255 results about "Feature transform" patented technology

The present invention relates to the technical field of expression recognition, and in particular, to an expression recognition method and system based on multi-scale features and spatial attention. The method includes: performing feature extraction on acquired facial image data by using an HNFER neural network model to obtain an original input feature map; performing pooling and concatenation on extracted features based on a CoordAtt attention mechanism to obtain a feature map; performing deep convolution processing on the feature map to obtain an attention map, and then performing element-by-element multiplication to obtain a final feature map; and performing feature transformation and normalization on the final feature map to obtain an expression category probability and output the expression category probability. In the present invention, by integrating scale perception and spatial attention technologies, the model can recognize and classify different emotional states more accurately and maintain high performance even under complex environmental conditions.

According to the motor anomaly detection model training method and the motor anomaly detection method provided by the invention, a plurality of audio samples and category labels thereof during motor operation are acquired, a plurality of disturbance audios are generated through speed disturbance for each audio sample, the disturbance audios are input into a frequency domain feature extraction sub-model and a time domain feature extraction sub-model, and the frequency domain feature extraction sub-model and the time domain feature extraction sub-model are extracted. Performing linear transformation and weighted summation on the enhanced wavelet time-frequency spectrum and the enhanced time-domain features through a full connection layer to generate fusion enhanced features, performing feature transformation on the fusion enhanced features through a deep neural network, inputting the fusion enhanced features into a classifier to obtain a classification probability, and calculating triple loss according to a category label and the classification probability; and updating parameters of the frequency domain feature extraction sub-model, the time domain feature extraction sub-model, the full connection layer, the deep neural network and the classifier through loss back propagation and a stochastic gradient descent algorithm to obtain a trained motor anomaly detection model. Therefore, the accuracy of motor state recognition of the model in a high-noise environment is improved.

The invention discloses an accurate analysis method for Internet user behaviors based on artificial intelligence, and the method belongs to the technical field of Internet, and comprises the steps: collecting user behaviors and basic attribute information through a multi-source data collection technology, and integrating data through a data fusion algorithm after data cleaning and preprocessing; designing and extracting a plurality of behavior features, screening key features by using a feature selection algorithm, and performing feature transformation processing at the same time; constructing a deep learning model fusing a recurrent neural network (such as LSTM and GRU), a convolutional neural network and the like, training the model by using a large-scale data set, and optimizing parameters; establishing a real-time analysis system, inputting new data into the model in real time, providing personalized service and recommendation according to an analysis result, and collecting feedback data to update the model online; and finally, evaluating the model by adopting multiple evaluation indexes, verifying the generalization ability through methods such as cross validation and the like, and further optimizing and adjusting the model.

The invention discloses a hysteroscope image color feature transformation method based on deep learning, particularly relates to the field of medical image processing, solves the problem of insufficient analysis accuracy caused by color feature redundancy in a hysteroscope image, and ensures high-quality image data through multi-level accurate screening. Extracting multi-dimensional color features by using nonlinear mapping and structural analysis, and enhancing the expression and discrimination of the features; the constructed multi-level index system is combined with spectral analysis and graph theory operation, and the macroscopic distribution difference and the microcosmic robustness of the features are accurately measured; the method comprises the following steps: successfully extracting a key candidate dimension subset with high discrimination and robustness in a complex scene through a Pareto-like leading edge search and topological data analysis technology, multi-round iteration and graph theory clustering screening; therefore, the problems of color feature redundancy, insufficient stability and poor adaptability are effectively solved, and the accuracy and reliability of hysteroscope image analysis are remarkably improved.

The invention discloses a multi-sensor feature level fusion target detection method and system, and the method comprises the steps: continuously collecting spatial data information and image data information in a vehicle operation process; superposing first data information collected at a plurality of previous historical moments and the current moment, extracting first feature information based on the superposed first data information, and extracting second feature information based on second data information collected at the current moment; performing feature transformation on the first feature information and the second feature information to obtain a first feature aerial view and a second feature aerial view, and performing sampling, channel splicing and feature fusion on the first feature aerial view and the second feature aerial view to obtain a final fused feature aerial view; and performing analysis according to the final fusion feature aerial view to obtain a target detection result. According to the invention, the detection precision and the detection stability of a long-distance target can be improved.

