109 results about "Significance map" patented technology

The invention discloses an image significance detection method which uses confrontation training to generate a convolution neural network model, which belongs to the field of computer vision and image processing. The method comprises the steps of data preprocessing, network structure, suitable parameter selecting, and training with a random gradient descending method and an impulse unit. According to data preprocessing, a large amount of collected data and labels are preprocessed. According to network structure, a network structure and a specific kernel function are designed. Suitable parameters including learning rate, a momentum factor and the number of images inserted into the network each time are selected. The random gradient descending method and the impulse unit are used for training to reduce the possibility of network over-fitting. According to the invention, a significance map can be accurately acquired.

A method and apparatus for encoding bit code utilizing context dependency simplification to reduce dependent scans. The method includes retrieving at least one 2 dimensional array of transform coefficient, transforming the at least one 2 dimensional array of the significance map of the transform coefficient to a 1 dimensional coefficient scanning and determining at least one of scan direction, coding unit type and slice type assigned to transform coefficient, selecting neighbors based on at least one of scan direction and coding unit type and slice type, computing context index based on the values of the selected neighbors for context selection, and performing arithmetic coding to generate coded bit utilizing the computed context index and binarization.

The present invention relates to a semantic object segmentation method suitable for low field depth image, which includes: firstly introducing gradient histogram to figure out the distribution of image in the energy space, obtaining an energy focusing significance map in combination with the character of the low field depth map; using the two-sided filter and morphology instrument to rehandle the energy focusing significance map; and then setting self-adapting threshold value and processing to obtain an initial object mask map, combining with the edge information obtained by canny operator to obtain the corrected object mask, in order to raise the segmentation accuracy of the interesting object; finally using the Bayesian eclosion algorithm to obtain the ideal semantic object segmentation result, in order to delicately process the complicated image boundary, such as hairs. Accurate segmentation to the interesting object in the field depth scope in the image video sequence.

A method for generating a quality oriented significance camp for assessing the quality of an image or video, comprising the steps of extracting features of the image or video, determining a perceptual quality requirement of at least one extracted feature and integrating the extracted features and the perceptual quality requirement of the at least one extracted feature to form an array of significance level values, thereby generating the quality oriented significance map.

The invention discloses an offshore scene significance detection method. The method comprises the following steps of: 1, extracting an offshore scene image sequence; 2, transferring each frame image to the CIELab colour space, and extracting the characteristic pattern of luminance and colour passages; 3, using the absolute value of the difference between the extracted characteristics and the global mid value as the global significance map; 4, using the absolute value of the difference between the characteristics and the local mean value filtration as the local significance map; 5, combining the global significance map and the local significance map of the characteristics to obtain an overall significance map; 6, linearly combining the significance maps of the colour passages of frame images, and respectively combining the combined significance maps with the luminance significance maps to form an overall significance map; 7, accumulating by using each frame detection result as the centre and modifying the significance map of the current frame; and 8, converting the overall significance map into a binarization image to obtain an offshore scene significance target area. By adopting the method, the significance area in the offshore scene can be extracted rapidly, and the interference of sea noise wave can be favorably inhibited. The method is simple in implementation, and is suitable for real-time application.

The invention discloses a cooperative significance detection method based on superpixel clustering, and the method comprises the steps: constructing a superpixel pyramid, and replacing a common pixel through a superpixel block; speeding up the cooperative significance calculation; constructing the superpixel pyramid, so the feature information at different scale can be obtained, and the boundary accuracy of a cooperative significance target can be guaranteed. Based on the above, a clustering method can achieve the further classification of superpixel blocks, ad further speeds up the calculation of the cooperative significance. Finally, a method of fusion of a cooperative significance map and a saliency map is used for obtaining a final cooperative significance map, thereby guaranteeing the accuracy of the cooperative significance target. The boundary contour of the obtained cooperative significance target is more accurate in location, and the method has advantages in time and accuracy.

A data processing system includes a subband decomposition of a source signal by encoding of the significant pyramid-type subband representation. While encoding process, M-array is used for initialization of set of arrays Si (arrays of solution bits), which define the order of creating output data streams. Output data streams contain an encoded significance map and values of (significant) non-zero subband decomposition coefficients adapted for successive quantization and entropy encoding.

The invention discloses a video saliency detection method based on deep fusion. The method comprises a deep feature extraction network, a deep feature fusion network and a saliency prediction network.Firstly, a depth feature extraction network extracts multistage depth features to generate depth space features and depth time features; then, an attention module is adopted to respectively reinforceand learn the depth features generated by each stage in the two branches, and the depth features are fused with the multi-stage depth features in a depth feature fusion network grading mode; and thedepth features obtained by fusion is combined with boundary information and prediction is carried out by a salient prediction network to generate a final significance map of the current frame. According to the network model provided by the invention, spatial information and time information can be fully and effectively utilized to predict the salient target in the video.

