65results about How to "Improved loss function" patented technology

The invention discloses a remote sensing image scene classification method based on multi-branch convolutional neural network fusion. The method comprises the following steps: firstly, randomly dividing a scene data set into a training set and a test set in proportion; performing preprocessing and data amplification on the data set; obtaining an object mask graph and an attention graph according to the processed data through an object detection network and an attention network respectively; respectively inputting the original image, the object mask image and the attention image training set into a CNN network for fine adjustment; obtaining three groups of test sets, respectively obtaining optimal classification models, respectively obtaining outputs of Softmax layers through the optimal classification models by taking the three groups of test sets as inputs, and finally obtaining a final prediction result by fusing the outputs of the three groups of Softmax layers through a decision-making level. The classification accuracy and the classification effect can be improved.

The invention relates to a three-dimensional liver image semantic segmentation method based on a context attention strategy. The method comprises the following steps: selecting a medical image data set to be subjected to liver segmentation, and dividing the medical image data set into a training set and a test set; preprocessing the three-dimensional liver image in the training set; in the encoding stage, obtaining a feature map of the liver by using a residual structure, a convolutional network and cavity convolution; in a decoding stage, obtaining a segmented image of the liver by using a context attention strategy module, transpose convolution and a depth supervision mechanism; and carrying out post-processing on the liver image obtained after semantic segmentation. The method has the characteristic of improving the semantic segmentation effect of the three-dimensional liver image, realizes a better automatic segmentation effect, and can assist doctors in diagnosis.

The invention discloses a traffic sign detection algorithm based on an improved YOLO v3 algorithm. According to the method, a feature extraction network capable of maintaining high-resolution representation is designed to replace DarkNet-53 in an original YOLO v3 algorithm, so that the detection precision of a small-size target traffic sign is improved, and the parameter quantity of the algorithmis reduced; attention of the detection algorithm to small and medium-sized targets is increased by fusing the feature maps participating in prediction; and optimizing the loss function by using a GIOUalgorithm and a local loss algorithm. According to the method, the detection accuracy of the small-size traffic sign is improved, and the traffic sign can be rapidly and accurately detected and identified on a complex traffic road.

The invention discloses a network traffic anomaly detection method based on a small amount of annotation data, and the method comprises the steps: carrying out the dimension reduction of a feature vector through employing double auto-encoders, and then carrying out the supervised training through employing a deep neural network; dividing the network traffic into two types of positive samples and negative samples, and finally screening out a part of important samples in unlabeled data and submitting the samples to experts for labeling, increasing the number of labeled samples, iteratively updating an auto-encoder and a classifier, and then employing the trained classifier for detecting network traffic abnormality. According to the invention, a double-auto-encoder architecture is proposed, pure positive and negative samples are used for respectively training the auto-encoders, and the stability of the classifier is improved. Meanwhile, the loss function of the deep neural network is improved, the sample weight is adjusted in a finer-grained manner, the problem of overfitting caused by imbalance of positive and negative samples and small training sets is solved, a new method for calculating the marking value of the unmarked data is provided, the samples with high marking value are selected to be delivered to experts, and the marking cost is reduced.

The invention discloses a liver image semantic segmentation method based on an edge attention strategy, and the method comprises the steps: S1, carrying out the format conversion of an original data set; s2, preprocessing the liver image data; s3, analyzing edge features of the liver image and designing an Encoder-Decoder deep learning framework model, wherein the Encoder-Decoder deep learning framework model comprises an encoding stage and a decoding stage; s4, in a coding stage, obtaining an edge attention graph by utilizing a ResNet34 residual network and an edge attention strategy module;s5, in a decoding stage, performing feature extraction and segmented image generation by utilizing a deconvolution and edge attention strategy module; and S6, performing noise reduction processing onthe image after semantic segmentation. Through training and fusion of the edge attention strategy module and the deep convolutional network, the loss function is optimized, the depth features of the liver image are extracted and segmented, and the method has the characteristic of improving the semantic segmentation effect of the liver image.

The invention relates to a compression method and system for a deep convolutional neural network, and the method comprises the steps: an unimportant filter in a to-be-compressed deep convolutional neural network is determined according to a filter importance selection mode and / or a model compression rate; progressive sparse constraints are applied to the unimportant filters and added into a loss function of network training as regular terms, and an optimized loss function is obtained; according to the regular term, a threshold iterative algorithm and a back propagation algorithm are adopted for combined solving, and updating parameters of the to-be-compressed deep convolutional neural network are obtained; a convolutional neural network model with a filter sparse form based is obtained onthe optimization loss function and the updating parameters; and the convolutional neural network model is pruned with the filter sparse form by using a structured pruning algorithm to obtain a compressed convolutional neural network model with relatively high network precision.

