146 results about "Pyramid (image processing)" patented technology

The invention belongs to the technical field of remote sensing image processing, and particularly relates to a remote sensing image building change detection method, which comprises the following operation steps: (1) reading an image and preprocessing the image; (2) making a sample data set; (3) constructing a network model based on an attention mechanism and a feature pyramid; (4) selecting a training sample to train the network model; (5) selecting a verification sample to verify the network model; and (6) classifying by using the trained model, and outputting a final change detection result. According to the method, a feature pyramid network is introduced, an attention mechanism is used in the multi-scale feature fusion process of all levels, and features are enhanced layer by layer tobe used for target detection of different scales; through application of deformable convolution and cavity convolution, the network obtains the feature expression capability of automatically adaptingto object deformation, the feature size is reserved while the receptive field is expanded, multi-scale information is obtained, the false alarm rate is effectively reduced, and the detection precisionis improved.

The invention provides a method for detecting an X-ray mammary gland lesion image based on feature pyramid network under transfer learning. The method comprises the steps: 1, establishing a source domain data set and a target domain data set; 2, establishing a deformable convolution residual network layer by a deformable convolution and extended residual network module; 3, establishing a multi-scale feature extraction sub-network based on a feature pyramid structure through a feature map up-sampling and feature fusion method in combination with a deformable convolution residual network layer;4, establishing a deformable pooling sub-network sensitive to the focus position; 5, establishing a post-processing network layer to optimize a prediction result and a loss function; and 6, migratingthe training model to a small sample molybdenum target X-ray mammary gland focus detection task so as to improve the detection precision of the network model on the focus in the small sample image. According to the method, a transfer learning strategy is combined to realize focus image processing in a small sample medical image.

The invention provides a static tile pyramid parallel building method based on MapReduce, and belongs to the technical field of remote sensing image processing. The static tile pyramid parallel building method based on the MapReduce aims at fast performing batch remote sensing image parallel processing on the basis of a parallel calculation model of the MapReduce by aiming at the features of remote sensing images. The method is completed by four steps including Map tile processing, creation of Lmax layers of tiles, tile image generation and Reduce tile merging. The invention provides the improved static tile pyramid parallel building method based on the MapReduce, and the combination of fast parallel building of batch remote sensing image tile pyramid and tile pyramid large file generation is realized. Through structures of tile pyramid large files and HBase tile index tables, a fast retrieval model of global tile images is built.

The invention discloses a video pedestrian detection method fusing multi-target tracking clues, and belongs to the field of image processing of computer vision.A pedestrian detection module (improvedFaster R-CNN) and a multi-target tracking module are included. According to the method, the feature pyramid is introduced into the feature extraction network, pedestrians are detected on more scales,features of different layers are fused to improve the prediction effect, and the detection rate and accuracy of small targets are improved. According to the invention, the multi-target tracking moduleis used for assisting pedestrian detection, the inter-frame information of the front and rear frames of the video is introduced, the detection stability can be improved, and the historical frame target track and the target number obtained through the tracking module are used for relieving the detection instability caused by deformation, shielding and targets located at the edge of the picture. According to the invention, the anchor point box is modified in the pedestrian detection network part, so that the pedestrian detection network part better conforms to the characteristics of pedestrians, thereby improving the pedestrian detection precision.

The invention belongs to the technical field of image processing, and particularly relates to a face recognition method based on cosine loss non-constraint conditions. The method comprises the following steps: S1, obtaining a to-be-recognized image, and performing multi-scale transformation on the to-be-recognized image to obtain an image pyramid; S2, inputting the image pyramid obtained in the step S1 into an MTCNN network, and processing the image by the MTCNN network to obtain facial feature points; S3, performing face correction according to the facial feature points in the step S2; S4, training an Inception-ResnetV1 convolutional neural network by using the data processed in the step S3, training a classifier model by using a cosine loss function as a supervision signal to obtain a feature extraction model, and performing verification and recognition of face data by using the feature extraction model.

