37results about How to "Rich image features" patented technology

The invention discloses a speckle-based three-dimensional-face reconstruction method and device. The method and the device can reduce matching calculation quantity while image feature richness is improved, and realize synchronized improvement of three-dimensional-face reconstruction efficiency and reconstruction precision. The method comprises: inputting a plurality of speckle image pairs conforming to binocular stereo vision; carrying out polar line correction on the input speckle image pairs to enable corresponding matching points to be at the same horizontal lines; carrying out face detection on the speckle image pairs after polar line correction, and extracting region-of-interest image pairs of to-be-reconstructed faces; carrying out first matching processing on the region-of-interestimage pairs to obtain first parallax graphs; setting a matching point searching range according to parallax in the first parallax graphs, and carrying out second matching processing on the region-of-interest image pairs to obtain second parallax graphs; and generating corresponding three-dimensional point cloud data according to the second parallax graphs and camera calibration parameters to obtain the reconstructed three-dimensional faces.

The invention discloses a welding seam defect identification method based on an improved LeNet-5 model. Firstly, input of traditional convolution kernel channels of the LeNet-5 model is improved for the welding seam grayscale image, the grayscale image is converted to a color image through pseudo-color enhancement technology, and the obtained color image is used as input of a neural network; thenconvolution kernels of the LeNet-5 model are improved, and convolution kernel channels with Gabor filters are added; features obtained by the multiple channels are fused in a sixth layer of the neuralnetwork to obtain a feature set T; and finally, a SoftMax classifier is used in a seventh layer (output layer) of the neural network to obtain the defect type of a welding seam and probability of each category, to which the same belongs, to use the same to provide a reference for negative-film type determination of a film evaluator and site rework scheme formulation. According to the method, feature extraction capability of the neural network is improved, and thus a correctness rate of defect identification is improved; and an identification result is given in a form of the probability of thecertain categories to which a defect belongs, and more sufficient reference information is provided for the film evaluator.

The invention provides a dual-channel output contour detection method based on encoding and decoding structures. In the encoding stage, image feature information is extracted through an improved VGG16network, in the decoding stage, feature information of different scales is fused from bottom to top, and the same label is used for carrying out deep supervision on output contour images of two channels. According to the method, the feature maps of different scales are fused in a bottom-up layer-by-layer decoding mode, so that the extracted image feature information is richer; adding a channel attention structure in a feature fusion stage, and performing feature map sampling by using sub-pixel convolution; a proper loss function is designed to solve the problem of imbalance of training samples; and a data set is amplified by using a data enhancement method, so that the generalization ability of the model is improved. According to the method, the target contours of the BSDS500 public dataset and the customized woodcarving contour detection data set can be effectively extracted, and the detection contour lines are fine.

The invention discloses a target detector and a construction method and application thereof, and the method comprises the steps: building a Faster R-CNN target detection model framework which comprises a regional suggestion network module RPN and a plurality of cascaded multi-core multi-background detection structures; adopting RPN to generate a training sample set; based on the training sample set and the weight distribution thereof, iteratively training a plurality of cascaded multi-core multi-background detection structures by adopting a loss function to obtain a Faster R-CNN target detection model; wherein after each multi-core multi-background detection structure is trained in each iterative training, the weight distribution is updated, the weight of the training sample with a large loss function value is large, and the cascaded next multi-core multi-background detection structure is trained based on the updated weight distribution and the regression sample generated by the current multi-core multi-background detection structure. According to the invention, a plurality of cascaded multi-core multi-background detection structures are introduced into the Faster R-CNN, and training is carried out based on weight distribution and updating thereof, so that the classification precision of the whole detector is improved, and the detector has relatively good detection performancein a complex background.

The invention relates to an image retrieval method and system, and relates to the field of electronic data processing. The method comprises the following steps: S1, acquiring n target images and extracting features of all the target images; S2, clustering the features and constructing a dictionary tree T1; S 3, calculating a frequency vector Fj according to the dictionary tree T1, and then obtaining a dictionary vector dj of each target image according to the frequency vector Fj; S4, acquiring a query image, calculating a dictionary vector qj of the query image, and calculating the similaritysj of the dictionary vector qj and the dictionary vector dj; and S5, obtaining a query result according to the similarity sj. According to the scheme, the technical problem of how to complete rapid retrieval of tens of thousands of levels of images is solved, and the method is suitable for rapid retrieval of tens of thousands of levels of images.

The invention relates to an image segmentation method and device. The image segmentation method comprises the following steps: inputting a to-be-processed image into a first image feature extraction network to obtain a first image feature; preset reference information, the first interaction information and the to-be-processed image are input into a second image feature extraction network to obtain a second image feature, the information amount of the image feature extracted from the image by the second image feature extraction network is smaller than the information amount of the image feature extracted from the image by the first image feature extraction network, and the second image feature is extracted from the to-be-processed image by the first image feature extraction network. The first interaction information is used for indicating position information of a to-be-segmented object in the to-be-processed image; obtaining a first target mask for the to-be-segmented object based on the first image feature and the second image feature; and performing segmentation processing on the to-be-processed image based on the first target mask to obtain a first segmentation result for the to-be-segmented object.