The invention relates to the technical field of image processing, in particular to a multi-modal remote sensing image semantic segmentation method and system, and the method comprises the steps: building a data set which comprises a plurality of image pairs, and each image pair comprises an SAR image and an optical image; constructing a multi-modal semantic segmentation model, wherein the multi-modal semantic segmentation model comprises a basic feature transformation layer, a multi-scale semantic acquisition layer, a significant feature cross fusion layer, a feature deep coding layer and a feature decoding layer which are connected in sequence; training and optimizing a multi-modal semantic segmentation model based on the data set to obtain a target multi-modal semantic segmentation model; and performing joint semantic segmentation on the SAR remote sensing image and the optical remote sensing image by adopting the target multi-mode semantic segmentation model to obtain a semantic segmentation result graph. The embodiment of the invention is at least beneficial to reducing the parameter quantity and calculation quantity of semantic segmentation of the multi-modal remote sensing image.

The invention relates to the technical field of augmented reality, in particular to an AR real-time scene reconstruction and illumination matching system and method based on neural rendering, and the system comprises a real-time scene understanding module, a real-time neural radiation field module and a real-time reflection inference and enhancement module. Comprising a forward network, a feature transformation layer and a backward network, the forward network maps a high-dimensional rendering space to a feature manifold, the feature transformation layer executes mapping from the feature manifold to a rendering manifold, and the backward network performs rendering calculation based on a Riemannian metric driven adaptive ray tracing mechanism; the real-time reflection inference and enhancement module performs material prediction and illumination information calculation based on an illumination dynamic adaptation system on a differential manifold, real-time application of a neural radiation field technology in an AR scene is realized through an innovative differential geometry framework, and a real-time rendering frame rate of 30-60 fps is achieved while a high-quality rendering effect is kept.

The invention belongs to the technical field of image processing, and particularly relates to a lightweight unmanned aerial vehicle target tracking method based on separable convolution. The method comprises the following steps: step 1, preparing two remote sensing image data sets which are respectively used for training and testing; 2, inputting a frame into a separable convolution block for feature extraction to obtain a feature sequence; 3, inputting the feature sequence into an inverted bottleneck block and a forward feedback network, carrying out feature transformation and information mixing, and enhancing the feature expression capability; 4, the extracted category number is input into a fusion module to be processed, and a fusion sequence is obtained; and step 5, obtaining a classification regression vector through loss calculation, and then outputting a result graph. According to the method, a lightweight remote sensing target tracking network architecture is innovated, an improved feature extraction module UIB-P and a fusion module ICA are added, and an innovative loss function is adopted, so that the model greatly improves the global representation capability and precision of target tracking while reducing the calculation amount.

The present invention relates to the field of pipeline leakage monitoring. Disclosed is an oil and gas pipeline leakage wave identification and monitoring system. In the present invention, an mCNN is combined with LFLBs for performing feature extraction on an acoustic wave signal collected by a DFB, and the collected data improves information completeness; a three-way parallel one-dimensional CNN used in the present invention exhibits good temporal resolution and sensitivity to high-frequency feature transformations in signals; and the present invention integrates advantages of different scales, enabling the algorithm to learn more features, and incorporating the LFLBs to further extract high-level local features. An mCNN-LFLBs network model of the present invention exhibits significant innovation and advancement on the technical level, and also demonstrates extremely high value in actual application. The network model not only provides a novel and efficient technical means for critical fields such as natural gas pipeline inspection, but also introduces new ideas and methods to research fields related to deep learning and signal processing.

The invention discloses an image defogging method based on improved DEA-Net. The method comprises the following steps: (1) obtaining real image foggy data of an open source website and a corresponding clear image as a training set and a test set; (2) the training set images are cut and then preprocessed; (3) constructing a deep learning model, wherein the model comprises a coding part, a feature transformation part and a decoding part; (4) inputting training image data in a training set into the created network model, and training by using a GPU (Graphic Processing Unit) to obtain model parameters with optimal accuracy; (5) taking the test image as input to obtain a defogged image; according to the method, the detail texture features of the original image can be better reserved in the defogging process, and distortion caused by model defogging is reduced. Meanwhile, the structure of the model is simplified, and model complexity and training time are reduced.