The present invention is to provide an image processing apparatus using the difference among scaled images as a layered image and a method thereof, which utilize the Gaussian and Laplacian pyramid theory to convert an original image into a plurality of scaled images of different scales, and the difference among scaled images of two adjacent different scales as a layered image of the corresponding layer, so that the edge and line characteristics of a scene of the original image for each layered image can be displayed in different levels sequentially from a clear level to a vague level, and provide a layered image display interface and an image characteristic editing interface for users to examine each layered image through the layered image display interface and edit or perform special effect to each layered image, so as to simulate different visual effects based on different vision models.

The invention provides an image processing method for automatically judging fetal hydrocephalus from ultrasonic images, comprising the following steps: firstly extracting skull ellipse of a new fetal brain image on the second plane, lateral ventricle horizontal plane; establishing a space histogram pyramid in the interior zone of the ellipse in the image; respectively calculating the distances among the space histogram pyramids of the given fetal ultrasonic image, the normal fetal brain images in the same gestational week and the image of hydrocephalus in the brain on the lateral ventricle horizontal plane in the same gestational week; and setting a threshold and giving the possibility of fetal hydrocephalus from judgment according to the relationship between the distances among the space histogram pyramids and the threshold.

The invention discloses a garage pedestrian detection method based on an improved Effective Det model, belongs to the technical field of target detection in image processing, and aims to enrich the background information of pedestrian detection by using a mosaic data enhancement method and calculate the data of four images at one time during Batch Normalization calculation. A feature shunt networkCSPNet is introduced into an EfficientNet, so that the learning ability of the CNN is enhanced, the detection accuracy can be maintained while the model is lightened, and the calculation bottleneck and the memory cost are reduced; a spatial pyramid pooling module SPP is introduced to the top of the feature extraction network, the receptive field of the network is increased, and pedestrian detection can be completed accurately and quickly in a complex and changeable garage environment.

The invention relates to a head posture estimation method combined with YOLO-MobilenetV3 face detection, belonging to the field of machine vision image processing. The method comprises the following steps: constructing a face and face key feature point detection module, improving an original YOLOV3 loss function, preprocessing an image containing a head posture, training an improved YOLOV3 detection model, inputting a test image into the trained model, and calculating an Euler angle and estimating the head posture through combination with a PnP algorithm. The method has the beneficial effects that the MobileNetV3 structure is used for replacing the darknet53 structure in an original YOLOV3 to realize light weight of the model; a spatial pyramid pooling module is adopted to improve the acquisition of global semantics of the image by the model, and the spatial position information of a target is processed more effectively, so detection precision is improved; and a YOLOV3 output layer and the loss function are improved, so the model can detect key points of face, practical application is easy, and the head posture can be estimated more quickly and accurately.

A processor performs pattern matching on a search target image using a reference image. The processor uses the reference image to calculate a model pyramid, which has model edges and different layers, and uses the search target image to calculate the search target pyramid, which has search target edges and layers having size reduction factors which are the same as those of the model pyramid. The processor performs pattern matching on the search target pyramid using the model pyramid. Calculation of the model pyramid includes (i) extracting and calculating sizes of edges from the reference image, (ii) obtaining maximum size reduction factors of the individual edges, and (iii) setting an edge (a) which has a size reduction factor equal to or larger than a target layer size reduction factor and (b) which has been subjected to size reduction as a model edge of the target layer.

An embodiment of the invention discloses an image processing method, intelligent equipment and a computer readable storage medium. The image processing method comprises the steps of: acquiring an original image to be subjected to moire processing; and running an image processing model to carry out moire removal processing on the original image to obtain a target image, wherein the image processingmodel is a network model which is obtained by training in advance according to a moire training data set and can remove moire included in the image. The image processing model comprises a multi-bandmodule, wherein the multi-band module is used for processing the original image to obtain N layers of Laplace pyramids about the original image, the multi-band module obtains a first processing resultfeature map based on feature maps corresponding to N spatial frequency bands of N layers of Laplacian pyramids, the target image is obtained according to the first processing result feature map, andN is a positive integer greater than or equal to 2. By adopting the image processing method, the intelligent equipment and the computer readable storage medium, a better moire removing effect can be achieved.