The embodiment of the invention discloses an image processing method and device, equipment, a storage medium and a computer product, which can be applied to various scenes such as cloud technology, artificial intelligence, intelligent traffic and auxiliary driving. The image processing method comprises the steps of obtaining input features of a to-be-processed image; the dimension of the space where the input features are located is a first dimension; obtaining an image processing network, the image processing network comprises N convolution layers, and each convolution layer comprises at least one convolution kernel; the dimension of the space where the weight value corresponding to the convolution kernel in each convolution layer is located is the first dimension; calling N convolution layers to map the input features and the weight value corresponding to the convolution kernel in each convolution layer to a mapping space, and then performing convolution operation; the dimension of the mapping space is a second dimension, and the second dimension is greater than the first dimension; and performing image processing on the to-be-processed image according to the convolution operation result to obtain a processing result. According to the method of the invention, the accuracy of image classification or recognition can be improved.

The invention relates to a hash center-based continuous learning method. The method comprises the steps that similar data pairs and non-similar data pairs in an image data set are input into a convolutional neural network for feature learning; a feature learning result passes through a hash layer, the hash layer comprises three full connection layers, and each full connection layer comprises a rowof neurons and an activation function; the hash codes of the similar data pairs converge to a public hash center after being subjected to center similarity constraint, the hash codes of the non-similar data pairs converge to different hash centers after being subjected to center similarity constraint, and the hash layer converts continuous depth representation into K-dimensional representation; after the hash layer outputs a real number vector, the K-dimensional representation is binarized into a K-bit binary hash code by using an activation function. According to the method, richer image features can be learned, less detail information of data is lost, the generated binary hash code is higher in accuracy and high in distinguishability, and the image hash retrieval performance can be improved.

The invention relates to a high-resolution remote sensing image change detection network. The network comprises a front time phase feature extraction branch and a rear time phase feature extraction branch which are arranged in parallel; wherein the front time phase feature extraction branch comprises a first convolution module and a front time phase feature fusion module; a front time-phase feature fusion module acquires low-level front time-phase features output by a low-level convolution layer and high-level front time-phase features output by a high-level convolution layer in the first convolution module, and fuses the low-level front time-phase features and the high-level front time-phase features to obtain front time-phase feature data; the rear time phase feature extraction branch comprises a second convolution module and a rear time phase feature fusion module; and the post-time-phase feature fusion module obtains low-level post-time-phase features output by a low-level convolution layer and high-level post-time-phase features output by a high-level convolution layer in the second convolution module, and fuses the low-level post-time-phase features and the high-level post-time-phase features to obtain post-time-phase feature data. The accuracy of a detection result is effectively improved.

The invention discloses a model training and table recognition method and device, and the method comprises the steps: determining a plurality of images containing a table as training samples, determining the mark of each training sample according to the structure and position of the table in the training sample, inputting the training sample into a feature extraction layer of a recognition model, and carrying out the recognition of the table. Determining an image feature pyramid corresponding to the training sample, for each feature map in the image feature pyramid, determining a reconstruction code corresponding to the feature map, performing up-sampling on the reconstruction code corresponding to the feature map, fusing the reconstruction code with other feature maps with the size larger than that of the feature map, and taking a fusion result corresponding to each feature map as input; and inputting a recognition layer of the recognition model to obtain a recognition result of the training sample. According to the method, fusion is carried out based on the feature maps of different sizes, the recognition result of the training sample is determined, the obtained image features are more comprehensive, abundant information can be obtained when the collected image is recognized, and the efficiency is high.

The invention relates to a generative adversarial network image defogging method fusing multi-level features. The method is based on an end-to-end defogging algorithm of deep learning, during down-sampling, feature extraction is carried out on foggy images with different resolutions, the extracted different feature maps are learned through an SE-ResNet module, information between channels is better fitted, and performance degradation caused by network deepening is prevented. And splicing the learned multi-level feature maps, and fusing more image features. And then, during up-sampling, putting a feature map after down-sampling fusion into an SE module for learning so as to better distribute channel weights and enhance useful features. And splicing the learned feature map with the up-sampling feature map, and fusing more image information. And finally, adding the residual image of network learning and an input foggy image to obtain a final image defogging result. Experimental results show that the image defogging method provided by the invention is better in defogging performance.