The invention discloses a meteorological prediction method and system for multi-resolution mixed data, and the method comprises the steps: carrying out the preprocessing of the obtained multi-resolution mixed data, obtaining standard data, carrying out the pyramid grading of the standard data according to the resolution, and obtaining first pyramid data, second pyramid data and third pyramid data, the method comprises the following steps: obtaining a plurality of feature values, respectively inputting the feature values into a corresponding first feature transformation network, a second feature transformation network and a third feature transformation network to obtain a first feature value, a second feature value and a third feature value, inputting all the feature values into a structure aggregation network to obtain a fourth feature value, and inputting the feature values into a multi-scale convolutional neural network to obtain a prediction result. According to the method, a hierarchical pyramid is constructed, the feature extraction precision and the prediction accuracy are improved, deep data fusion and potential association mining are performed through feature transformation and a structure aggregation network, and accurate meteorological prediction is realized through a multi-scale convolutional network and an adaptive fusion mechanism.

The invention relates to the technical field of computer vision and deep learning, and discloses an image classification system and method fusing a space attention mechanism and long and short-term memory network sequence modeling, which combines the feature extraction capability of a convolutional neural network and space attention and time sequence attention mechanisms. And processing the spatial position sequence by using a long-short-term memory network. Firstly, advanced spatial features of an image are extracted through a feature extraction module by adopting a pre-trained convolutional neural network with a frozen weight, then a spatial attention module is introduced, a spatial attention graph is generated through channel dimension statistics, and important region features are enhanced. After the spatial attention is weighted, a dual-path feature is utilized, one path enters a feature transformation module, and a convolutional layer is used for reducing dimensionality and enhancing feature expression ability. Then, a sequence modeling module is carried out, the spatial features are flattened into a position sequence, a spatial position dependency relationship is modeled by adopting a bidirectional long-short-term memory network, and learnable attention vector dynamic aggregation key position features are introduced; the other path retains spatial global features. And outputting the dual-path features to a feature fusion module, extracting spatial global features and sequence aggregation features in parallel, and designing a gating mechanism to adaptively fuse the dual-path features. And finally, entering a classification module, and realizing end-to-end image classification based on fusion features.

The invention relates to an embankment slope stability assessment method based on artificial intelligence, and belongs to the technical field of embankment slope monitoring and assessment. The method comprises the following steps: acquiring and marking embankment slope stress sensing data; dividing the data into a plurality of spatio-temporal data blocks; constructing a stability evaluation model; extracting multi-scale convolution features by adopting multi-scale cavity convolution of a high-frequency channel and pooling-deconvolution operation of a low-frequency channel to obtain a fused multi-scale feature matrix; a hidden state sequence is obtained through a space attention mechanism and a double-door-setting mechanism; calculating time interval saliency based on the hidden state vector, then calculating a weighted feature vector, further obtaining a weighted feature matrix, and processing through deep convolution and point-by-point convolution to obtain a pooling feature vector; carrying out stability evaluation grade classification through learnable category prototype and gating feature transformation; and dynamically adjusting sample weight and constraint attention distribution by adopting a total loss function. According to the method, the progressive instability identification capability can be improved.

The embodiment of the invention provides an image decoding method and device. The method comprises the following steps: firstly, receiving a code stream; then, acquiring first image information based on the code stream; performing feature transformation on the first image information to obtain features of the first image information; then, inputting the second image information into a diffusion model to obtain a target feature; wherein the features of the first image information are used for adjusting intermediate features generated in the feature generation processing process of the diffusion model, and the second image information comprises one of a preset noise image, an initial reconstruction image determined based on a code stream or an image obtained by adding noise to the initial reconstruction image; and then, performing reconstruction processing on the target features to obtain a target reconstruction image. In this way, the subjective quality of the reconstructed image can be improved.

The invention provides a network intrusion detection method based on depth time sequence clustering. The method comprises the following steps: performing dimensionality reduction and reconstruction on high-dimensional data by using a convolutional automatic encoder CAE; the self-supervised learning module performs feature transformation on the data, the training model classifies and identifies the transformation type of a sample as an auxiliary task, and learns generalized feature representation; global and local memory modules learn universal and specific features of data; the adaptive fusion module fuses the features to obtain a new feature representation for reconstruction; calculating the most suitable clustering feature of the positive sample data set; and identifying data different from the positive sample data set in the low-dimensional space by using a BIRCH clustering algorithm to realize anomaly detection. The self-supervised learning module is introduced to avoid the problem of data source limitation, the global and local memory modules are utilized to avoid the problems of poor model generalization ability and feature expression limitation, the BIRCH clustering is utilized to avoid the problem that the quantity of k needs to be determined in the traditional K-MEANS algorithm, and the method has higher efficiency.