The invention belongs to the technical field of image processing. The invention discloses a sample extraction method and an image classification method based on hole convolution and residual link. According to the sample extraction method, an image pyramid of an original image is constructed, then convolution is carried out on the image pyramid, an identification result is extracted, a region identification image with a category label is finally obtained through methods of binarization, fuzzy clustering and the like, and finally a homologous reliable classification sample is obtained through image mapping. According to the image classification method based on cavity convolution and residual link, a completely self-adaptive deep learning classification network is constructed by a method ofsimulating clear vision in the central fovea, different types of unique feature modes are learned from homologous reliable samples of the images, and completely self-adaptive classification is carriedout on the images according to the learned feature modes.

The embodiment of the invention relates to the field of image processing, and discloses a monocular structured light depth recovery method, electronic equipment and a storage medium. The method comprises the following steps: acquiring an object speckle image and a reference speckle image of a target object as a monocular structure light image; generating an image pyramid comprising N layers of pyramid images according to the monocular structured light image, wherein N is an integer greater than 1; processing the image pyramid in such a way of: acquiring a parallax image of the (i-1) th layer according to the parallax image of the ith layer and the pyramid image of the (i-1) th layer until the parallax image of the first layer serves as the target parallax image of the target object, wherein the initial value of i is N, and the parallax image of the Nth layer is acquired according to the pyramid image of the Nth layer; and determining depth information of the target object according to the target parallax image. By using the embodiment of the invention, the calculation amount of monocular structured light measurement depth information can be reduced, the calculation speed is improved, and the requirement on deployment equipment is reduced.

The invention relates to an image processing technology, is applied to the field of smart medical treatment, and discloses an abnormal cell automatic labeling method. The method comprises: performingGaussian convolution smoothing processing on a cell pathology picture to obtain an image pyramid; performing detection fitting on the image pyramid through a Hough transform circle detection method toobtain a low-power fitting region of interest; mapping the low-power fitting region of interest to the acquired cell pathology picture to generate a fitting high-power image; segmenting a region of interest in the fitting high-power image to generate a segmented high-power image; labeling and cutting the segmented high-power image through an adaptive threshold segmentation algorithm to obtain anabnormal cell labeling set; and mapping the abnormal cell labeling set to the cell pathology picture to obtain an abnormal cell labeling picture. The accuracy of abnormal cell labeling is improved, and the calculation and storage pressure is reduced. In addition, the invention also relates to a blockchain technology, and the abnormal cell label set can be stored in a blockchain.

The invention discloses an underwater image splicing method and device, and belongs to the technical field of image processing. The method comprises the following steps: acquiring at least two underwater images to be spliced; sequentially performing red channel attenuation compensation processing and white balance processing on each underwater image to obtain a preprocessed underwater image corresponding to each underwater image; performing image enhancement processing on each preprocessed underwater image based on the multilayer image pyramid to obtain an enhanced underwater image corresponding to each preprocessed underwater image; performing image registration processing on the at least two enhanced underwater images based on grid optimization to obtain registered underwater images; performing image fusion processing on the registered underwater images based on a Laplace pyramid multi-scale image fusion algorithm to obtain a fused underwater image; and determining that the fused underwater image is a spliced underwater image, and outputting the spliced underwater image. The invention realizes underwater image splicing method.

The invention relates to the field of image processing, in particular to an image processing method and device, electronic equipment and a storage medium. The image processing method comprises the steps: obtaining a to-be-processed original image and a reference image corresponding to the original image, wherein the resolution of the reference image is greater than that of the original image; performing feature extraction on the original image to obtain a first feature set; performing feature extraction on the reference image to obtain a second feature set; performing feature alignment according to the first feature set and the second feature set to obtain a plurality of alignment feature data with a plurality of different image scales; and performing feature fusion on the plurality of alignment feature data and the original image to obtain a processed target image. Characteristic alignment is carried out layer by layer through pyramid cascading according to characteristic data of different scales, so reference image characteristics are effectively matched to corresponding positions of an original image, influences caused by different shooting angles or shooting time are reduced oreliminated, and the reference image characteristics are more effectively combined in the image super-division process.