The invention discloses a target tracking method and device based on video denoising, computer equipment and a storage medium, and the method comprises the steps: carrying out the video denoising processing of a noise video according to a video denoising processing model, and outputting a denoised video; according to a target tracking model and the denoised video, tracking the target object to obtain tracking data of the target object; and processing the tracking data to obtain the processed tracking data, and generating the trajectory video of the target object based on the processed tracking data, therefore, by adopting the embodiment of the invention, the effective video noise reduction processing can be performed on the noise video through the video noise reduction processing model, and the video noise reduction processing efficiency is improved. According to the method, more image features can be reserved while image noise information is removed, so that the output denoised video has better definition, detection and tracking of the target object are facilitated, and the tracking accuracy of the target object is effectively improved.

The invention relates to the technical field of artificial intelligence, and discloses an image feature extraction method and device, equipment and a storage medium, and the method comprises the steps: obtaining a to-be-extracted image; inputting the to-be-extracted image into a target image feature extraction model for image feature extraction, the target image feature extraction model being a model obtained by training based on a VGG model, a residual module and a pyramid pooling module; and obtaining an image feature output by the target image feature extraction model as a target image feature corresponding to the to-be-extracted image. Therefore, the target image feature extraction model is obtained through training based on the VGG model, the residual module and the pyramid pooling module, the problem that time is often consumed when SIFT or ORB is adopted to extract image corner features is avoided, and the accuracy of image feature extraction is improved. The method can be applied to the fields of intelligent government affairs, digital medical treatment, science and technology finance and the like.

The invention provides a head attitude estimation method, system and device based on a multi-scale lightweight network, and a medium. The method comprises the following steps: obtaining a data set containing a head attitude, and preprocessing the data set; extracting the preprocessed data set by using a multi-scale convolutional network to obtain a corresponding feature map; training a lightweight network based on the feature map to obtain a MobileNet regression device model; and obtaining a head image of an image to be detected, and inputting the head image into the MobileNet regression device model for head posture prediction to obtain head posture information of the image to be detected. According to the method, the feature map in the data set is extracted by adopting the multi-scale convolution kernel, and the convolution kernels of different scales are used for extracting features of the input head posture image, so that the image features are enriched, the image information is reserved, and the accuracy of head posture estimation is improved; and meanwhile, the MobileNet regression device model is trained based on the lightweight network, and the calculation amount is greatly reduced on the premise that the network performance is not lost.

The invention provides a tourism road book generation method and system, a medium and an intelligent terminal. The generation method comprises the following steps: obtaining a plurality of user findings associated with corresponding position information; wherein the user discovery comprises pictures, videos and characters shot by the user in the travel process; carrying out information point association operation on each user discovery so as to recommend travel itinerary service to a target user; generating a travel route book based on the user discovery associated with the information point and the position information and the current position of the target user. According to the invention, interested travel information can be collected from each platform, so that when a user wants to travel, the travel information can be directly utilized to generate a travel road book for the user to refer to, thereby reducing the blindness of travel of the user, and greatly reducing the time and effort of the user for planning travel at the same time.

The invention provides a key point detection method of a human body posture image. By using an estimation network model provided by the invention, the characteristics are divided into different channels by designing the segmentation channel blocks, so that the image characteristics become richer, and the accuracy of human body posture detection can be effectively improved. Besides, an attitude correction machine is designed, and by improving a channel attention mechanism and a space attention mechanism, a context attention mechanism is introduced to improve the modification effect. The accuracy of human body posture estimation can be effectively improved.

The invention discloses an image and video processing method and system, a data processing device and a medium, and the image processing method comprises the steps: extracting the image features of an original image, and obtaining a first detail image; selecting a target pixel and a local area from the first detail image, wherein the local area comprises the target pixel; based on a statistical relationship between other pixels in the local area and the target pixel, calculating to obtain corresponding statistical feature information, and updating the statistical feature information into color information of the target pixel in the first detail image to obtain a second detail image; and combining the second detail image and the original image to obtain a composite image. According to the scheme, the image quality and the processing efficiency can be improved.

The invention provides a liquid state recognition method in liquid separation and a liquid separation system. The liquid separation system executes the liquid state recognition method, and the method comprises the steps: starting to acquiring a video of liquid in a pipeline from the liquid separation process, and extracting image frames in the video at an interval of a preset period; for each image frame, acquiring the image features of a to-be-recognized area in the image frame, inputting the image features into a trained liquid state recognition model, and obtaining a prediction state corresponding to the image frame, wherein the trained liquid state recognition model is obtained through training of a sample image frame with a liquid state label; and determining the liquid state in the pipeline based on the prediction state corresponding to each image frame. The liquid state recognition model adopted in the scheme is particularly suitable for the liquid separation process of the rapidly flowing liquid, the machine learning model is adopted to participate in liquid state prediction, so the analyzed image features are richer, the robustness of the algorithm is higher, and the accuracy and universality of liquid separation are better.