The invention belongs to the technical field of multimedia information security, and particularly relates to a video object removal tampering time-space domain positioning method based on deep learning, which comprises the following steps: S1, training a time domain positioning model and a space domain positioning model; S2, inputting a to-be-tested video into the time domain positioning model toobtain a tampered frame sequence; and S3, inputting the tampered frame sequence into the spatial domain positioning model to obtain a tampered area positioning result in the tampered frame. Accordingto the method, the tampered frames can be detected from the tampered video, and the tampered area can be positioned in each tampered frame.

The invention relates to an image content automatic description method constructed by a Chinese visual vocabulary. The method comprises the following steps of: a, performing word segmentation processing on a plurality of description sentences corresponding to a single picture by using a Chinese word segmentation tool, selectively reserving nouns, verbs and adjectives in a word list according to statistical word frequencies, and forming a Chinese visual vocabulary by using the reserved words; b, predicting the Chinese visual vocabulary based on a Chinese vocabulary prediction network to obtainimage annotation information; c, based on the automatic image description model, extracting image convolution features by using an encoder, and decoding the image convolution features as initial inputinto Chinese description statements by using a decoder; image annotation information can be obtained by predicting the image vocabulary through the vocabulary prediction network, a residual structureis added to the Chinese visual vocabulary prediction network, and the problem that the number of layers of the Chinese visual vocabulary prediction network is deepened can be effectively solved.

The invention discloses an electromechanical equipment state data completion and prediction method based on multi-scale sampling, and relates to the technical field of data processing. The electromechanical equipment state data completion and prediction method based on multi-scale sampling comprises the following steps: S1, acquiring working condition data in a working process of electromechanicalequipment by adopting an intelligent sensor, and constructing a working condition data set D. According to the electromechanical equipment state data completion and prediction method based on multi-scale sampling, a multi-scale sampling thought is adopted to extract a plurality of time sequences from the data set, feature learning is carried out from different time scales; the prediction precision and stability are improved through a voting scoring strategy, meanwhile, a generative adversarial network is adopted, sample completion is conducted on a data set, a loss function of the generativeadversarial network is improved, training stability and efficiency are improved, completed samples are selected based on the voting scoring strategy, and low-quality generated samples are removed.

The invention discloses a CNN hyperspectral image classification method based on edge preserving filtering, and in order to enhance the classification effect, edge preserving filtering is adopted to carry out spatial feature extraction. Meanwhile, in order to solve the problem of unbalanced features of the classified samples, the original loss function is changed, the classification penalty of a small number of samples is increased, and the classification effect is further improved.

The invention discloses an artificial intelligence-based composing method and system, a computer device and a storage medium, wherein the method comprises the following steps: obtaining note information, wherein the note information comprises a playing start time, a playing duration time and a pitch value of each note; creating a music time sequence, and performing autoregressive integral moving average ARIMA model prediction to obtain a time sequence prediction result; constructing a music prediction model according to the note information and the time sequence prediction result; determininga topological structure of a music prediction model, and obtaining and realizing automatic composition predicted music, according to the note information and the time sequence prediction result, combining the music prediction model and the determined topological structure. The method realizes automatic composition, greatly simplifies the production process, allows users to participate in the creation, generates different tracks through different inputs, does not require midway intervention, and is flexible and rich.

The invention discloses an incomplete codeword recognition method for a gas meter. On the basis of common features of a complete codeword and an incomplete codeword of a gas meter, a loss function ofa network model based on a convolutional neural network is improved to achieve an objective of training and identifying an incomplete codeword network model of the gas meter by using the complete codeword of the gas meter. Therefore, a problem that is caused by insufficient training of the deep learning model because of insufficient data volume of the incomplete codewords of the gas meter is solved completely.