The invention discloses a real-time video stitching algorithm based on optimal stitching line updating, and belongs to the technical field of digital image processing. According to the real-time videostitching scheme provided by the invention, an image registration technology based on feature points is used, and internal and external parameters of each camera are estimated through matching feature points of different view angle images. When image fusion is carried out between different visual angles, an image fusion algorithm based on an optimal suture line is adopted, and a background elimination algorithm based on KNN and a suture line updating algorithm based on dynamic programming are used for avoiding discontinuity and ghosting caused by the fact that a moving object penetrates through the suture line in order to solve the problems of ghosting and discontinuity possibly generated by the moving object. And finally, splicing seams are eliminated by using an image fusion algorithm based on a convolution pyramid. By adopting the video stitching scheme designed by the invention, a good large-view-field image can be provided under the condition of ensuring the stitching real-time performance.

The invention belongs to the technical field of remote sensing image processing, and discloses a multi-scale nested remote sensing image change detection method and system and a computer terminal.The multi-scale nested remote sensing image change detection method comprises the steps that a public data set is selected, and the data set is preprocessed; constructing a network model based on an ECA attention mechanism to extract single-time-phase image features; designing and constructing a network model based on multi-scale double superposition; and selecting a training sample to train the network model and selecting a test set to test the network model. According to the method, change detection is carried out on a remote sensing image by using a feature pyramid, a twin network and an attention mechanism, the problem of class imbalance is solved by weighting two loss functions, and the influence of extraction of network fusion features and class imbalance on a detection result is enhanced by considering multi-scale superposition. The running speed is increased to a great extent; and meanwhile, the resource consumption is reduced.

The invention belongs to the technical field of medical image processing, and particularly relates to a deep detection network for quantifying esophageal mucosa IPCLs vascular morphological distribution. The deep detection network comprises a feature extraction network, a feature pyramid, a region candidate network, an interest region pooling and clustering distribution priori self-embedded cancerlesion classification network and a system for visualization on a narrow-band imaging endoscope image. The feature extraction network extracts a feature map of the input image; the feature pyramid fuses the features of different scales; the region candidate network proposes a possible lesion region; the region of interest is pooled, and the features are pooled to a suspicious lesion region; the cancer lesions are classified by a clustering distribution priori self-embedded cancer lesion classification network; and finally, visualizing is carried out on a narrow-band imaging endoscopic image,and frame selection marking is carried out on the cancer lesion by using different colors. The cancer focus of the early esophageal squamous cell carcinoma existing in the image is detected and diagnosed, the diagnosis efficiency can be effectively improved, and a doctor is assisted in obtaining higher diagnosis precision.

The invention discloses a two-dimensional code recognition method, and the method comprises the steps: selecting QR two-dimensional codes in different backgrounds according to the actual recognition demands, adding motion blur with different blur lengths and blur angles, and designing and manufacturing a two-dimensional code data set based on motion blur; secondly, an image motion blur removing algorithm based on a generative adversarial network is adopted; according to the image deblurring method based on the generative adversarial network, the feature pyramid-based image feature extraction network is adopted to extract feature mapping of the image, and a new GAN loss function is designed, so that the image deblurring algorithm based on the generative adversarial network is faster in training speed and higher in robustness, and a very efficient image deblurring effect can be provided. An image binarization algorithm based on an adaptive threshold and a series of algorithms based on image processing are designed, so that an image edge region is better extracted and an image is better corrected. The method can be applied to two-dimensional code recognition environments with different backgrounds.

The invention discloses an image processing method, device and system based on an alignment feature pyramid network. The method comprises the steps of S10, performing up-sampling convolution on upper-layer features; S20, performing alignment operation on the upper-layer features which are subjected to up-sampling convolution and the lower-layer features; and S30, fusing the aligned upper-layer features and lower-layer features. According to the method, the offset vector field between the upper-layer feature and the lower-layer feature is estimated, and alignment operation is performed before addition and fusion, so that it is ensured that the upper-layer feature can be fused to the correct position of the lower-layer feature, and the network accurate identification capability is improved.Besides, due to the fact that the low-resolution high-level features are used, the whole network can process images with very high resolution under the condition of limited calculation and storage expenditure, meanwhile, the recognition precision of the method using the high-resolution high-level features can be achieved, and possibility is provided for practical application.