination of rigid regions by global deformation, and significantly improves the robustness and accuracy of cross-cycle acupuncture target localization.

The invention discloses a micro-expression recognition method based on a position-motion double-flow fusion network. The micro-expression recognition method comprises two parts: designing a position extraction branch PEM based on Vision TransformerVIT; designing a motion feature extraction branch based on dynamic cross attention; a depth separable convolutional layer is embedded in front of the multi-head self-attention MSA module. A spatial adaptive feed-forward network is designed to replace a multi-layer perceptron in a standard VIT. A grouping processing strategy is adopted, and input features are divided into a plurality of sub-feature groups according to channel dimensions. A parallel sub-network is constructed to focus on feature extraction of different dimensions. And finally, generating a gating signal through a Sigmoid activation function by the interacted features, and performing element-by-element multiplication on the gating signal and the original input features. Through the dynamic weighting mechanism, which key features are reserved and which irrelevant noise is inhibited can be adaptively determined, so that accurate extraction of weak motion features is realized.

The invention provides a load data missing value dynamic filling method and device, equipment and a storage medium. The method comprises the steps that firstly, a real-time power grid parameter input vector of a clean heating transformer area is acquired, a space estimation result is generated through a KNN algorithm, and a time sequence prediction result is generated through an LSTM model; and the meta-model generates output through a full-connection ReLU hidden layer and transmits the output to an output layer, and a dynamic fusion weight is generated through Sigmoid activation function processing. And fusing the results to generate a final filling value, and inputting the final filling value into a load prediction model. According to the method, the KNN and LSTM results are dynamically fused, the load prediction precision is improved, optimal dispatching of the power distribution network is supported, the clean heating operation cost is reduced, and the wind and light abandoning phenomenon is reduced.

The lithium ion battery SOC prediction method based on the CEEMDAN and the optimized BiLSTM comprises the following steps: decomposing preprocessed operation data by adopting adaptive noise complete ensemble empirical mode decomposition CEEMDAN to obtain a plurality of intrinsic mode functions (IMF) and a residual component; constructing a BiLSTM model, wherein the model comprises a forward LSTM unit, a backward LSTM unit and an output layer; a dropout optimization algorithm is introduced into the BiLSTM model, and a combination of a two-parameter sigmoid activation function and a softsign activation function is adopted as a neuron activation function; and inputting the intrinsic mode function IMF and the residual component into the optimized BiLSTM model, optimizing model parameters through training, and predicting the state of charge SOC of the lithium ion battery by using the trained model. Aiming at the problems of weak data nonlinear processing capability, model overfitting, gradient dispersion and the like in the existing SOC estimation method, the invention provides a lithium ion battery SOC prediction method based on CEEMDAN and BiLSTM optimization, so as to improve the SOC prediction precision and model stability.

This invention belongs to the field of image processing and relates to a Retinex-based low-light image enhancement method. It reveals degraded content in the image through a degraded appearance restorer and performs Retinex decomposition using two branches consisting of convolution and a sigmoid activation function to estimate the reflectance and luminance components, respectively. After obtaining the estimated reflectance and luminance components, gamma correction is used to adjust the luminance component. This invention decouples the low-light image into a three-channel color map and a grayscale detail map to maintain consistency with the target image in color and detail representation. Furthermore, this application provides an unsupervised loss function to constrain the solution space of the Retinex decomposition, thereby improving adaptability in unknown and complex scenes. Extensive experiments demonstrate that this method outperforms state-of-the-art methods.

A gas overflow detection method based on contrast learning, comprising the following steps: S1, a contrast learning training stage, a model weight of picture feature extraction is trained, the model comprises three branches respectively for feature extraction, then three features are fused to obtain a final picture feature I o ; S2, the last model in step S1 is frozen, and two fully connected layers are connected thereafter, a nonlinear feature is obtained through a sigmoid activation function between the fully connected layers, finally a feature map of a specific dimension size of the feature is output, an output of positive and negative classification is obtained, representing the classification of whether there is gas overflow; S3, a preset gas leakage confidence threshold value, a picture and a thermal map acquisition device are fixed at a certain angle to fixed-point shoot the device to be acquired, an RGB three-channel picture corresponding to the device and a corresponding single-channel thermal map are acquired, the two fused pictures are input into the model obtained in step S2 for gas overflow classification prediction. Through the present application, gas leakage detection can be better recognized.