The invention discloses an electromagnetic field intelligent calculation method based on deep learning, and the method specifically comprises the steps: inputting a space-time input vector into an MFF-PINN neural network, the MFF-PINN neural network comprises parallel sub-networks and a linear superposition module, the sub-network comprises a scale transformation module, a Fourier feature transformation module and an MLP processing module, and the MFF-PINN neural network comprises a linear superposition module; firstly, scale transformation is carried out on a space-time input vector, then Fourier feature transformation is carried out on the vector after scale transformation, and the Fourier feature transformation module carries out Fourier transformation on the vector after scale transformation based on an effective frequency matrix; all the sub-networks share the Fourier feature transformation, and the output obtained by the Fourier feature transformation is input to the MLP processing module in the first sub-network; and carrying out linear superposition on the output of the sub-networks. According to the invention, the expression capability of the network on the high-frequency component and multi-scale characteristics of the electromagnetic field is obviously enhanced.

The invention relates to a construction method of an adaptive temperature guided mixed attention network for hyperspectral image classification, and provides a novel mixed attention network for hyperspectral image classification. A deterministic-probabilistic mixed attention mechanism is designed, and a traditional single-mode attention mechanism is expanded into a double-branch structure. The deterministic branch generates stable feature weights through explicit feature transformation and normalization, and the probabilistic branch introduces a soft attention mechanism of temperature parameter adjustment to realize flexible feature selection. The robustness and adaptability of feature selection are remarkably improved through the complementary effect of the two mechanisms. Meanwhile, a multi-level feature fusion strategy based on a learnable fusion coefficient is innovatively designed, and adaptive adjustment of feature importance is realized by dynamically balancing contributions of two types of attention.

The invention provides a 3D digital human and scene processing method, device and equipment, a medium and a product, and the method comprises the steps: obtaining 3D data based on 3D Gaussian sputtering 3DGS; and performing representation generation on the 3D data, specifically, processing the 3D data to obtain a reference point position and a reference point attribute of each reference point, and performing feature transformation based on the reference point position and the reference point attribute in combination with a Gaussian ball position correspondingly generated by the reference point, obtaining an attribute context feature of each datum point; quantizing the datum point attribute of the datum point by using the attribute context feature; and performing entropy coding on the quantized reference point attribute according to the attribute context feature to obtain a coding result.

The invention discloses a ship identification method based on multiple acoustic feature fusion. The ship identification method comprises the following steps: acquiring and preprocessing underwater noise signals of a ship; carrying out MEL frequency spectrum, CQT frequency spectrum and STFT phase diagram feature transformation on the preprocessed ship underwater noise signals, and carrying out stacking on channel dimensions to obtain three-channel feature images; constructing a recognition neural network; performing feature extraction on the three-channel feature image through an asymmetric convolution feature extraction module to obtain a preliminary feature; strengthening the preliminary features through a feature strengthening module to obtain strengthened features; and performing classification identification on the reinforced features through a classification module to obtain an identification result. Through multiple acoustic feature extraction, deep information in original data is mined and utilized to the greatest extent while the calculation cost is reduced, the noise robustness is excellent, the model structure is simple, and the recognition effect can be effectively improved.

The invention discloses a ground feature classification method, system and device based on time-space spectrum feature collaboration, and the method comprises the steps: obtaining a time sequence remote sensing image set, and carrying out the processing of the time sequence remote sensing image set through a graph attention model, and obtaining key time spectrum features; processing the time sequence remote sensing image set and the key time spectrum features through a space attention model to obtain space features; converting the key time spectrum features and the spatial features through a feature transformation network to obtain space-time spectrum three-dimensional features; constructing a ground feature classification pre-training model, and training the ground feature classification pre-training model based on the space-time spectrum three-dimensional features to obtain a ground feature classification model; and reasoning the time sequence remote sensing image of the to-be-measured target based on the ground feature classification model to obtain a ground feature classification result. According to the method, high-precision classification based on time-space spectrum multi-dimensional features is realized, and the ground feature classification model not only can realize high-precision classification on a limited data set, but also has relatively strong generalization ability, and can be suitable for ground feature classification tasks in different scenes.