The invention discloses an AI thermal imaging all-weather intelligent monitoring method, which comprises a control module and comprises the following steps: S1, extracting the working time of a systemby the control module, and starting a high-definition camera or a thermal imager according to a preset time threshold section H; S2, the control module obtains thermal imaging or high-definition camera video data, and transcodes thermal imaging AV format video output into video images; S3, performing motion detection on the video image by using a motion detection algorithm; Setting a remarkable motion detection threshold, and judging whether the current motion is remarkable or not, so as to filter out false detection caused by noise and light and shadow changes; Meanwhile, the detected movingtarget is subjected to screenshot; And S4, performing interested target segmentation on the detected video image with significant motion by using an instance segmentation algorithm to determine whether the target causing the current motion is a monitoring target interested in the system, and entering the next step if the segmentation result is the interested target.

The invention provides a fan multi-fault diagnosis method based on a deep metric network difficult for sample mining, and relates to the technical field of wind turbine fault diagnosis. The method comprises the following steps: firstly, constructing a difficult sample data set for z different types of SCADA data of a wind turbine; subjecting the constructed difficult sample data set to imaging processing, optimizing various SCADA data variables so that arrangement of the data variables is made to be continuous in time and space, finally acquiring optimized training samples, and creating a depth measurement network model based on an improved triple training sample set loss function; and finally, taking a feature vector obtained by mapping the triple sample through a depth measurement network as a training set and inputting the training set into an SVM model to carry out wind turbine fault diagnosis. Sample generation is carried out on the basis of a triple model so that a method for diagnosing multiple faults by using a single model is provided, and an improved triple loss function training model is used, so that the accuracy of multi-fault diagnosis is improved.

The embodiment of the invention provides an image processing method and device and a computer readable storage medium. The method comprises the steps that an image processing request of a client is received, the image processing request comprises projection data to be processed, and the image processing request is used for requesting to reconstruct an image according to the projection data to be processed; domain transformation processing is carried out on the to-be-processed projection data to obtain image data; an image processing model is called to perform image reconstruction processing onthe image data to obtain an image result, the image processing model being obtained by training the image processing model according to the image sample data and the target sample data; and image results are sent to a client. By adopting the embodiment of the invention, the imaging quality of positron emission tomography is improved while the injection dosage of the tracer agent is reduced and the radiation is reduced, and the diagnosis and screening of diseases are facilitated.

The invention discloses a pavement crack detection method based on a pseudo twin dense connection attention mechanism. The pavement crack detection method comprises the following steps: S1, acquiring a data set; S2, preprocessing the pictures in the training set; s3, constructing a pseudo twin residual network; s4, designing a loss function of the pseudo twin residual network, training the pseudo twin residual network until the loss function converges, and storing the model; and S5, detecting the crack of the picture in the test by using the model obtained in the step S4. According to the method, a traditional Encoder-Decoder model is improved, so that a detection result under a mixed background, namely a mixed data set, can be effectively detected; and the loss function is optimized, so that the method is more suitable for a pavement crack background.

The invention belongs to the technical field of underground coal mine conveying belt detection, and discloses a method for detecting a foreign matter target of a underground coal mine conveying belt, system, equipment and a terminal, and the method comprises the steps: making a training sample and a test sample through an underground coal mine conveying belt monitoring video; intercepting and labeling foreign matters from the obtained video by using a Labelmg labeling tool, and carrying out adaptive histogram equalization on the conveyor belt image; under a YOLOv5 algorithm framework, introducing an attention model CBAM, simplifying network parameters by using depth separable convolution, optimizing a loss function, and constructing a detection model. According to the method, the foreign matters on the conveying belt can be detected under the conditions of coal dust interference, non-uniform illumination and high-speed movement of the conveying belt, so that higher detection precision is achieved, and the real-time requirement can be well met. Experimental results show that the accuracy and the recall rate of the foreign matter target detection algorithm for the underground coal mine conveying belt are the highest, and meanwhile the algorithm keeps good real-time performance.

The embodiment of the invention relates to the field of artificial intelligence, and provides an image main target detection method and device based on a convolutional neural network. The method comprises the steps of obtaining an input image; carrying out convolution layer processing on the input image to obtain a plurality of feature maps, and outputting the feature maps through a feature channel; according to the thermodynamic diagram of each feature channel, obtaining a saliency region on the feature diagram; carrying out clustering processing on the salient regions on the feature map to form a plurality of feature clusters; and determining the position of a main target on the input image based on the positions of all the feature clusters. According to the technical scheme, the problems that in the prior art, during model training, a specific target of an image needs to be manually labeled, and a trained model is only suitable for the specific target are solved.