The application discloses a bridge risk identification method and system fusing integrated learning and an attention mechanism, and the method comprises the following steps: processing an i-th input feature in an input data set to obtain a reconstructed feature vector; constructing XGBoost and RF-MLP models to perform deep feature extraction, and acquiring fusion features by using an attention mechanism; inputting the fusion features into a back-end multi-task perception decoder, and using parallel decoding branches to decouple and identify the fusion features, wherein a first full connection layer and a first Softmax activation function are used as a risk type decoding branch to output a risk type probability distribution; a second full connection layer and a second Softmax activation function are used as a position decoding branch to output a position probability distribution; a third full connection layer and a Sigmoid activation function are used as a damage degree decoding branch in cooperation with linear mapping to output a damage degree quantitative index; and the application has the advantage of high identification precision.

This invention relates to the fields of digital image processing and computer vision technology, and discloses an automatic adjustment method for ultrasound image enhancement, comprising: acquiring discrete ultrasound echo data streams; calculating in real time, based on a sliding window, the local grayscale range characterizing the physical contrast limit and the local coefficient of variation characterizing signal fluctuation characteristics; performing threshold dynamic drift determination based on physical contrast constraints, and constructing a dynamic threshold adjustment factor; using this factor to modulate the response center threshold of the sigmoid activation function in real time, and generating image enhancement weights through dual-parameter collaborative nonlinear mapping; and using this weight to perform multiplicative modulation on the original pixels. This invention establishes a dynamic coupling mechanism between local physical contrast and the response interval of the enhancement function, thereby achieving adaptive differential processing of high-depth attenuation regions and strong reflection regions, effectively improving the imaging signal-to-noise ratio and boundary sharpness in complex sound field environments.

The application discloses a kind of context-aware mask feature enhancement synthetic speech detection methods, it is related to speech signal processing technical field, method includes: to original audio data pre-processing, the speech data after pre-processing is input to SincNet filter extraction speech timing feature;The speech timing feature is input to CAM-ResNet extraction context-aware local feature;Context-aware local feature is input L-Transformer network calculation global feature;Again through fully connected layer and sigmoid activation function classification;After iteration training, select the best model for synthetic speech detection.CAM-ResNet network in the application can effectively weaken irrelevant features, focus on context-related features, thereby improving the perception of context-effective features, capturing the local details of speech context-aware features;Again after having L-Transformer deep network calculation global feature, to improve the detection accuracy of synthetic speech.

The application discloses an FPGA accelerator for realizing high-quality new view synthesis based on a neural radiation field; the FPGA accelerator comprises an FPGA and three operation modules, namely, a position coding module, an MLP module and a ray stepping module, and the operation modules are arranged on the FPGA; the MLP module is split into two-stage pipelines; a front-stage pipeline is used to derive body density and a hidden layer, and a rear-stage pipeline is used to derive color information; the front-stage pipeline has eight fully-connected layers, and there are ReLu activation functions between the fully-connected layers; a coordinate signal is subjected to the front-stage pipeline to obtain the body density; the rear-stage pipeline has two fully-connected layers, and input hidden layer information and a view direction, and then is subjected to a sigmoid activation function to obtain the color information. Experimental results show that the hardware architecture designed in the application can accelerate the neural radiation field calculation, and compared with the same algorithm running on a GPU, the running power consumption is far less than that of the GPU, and the generated image quality is similar to that of the GPU.

The present application relates to a sigmoid activation function circuit based on linear broken line fitting, comprising: a data processing module, a compensation module, a first addition module and a second addition module; the data processing module, the first addition module and the second addition module are connected in sequence, and the compensation module is connected with the first addition module; the data processing module inputs and processes input data and outputs to the first addition module; the compensation module inputs and processes input data and outputs corresponding offset to the first addition module; the first addition module accumulates the data output by the data processing module and the offset, and outputs the accumulated data; the second addition module processes the accumulated data to obtain an output value. The present application simplifies the circuit module and introduces the rounding calculation based on the broken line fitting method, effectively reduces the operation power consumption and propagation delay under the condition of ensuring the accuracy, and at the same time realizes smaller hardware resource consumption.