The invention relates to the technical field of intelligent control and reinforcement learning technologies, in particular to a trajectory tracking control method fusing multi-objective optimization and a physical sensing network, which comprises the following steps: acquiring a real-time state vector of a to-be-controlled object, decomposing the acquired state vector into a sphere dynamic flow and a platform attitude flow, coding features of different attitude flows are extracted, a fusion feature vector is constructed, and at the same time, an attention mechanism is used to carry out feature transformation to determine control decision features; and for the determined control decision features, utilizing a multi-objective optimization function to carry out cooperative constraint on the generated actions, carrying out feature training in combination with an experience playback mechanism and a self-adaptive stable learning mechanism, and after training is completed, determining a trajectory tracking control instruction of the object to be controlled through dynamic adjustment of learning parameters. According to the invention, by establishing an adaptive stable learning mechanism, the learning rate and exploration noise are dynamically adjusted based on performance stagnation detection, and the training stability and convergence speed are improved.

The invention discloses a personnel state detection system and method based on a wifi router, and the method comprises the following steps: S1, deploying the wifi router, collecting subcarrier CSI which changes due to the disturbance of the movement of personnel, and forming a CSI signal sequence; s2, configuring a neural network classifier and loading parameters; s3, preprocessing the CSI signal sequence to obtain a stable CSI time sequence signal; s4, constructing a graph structure taking the subcarriers as graph nodes, and generating a spectrogram; s5, performing Laplace feature transformation on the spectrogram, and extracting a frequency domain feature vector of the spectrogram; s6, constructing a multi-scale Sheng differential equation model, and generating a state trajectory under each time scale; and S7, splicing the state tracks to form a state evolution sequence, inputting the state evolution sequence into a neural network classifier, and generating a state recognition result. According to the method, CSI is used for modeling personnel disturbance, multi-scale dynamic representation and accurate recognition are achieved, and the method is suitable for a non-contact personnel state sensing scene.

The invention discloses a graph classification method based on a spiking neural network, and the method comprises the steps: carrying out the aggregation of the node features of a graph G (V, E) through a graph convolution network, and obtaining the embedded representation of nodes; performing dimension conversion on the embedded representation through linear projection, and inputting the converted node representation X into a standard Transform layer; performing weight screening and sparsification through a sparse weight selection module to generate queries, keys and values; and carrying out dimension expansion and information enhancement on the Q and the K by utilizing a Kronecker product, and calculating the element similarity between the Q and the K through a Hadamard product. And sending the result to a pulse sequence generated by the leakage integral-issuing neurons, averaging the pulse sequences of a plurality of time steps to obtain an attention score, processing the attention score to obtain attention output, performing feature transformation by using a feedforward network to obtain final node representation, and finally converting Z into a classification result by using a classification layer. According to the method, the attention mechanism is combined with the pulse neural network, so that the expression ability of the attention mechanism can be kept, the utilization of computing resources is further optimized through the sparsity and discreteness of pulse signals, the parameter quantity in the model is remarkably reduced, and the accuracy of graph classification is improved.

Provided in the present disclosure are a digital-human processing method and apparatus, and a device, a medium and a product. The method comprises: performing semantic analysis processing and / or representation generation processing on 3D data which is represented on the basis of a neural radiance field; and performing feature transformation, quantization processing and encoding processing on processed data.

The invention discloses an attribute feature driven scene-to-image cross-domain mapping sample set construction method, and belongs to the technical field of depth model compression, and the method comprises the steps: carrying out the augmentation of a basic sample set through employing an image feature transformation method, and obtaining a material library used for sample synthesis; the basic sample set comprises a simulation target material and a real shooting environment material; performing spatial position registration on an actual target and an actual environment in an actual application scene to obtain position information of the actual target in an environment coordinate system; selecting a corresponding simulation target material and a real shooting environment material from a material library based on an actual application scene; based on the position information of an actual target in an environment coordinate system, performing threshold segmentation on a selected simulation target material and fusing the selected simulation target material with a real shot environment material to obtain a high-confidence synthetic sample, so as to solve the problems of obvious segmentation of pixel blocks in a synthetic sample region, sample distortion and high reliability of the existing sample synthesis method. And therefore, a high-confidence sample set required by intelligent algorithm training cannot be effectively synthesized.