The invention discloses an image retrieval method based on a depth feature consistent Hash algorithm. The method comprises the following steps: acquiring multi-label or single-label image data including a training set and a test set; preprocessing the training set; optimizing the neural network by using the preprocessed training set; inputting the training set into the optimized neural network to obtain a hash code; and calculating Hamming distances between the Hash codes obtained by calculation and the Hash codes obtained by the test set, sorting according to the distances from small to large, and outputting the first k retrieval results to finish retrieval. Through verification, the model provided by the invention has better retrieval performance than other existing baseline methods. In the retrieval of single-label and multi-label image data sets, compared with the existing common method, the method disclosed by the invention has obvious advantages in retrieval precision and time.

The embodiment of the invention provides a pedestrian re-identification method and system based on unsupervised learning. The method comprises the following steps: firstly, obtaining two to-be-identified video frames containing a plurality of pedestrians; and inputting the two to-be-identified video frames into the pedestrian re-identification model, and determining whether the two to-be-identified video frames contain the same pedestrian or not by the pedestrian re-identification model. A pedestrian re-identification model adopted in the embodiment of the invention is constructed based on a deep convolutional neural network, and when the pedestrian re-identification model is trained, a pedestrian cyclic distribution matrix between two sample video frames containing a plurality of pedestrians is determined, and an optimization loss function is determined based on the cyclic distribution matrix. In the whole training process, no extra algorithm module or indirect supervision signal is needed, such as a pedestrian tracking module or a clustering algorithm, pedestrian features can be directly learned from the unlabeled sample video frame, pedestrian re-identification is realized, the whole training process of the pedestrian re-identification model is simplified, and the accuracy of pedestrian re-identification is higher.

The invention discloses a partial discharge network training method and device for phase discrimination of power equipment. The partial discharge network training method comprises the steps: collecting a partial discharge spectrum of phase discrimination of a partial discharge measurement signal, forming an original detection sample set of the partial discharge of the power equipment, and carryingout preprocessing on an original detection sample; adopting a whitening mechanism to reprocess the preprocessed detection samples, and inputting a part of the detection samples as training data intoa neural network input layer; training the neural network, and optimizing the output of the neural network according to the loss function of the neural network; and predicting the partial discharge fault classification of the power equipment by using the rest of the detection samples as test data. According to the partial discharge network training method, firstly, a sample is subjected to whitening mechanism processing, so that the dimension of the sample is reduced, redundant data of the sample are removed, and the overfitting problem of the neural network in training is prevented; and the loss function of the neural network is improved, and the training accuracy of the power equipment partial discharge defect diagnosis neural network is improved.

The invention discloses a Hash code image training model based on binary weight, and a model algorithm comprises the steps: selecting a loss function, determining a target equation, and performing binary coding of a classifier and training image features; performing unified learning of a binary code, updating the binary code, and optimizing the loss function; and deducing the Hash code training model. The invention also discloses a classification learning method employing the Hash code image training model based on binary weight, and the method comprises the steps: obtaining a Hash code of a to-be-searched image through the Hash code training model based on binary weight, and solving Hamming distances between the Hash code and a classifier binary code; searching in the minimum Hamming distance from the Hamming distances, and obtaining the classifier corresponding to the minimum Hamming distance, wherein the classifier is the category to which the to-be-searched image belongs. The method can be used for the image classification for various types of images in high-latitude scenes, improves the performance of an algorithm in a large-scale data set, is precise, efficient and quick, andis small in consumption of the memory.

The invention discloses an unmanned multi-laser radar collaborative simulation method based on GPU acceleration, and the method is characterized in that the method comprises the following steps: 1, obtaining the depth information of laser radars at the initial position of a vehicle, converting the depth information into point cloud output corresponding to the laser radars, and obtaining the depth information and point cloud of the laser radars at the initial position; step 2, based on the latest position of the vehicle, using a GPU to generate a depth map surrounding the vehicle; step 3, constructing an optimization loss function and an optimization problem based on the depth map; step 4, simplifying and optimizing a loss function; 5, solving the optimal solution of the optimization problem; 6, judging whether the optimization loss function reaches an optimization threshold value or not according to the optimal solution; and 7, converting the depth information of the laser radar into a corresponding point cloud.