The application relates to an image classification method based on a quaternion Sigmoid activation function, belonging to the technical field of deep learning, which utilizes an image classification network based on a quaternion Sigmoid activation function to perform image classification on an image to be classified. The image classification network comprises a parallel structure layer; the parallel structure layer comprises a splicing layer, a quaternion Sigmoid activation function branch and a convolution branch; wherein the quaternion Sigmoid activation function branch and the convolution branch are in parallel; the quaternion Sigmoid activation function branch performs forward propagation calculation on a quaternion feature matrix by using a quaternion activation function, and then performs convolution operation; the convolution branch performs convolution operation on the quaternion feature matrix; and the splicing layer splices two outputs. The application utilizes the quaternion Sigmoid activation function and the parallel structure matched therewith, can follow the quaternion operation rules, thereby retaining the integrity of image data information and improving the accuracy of image classification.

The invention provides a false news detection method based on progressive field adaptive multi-view learning. The method comprises the following steps: step 1, constructing a multi-modal field adaptive data set; step 2, single-mode feature extraction based on pre-training knowledge enhancement; step 3, constructing a visual and language cross-modal comparison visual angle; step 4, domain-aware dynamic routing and feature fusion; 5, carrying out progressive course learning optimization training; and step 6, inputting the final news representation subjected to field enhancement into a full-connection classifier, calculating the authenticity probability of the news sample, and outputting a final dichotomy prediction result through a Sigmoid activation function. According to the method, general knowledge of a pre-training model is reserved through efficient fine tuning of parameters, heterogeneity characteristics of a specific field are effectively captured, adaptive fusion of multi-modal information in different fields is achieved through a dynamic routing mechanism, and robustness and accuracy of false news detection in a cross-field scene are remarkably improved.

The invention discloses a real-time unknown target detection method based on feature-base interaction and double-flow reasoning. A detection model used in the method comprises a CSPDarknet backbone network, a neck network and detection heads, the detection heads comprise a pair of heads and a pair of multiple heads, and feature-substrate interaction object type branches are embedded in the multiple heads; multi-scale feature maps of an input image are extracted through a CSPDarknet backbone network, the multi-scale feature maps are processed through a neck network, and the multi-scale feature maps output by the neck network are processed through a detection head; in the feature-base interaction object property branch, calculating the cosine similarity of each point and each base vector in the input feature map to obtain the original object property response quantity of the point; and the original object property response quantity of the point passes through a linear mapping layer and a Sigmoid activation function to generate an object property score of each point. The model can be efficiently migrated from a known category to an unknown category, so that the omission ratio of targets without category finding is remarkably reduced, and the excellent generalization performance in the open world is ensured.

The invention discloses an unsupervised single-frame fringe analysis method based on regularization phase tracking, which comprises the following steps of: projecting a high-frequency fringe pattern by using a projector, synchronously acquiring a three-dimensional cosine fringe image by using a binocular camera, and performing three-dimensional correction on the acquired original cosine fringe image by using calibration parameters of the binocular camera. Background items in the fringe images are removed through a variational image decomposition algorithm, and then a set of cosine fringe images without background and modulation and orthogonal fringe images of the cosine fringe images are obtained through Hilbert spiral transformation processing. Inputting the cosine fringe image into a fringe analysis network of a UNet structure, introducing a reference plane phase obtained in advance and an orthogonal fringe image to solve phase ambiguity in phase estimation, and obtaining an initial absolute phase by using luminosity consistency and phase smoothing constraint based on regularization phase tracking; at the same time, the luminosity consistency constraint and the improved sigmoid activation function are utilized to constrain the network output to the range to obtain a wrapped phase, and then the initial absolute phase is utilized to unwrap the wrapped phase to obtain the absolute phase of the left and right cameras, thereby realizing the single-frame three-dimensional morphology measurement.