The embodiment of the invention provides a question and answer data processing method and device and a storage medium. In the question and answer data processing method, after a reply text generated by a target language model for an inquiry instruction is obtained, a target scoring model can be adopted to score question and answer pairs formed by the inquiry instruction and the reply text. Wherein a base model of the target language model is a pre-trained generative large language model, and a feature transformation module in the target scoring model is obtained by fine tuning according to an encoder and / or a decoder in the pre-trained generative large language model. In the embodiment, partial model structures and model parameters of the target scoring model and the target language model are from the same base model, so that knowledge migration and knowledge sharing can be realized on the basis of the structures and the parameters of the same base model, the scoring performance of the target scoring model is improved, and the scoring result is more reliable. And on the other hand, the scoring model of the target language model is generated by utilizing the base model of the target language model, so that the existing model structure and model parameters can be fully reused, and the construction cost of the target scoring model is reduced.

The invention provides a machine vision coding method based on feature distillation, and relates to the technical field of image processing.The method comprises the steps that an image to be processed is input into a machine vision coding model, and the model extracts first potential feature representation of multiple channels through an analysis encoder; the method comprises the following steps: quantitatively dividing into basic layer quantitative features containing semantic and spatial structure features and enhancement layer quantitative features containing detail and texture features; the hyper-priori correlation module encodes hyper-priori information and generates enhanced auxiliary features and basic auxiliary features, and the conditional entropy coding network realizes encoding and decoding of the basic layer quantization features and the enhanced layer quantization features based on the enhanced auxiliary features and the basic auxiliary features. A machine vision task result is obtained through a feature transformation and task processing module, and after splicing is conducted through a splicer, a reconstructed image is output through a synthesis decoder. According to the invention, both machine vision and image reconstruction can be considered.

The invention discloses a road snow-blown visibility recognition method and system based on image recognition, and the method comprises the steps: carrying out the sequential processing of an original snow-blown image through initial feature extraction, multi-level down-sampling and feature enhancement, and multi-scale context fusion, and directly outputting a visibility interval and confidence, according to the method, downsampling, feature transformation and feature enhancement processing processes are repeatedly executed for multiple times, deep feature maps with gradually reduced scales are sequentially obtained, automatic and objective recognition of the visibility of the whole road line blown snow is achieved, the adaptability to the scene of the blown snow which is high in burstiness and non-uniform in space is improved, and the visibility of the whole road line blown snow is improved. Therefore, the model can more accurately capture the key depth of field and texture degradation characteristics which influence the visibility, and the recognition result is more accurate.

The invention provides a video moment retrieval method and device, electronic equipment and a storage medium, relates to the technical field of artificial intelligence, and is suitable for the fields of financial science and technology and medical health. The method comprises the following steps: performing space-time coding on a target video to obtain an initial space-time feature map; performing feature transformation on the initial spatial-temporal feature map through a first hierarchical adaptive granularity converter to obtain a first hierarchical spatial-temporal feature map; performing feature transformation on the first-level spatial-temporal feature map through a second-level adaptive granularity converter to obtain a second-level spatial-temporal feature map; performing feature fusion on the first-level spatio-temporal feature map and the second-level spatio-temporal feature map to obtain multi-scale spatio-temporal features; retrieval demand information is obtained, and then text coding is carried out to obtain retrieval demand information features; and carrying out feature decoding on the retrieval demand information features and the multi-scale spatial-temporal features, and then carrying out moment prediction to obtain starting and ending moments of the target fragment. The video time retrieval efficiency can be improved.

The application discloses an aerial view generation method based on multi-scale feature transformation and time sequence context, and relates to the technical field of map generation. The method fully utilizes the complementarity of image and laser radar data by fusing the image and the laser radar data, improves the accuracy and robustness of aerial view generation, and can still remain stable under bad weather; a multi-scale space conversion module extracts different scale features, enhances the feature expression capability, and makes the aerial view clearer and more accurate; time sequence information is introduced, past time features are used to enhance current features, and dynamic perception capability is improved; an advanced backbone network and a feature alignment and fusion module are adopted, so that high efficiency and flexibility are ensured; and specific network structures such as Swin-T, PointPillars and random inactivation layers are applied, so that the accuracy and generalization capability are further improved.