The invention provides a super-resolution reconstruction method. The method comprises the following steps: establishing a picture data set; constructing a neural network structure, wherein the neuralnetwork structure is used for extracting features of the picture data set in a neural network training process; establishing a loss function of a neural network structure, wherein the loss function isused for guiding neural network training; training the picture data set to obtain a neural network model; and reconstructing a picture by using the neural network model, inputting a low-resolution picture, and outputting a high-resolution picture. According to the technical scheme of the present invention, the SRGAN (Super-Resolution Generation Advanced Network) is improved, so that the SRGAN (Super-Resolution Generation Advanced Network) is improved; according to the super-resolution reconstruction method, the network structure of the generation network G-NET is changed, the loss function isimproved, and after improvement, the generation network G-NET extracts more accurate features, so that the super-resolution reconstruction effect is better, and a better effect can be obtained duringdetection, recognition and semantic segmentation.

The invention is suitable for the technical field of speech recognition, and provides a speech recognition method and device, and terminal equipment, and the method comprises the steps: calculating afirst conditional probability of a sentence according to a pre-trained language model; adjusting a first loss function of a speech recognition model according to the first conditional probability to obtain a second loss function; and training the speech recognition model by using the second loss function, and performing speech recognition by using the trained speech recognition model. According tothe method, the speech recognition accuracy can be improved.

The invention provides an image super-resolution reconstruction method. The image super-resolution reconstruction method comprises the steps of obtaining a high-resolution image; performing image preprocessing on the high-resolution image to obtain a training set; training the training set by using a multi-scale loss convolutional neural network to obtain a training model; and performing super-resolution reconstruction on a to-be-reconstructed image by using the training module to obtain a super-resolution image. According to the invention, the training model is obtained based on the multi-scale loss convolutional neural network, the loss function in the model is improved, the difference between the SR image and the HR image under different sizes is comprehensively considered, the fusion of high-layer information and low-layer information is realized, and the image super-resolution reconstruction quality is effectively improved. The invention further provides an image super-resolutionreconstruction system, a computer readable storage medium and an image processing terminal, which have the above beneficial effects.

The invention discloses a network intrusion detection method for improving BYOL self-supervised learning, and the method comprises the following steps: 1, carrying out the preprocessing of a UNSW-NB15 intrusion detection data set, and carrying out the one-hot coding processing and data normalization processing of character type data; step 2, training an improved BYOL intrusion detection model, and step 3, testing the improved BYOL intrusion detection model, inputting a preprocessed test data set into a feature extraction encoder f theta to obtain a feature representation of each piece of data in the data set, and inputting the feature representation into a classifier to obtain a classification result of each piece of data. The method has the advantages that BoTNet of a multi-head attention mechanism is introduced to suppress features with small contribution to classification in intrusion detection data, and features with large contribution to classification are increased, so that various performance indexes of the model are enhanced; and a BYOL loss function is optimized, so that the model training process is more stable and the convergence speed is accelerated, and the stability and robustness of the model are enhanced.

The invention relates to artificial intelligence, and provides a character attribute identification method and device, computer equipment and a storage medium. The character attribute identification method comprises the steps of classifying to-be-identified character images according to an occlusion area to obtain image types of the to-be-identified character images; if the to-be-identified figureimage is an occlusion image, inputting the to-be-identified figure image and the image type into a multi-scale generative adversarial network, and generating a de-occlusion image of the to-be-identified figure image by the multi-scale generative adversarial network according to the image type; and inputting the de-occlusion image of the figure image to be identified into an attribute identification network for attribute identification. The method can effectively remove the shielding interference in the shielded figure image, recover useful information, and improve the image attribute recognition performance. In addition, the invention also relates to a blockchain technology, and the character attributes of the to-be-identified character image can be stored in the blockchain.

The invention discloses a road target detection and association method based on millimeter-wave radar and vision. The method comprises the following steps: 1, installing the millimeter-wave radar and a camera on a vehicle; 2, constructing an improved YOLOv3 network which comprises a main body network and three feature prediction layers, and constructing a loss function and training the network based on DIOU so as to obtain a trained visual detection model; 3, configuring parameters of the millimeter-wave radar, acquiring CAN signals by using the configured millimeter-wave radar, processing the CAN signals, outputting preset frames with different target preset lengths and widths for the millimeter-wave radar, and projecting the preset frames back to a pixel coordinate system; and 4, calculating cost values of the radar target preset frame and the visual target frame through a DIOU-based cost calculation method, and setting a reasonable threshold value for optimization solution through a Hungary algorithm to obtain a global optimal matching result. According to the method, the correlation precision of the millimeter wave radar detection target and the visual detection target can be improved, so that richer target information can be obtained.