The invention discloses a theoretical method for optimizing the performance of a vertical axis wind turbine, and belongs to the technical field of wind power generation. The method comprises the following steps: firstly, constructing a blade variable pitch control model based on wind direction dynamic adaptation, equivalently incorporating a wind direction angle into blade pitch angle calculation, simulating a jumping process of an optimal pitch angle under different azimuth angles by using a Sigmoid activation function, and generating a continuous and smooth adaptive pitch angle control function; therefore, the attack angle of the blade is adjusted in real time to be maintained within the optimal range. Secondly, a self-adaptive multi-flow-pipe model is provided on the basis of a double-multi-flow-pipe model, a wind flow field is subdivided into eight areas (equivalent to four parallel double-brake-disc multi-flow-pipe models), and interference factors and average power coefficients of all the areas are iteratively solved by combining the blade element theory and the momentum theorem; the dynamic stall phenomenon can be effectively restrained, the lift characteristic, torque output and operation stability of the wind driven generator in the full working condition range are remarkably improved, and the method is particularly suitable for a distributed wind power system in the urban environment.

This invention relates to the field of bioinformatics technology and discloses a method, apparatus, device, and medium for analyzing unspecified diabetes mellitus based on deep learning. The method involves acquiring an in vitro sample to be analyzed, which includes multiple protein expression features, including at least one important protein feature corresponding to unspecified diabetes mellitus. The in vitro sample is input into a trained deep learning model to obtain the real-valued logarithmic probability corresponding to the matching result between the in vitro sample and the binary classification result. A sigmoid activation function is used to map the matching probability value between the in vitro sample and the binary classification result. Based on the matching probability value, the prediction result for unspecified diabetes mellitus corresponding to the in vitro sample is determined. Based on deep learning technology, important protein features related to unspecified diabetes mellitus can be automatically identified from multiple protein expression features of the in vitro sample, resulting in a biologically interpretable deep learning model that significantly improves the accuracy and stability of model predictions.

The invention relates to the technical field of deep learning and recommendation systems, and particularly discloses a generative recommendation-oriented hybrid expert model sparse reasoning method and system, and the method comprises the steps: obtaining original input data, and obtaining an input vector through an embedded layer; carrying out normalization and maximum value selection operation on the input vector to obtain a mixed expert weight; on the basis of a hierarchical attention mechanism, according to the mixed expert weight, attention enhancement features are obtained through calculation; inputting the input vector and the attention enhancement feature into a hybrid expert model, and calculating to obtain a recommendation result; wherein on the basis of a parallelization gating mechanism activated by Sigmoid, according to attention enhancement characteristics, expert weights are obtained through calculation; and selecting the activated expert layer according to the expert weight, and calculating to obtain a recommendation result by using the input vector. According to the method, the recommendation precision equivalent to or even higher than that of an advanced dense model is achieved under the conditions of low time overhead and low calculation complexity.

The invention discloses an embryo transplantation outcome prediction method based on single blastocyst image data, belongs to the field of medical artificial intelligence, and is suitable for single blastocyst transplantation. Firstly, a high-resolution embryo image of a patient is obtained. The image resolution is unified by fixing the aspect ratio of an original image and properly cutting and filling. A standard picture enhancement technology is adopted for model training, and adaptability to variable imaging conditions is improved. During prediction, a preprocessed image is input into a model which takes ResNet-50 as a trunk and combines SimCLR comparative learning to perform self-supervised pre-training, the model is used for extracting morphological features of embryos, and the recognition capability of the model for specific tasks is further optimized through supervised fine tuning. The extracted features are averagely pooled and then connected with a newly defined full-connection network, then pregnancy probability is output through a Sigmoid activation function, and finally a prediction result of single blastocyst transplantation is obtained through threshold mapping. Besides, in order to enhance clinical interpretability, gradient weighted class activation mapping (Grad-CAM) is further introduced, and a key region concerned by the model is visualized so as to verify whether the model is judged based on structures with biological significance, such as an internal cell mass and a trophoderm. According to the method, the transplantation outcome can be predicted under the condition of only depending on a single blastocyst image, and the method can be used for assisting clinical embryo selection and decision making.

The application belongs to the technical field of convolutional neural network (CNN) in deep learning, and discloses an image classification method and system based on an improved light-weight channel attention mechanism, which inputs a feature map, compresses the spatial dimension into 1*1 through a global average pooling operation, retains the global information of the channel dimension, and obtains a feature map after global average pooling; the feature map after global average pooling is sequentially subjected to a full connection layer and a Sigmoid activation function operation, and a channel attention weight W att is obtained; a residual connection is introduced, a residual connection weight alpha is set, an input feature map is multiplied by the residual connection weight to obtain an output feature map F x . By introducing the channel attention mechanism, the model can better focus on important channel information in the input feature map, improve the distinguishing ability of feature expression and the generalization ability of the model, and improve the performance of the convolutional neural network in the image classification task.

The invention discloses a drug discovery method and system based on a multi-modal graph neural network and a storage medium in the technical field of artificial intelligence and bioinformatics, and the method comprises the following steps: S1, molecular graph coding: converting SMILES character strings of drug molecules into graph structure representation, and extracting topological structure features of the molecules through a graph convolutional neural network; s2, protein sequence coding: coding the amino acid sequence of the target protein into a feature vector of a fixed dimension through a bidirectional long-short-term memory network; s3, multi-modal fusion: splicing the molecular map features and the protein sequence features, and realizing cross-modal feature fusion through a multi-layer perceptron; s4, interaction prediction: based on the fused feature vector, outputting a probability value of drug-target interaction through a sigmoid activation function; the method is used for predicting the interaction between the drug and the target protein and realizing efficient drug screening and discovery.

The invention provides a junk mail detection method, electronic equipment, a chip, a storage medium and a program product. The method comprises the following steps: acquiring a to-be-detected mail, and extracting label type features and text type features of the to-be-detected mail; converting the label class features into embedding vectors with fixed dimensions, and generating binary values; utilizing a Skip-gram word embedding model to convert the text type features into a word embedding sequence; and fusing the embedded vector, the input scalar and the feature word embedded sequence by using a pre-constructed convolutional neural network to form a fused feature vector, predicting a probability value of a junk mail by using a Sigmoid activation function in combination with the fused feature vector, and judging that the mail is a junk mail if the probability value is greater than a preset threshold.

The application discloses a high-speed motor train unit idling and sliding trend prediction method, which comprises the following steps: step one, screening and standardizing the model input data; step two, obtaining a feature matrix containing position information through value encoding and position encoding; step three, inputting the feature matrix obtained in step two into a ProbSparse self-attention network to obtain the attention relationship of different feature vectors at different positions, and performing parallel calculation through a multi-head mode; step four, using a Mask self-attention network for the decoder to associate the sliding parameter information at different moments with the sliding state, and finally performing classification prediction on the future moment through a full connection layer and a Sigmoid activation function. The application can predict the idling and sliding trend of the train more than 1 second in advance, the prediction accuracy meets the train operation demand, the prediction speed is fast, and the idling and sliding trend states of multiple axles of the train can be simultaneously and on-line predicted in parallel.

A spectral image detection model, a detection method, and a camera processor capable of on-chip computing. The spectral image detection model comprises N KAM blocks connected in series and a classifier block connected after the N KAM blocks; each KAM block comprises a depthwise separable convolution layer, a batch normalization layer, an improved triple attention layer, and a ReLU activation function layer which are connected in sequence; the depthwise separable convolution layer is used for extracting image features; the improved triple attention layer is used for enhancing information interaction within and between feature maps; and the classifier block comprises a depthwise separable convolution layer, a batch normalization layer, an ordinary convolution layer, and a Sigmoid activation function layer which are connected in sequence. The spectral image detection model has the characteristics of light weight, low computational complexity, etc. and can be deployed in a camera processor in a camera, so as to meet the requirements for hyperspectral image processing at mobile devices and edge devices.

The invention discloses a risk early warning method in a multi-modal fusion ERCP operation. The method comprises the following steps: synchronizing multi-modal data; preprocessing the multi-modal data by adopting gray correction, Kalman filtering and other algorithms, and removing noise and abnormal data; image features are extracted through CNN, pressure time sequence features are extracted through TCN, and a fusion feature vector is output in combination with an attention fusion algorithm; inputting the fusion feature vector into a Transform dichotomy model, calculating a risk probability through a sigmoid activation function, judging a risk level based on a preset threshold value, and triggering early warning; feedback information is received, and the model is iteratively optimized through a parameter updating algorithm. The system comprises a corresponding data acquisition module, a preprocessing module, a feature fusion module, a risk early warning module, a feedback iteration module and a storage module. According to the method, the accuracy and timeliness of risk early warning in the ERCP operation are remarkably improved, the occurrence rate of complications is reduced, the operation safety is improved, and the workload of medical staff is relieved.