A coder for use in encoding and decoding a data set representing an image includes a first device which partitions the subband transformation into first and second sets, which adds the first set into a list of insignificant sets (LIS), and which initializes a list of significant pixels (LSP), a second device which tests the first and second sets for significance with respect to a threshold value, which partitions significant members of the first and second sets in accordance with first and second partitioning functions, respectively, and which adds significant pixels to the LSP, a third device which refines the quantization of the pixels in the LSP. In operation, the threshold value is decrement as the second and third devices are operation in seriatim until the coding endpoint has been reached. An optional fourth device entropy codes a significance map cooperatively generated by the second and third devices. Methods for encoding and decoding the subband transformation of a data set and a computer program for converting a general purpose computer into a hierarchical image coder are also described.

The invention discloses a method for recognizing audio based on spectrogram significance test. The method is characterized by comprising the following steps: 1, acquiring spectrograms of different sound sources, and extracting characteristics to obtain a basic characteristics set; 2, obtaining a significance map by the GBVS algorithm, and extracting a main map by the main map separating method; 3, extracting a hierarchy correlation map; 4, acquiring a PCA characteristics map; 5, building GCNN sound source models of different sound sources; 6, recognizing the sound sources of which the spectrograms are to be tested according to the GCNN sound source models. With the adoption of the method, the characteristic information of unknown audio type under a complex environment can be effectively represented, and meanwhile, the audio can be quickly and automatically recognized.

The invention discloses an RGB-D saliency target detection method, and the method comprises the following steps: respectively obtaining the single-mode saliency features of each stage of an RGB imageand a depth image through single-mode feature extraction; selecting RGB and depth single-mode saliency features of each level through cross-mode joint feature extraction, capturing complementary features of an RGB image and a depth image level by level, and generating cross-mode joint features; inputting the cross-modal joint features and the single-modal saliency features into a saliency target detection part. By designing low-value fusion loss and network overall loss, fusion of RGB flow and deep flow detection results and supervised learning of the network are realized, and a final saliencydetection result is output. According to the method, valuable cross-modal joint features are extracted and captured through the cross-modal joint features, the network pays attention to a low-value significance region of the significance graph through designed low-value fusion loss, and the lower bound of the significance value is improved.

The invention particularly relates to a method for detecting significance of a stereo image based on human eye stereo visual characteristics. The method comprises the following steps of calculating the significance of view information of two different dimensions (space and depth) of a stereo image; firstly, using an SLIC (simple linear iterative clustering) algorithm to perform ultrapixel segmentation on a single-viewpoint view, and combining similarity of areas; then, using a GBVS (graph-based visual saliency) algorithm to calculate a two-dimensional space significance map; combining the absolute parallax feature, local parallax feature and depth contrast feature in a parallax map to perform depth significant calculation; finally, combining with human eye visual fatigue characteristics to reasonably process the parallax distribution, clustering the significance of the two different dimensions of significance maps by a linear weighting type, and generating a stereo image significance map. The method can be used for effectively detecting the significance area of the stereo image under different scenes, and is suitable for the fields of doubtful object calculation, image searching and the like.

The invention discloses a salient target detection method based on a cascade convolutional network and adversarial learning. The method comprises the following steps: 1, designing a global saliency estimator E; 2, designing a local saliency refiner R; 3, combining the global saliency estimator E and the local saliency refiner R into a generator G based on a cascade convolutional neural network forgenerating a saliency map; 4, optimizing the generator G; 5, designing an adversarial learning discriminator D to distinguish a real saliency map from a predicted saliency map generated by a generator G; and 6, the generator G and the adversarial learning discriminator D follow the CGAN strategy and are trained in a complete end-to-end manner, so that the generator G can better understand the structure information of the salient object, and a good saliency detection result is obtained. According to the method, the structural information is learned implicitly through confrontation learning, sothat significance target detection can be well carried out, and a best result is obtained on a plurality of databases.

The invention discloses a depth significance-based stereopicture just noticeable difference (JND) model building method. The method comprises the steps: calculating the horizontal parallax of a stereopicture pair to obtain the horizontal parallax map of the stereopicture pair; calculating the depth value of the stereopicture pair to obtain a depth map of the stereopicture pair; calculating the depth significance of the stereopicture pair to obtain the depth significance map SD of the stereopicture pair; and building a depth significance-based stereopicture JND model. The method fully considers the depth significance influence factors in stereopicture noticing, the model obtained by adopting the method can more accurately reflect the feeling of eyes, the stereopicture processed by the guidance of the model can be added with more noise under the condition of keeping the subjective quality basically unchanged, thus being capable of removing vision redundancy in the stereopicture videos.

The invention provides an image zooming method based on a significance model. The method comprises steps of firstly, preprocessing an input colorful image; then, using a method based on superpixel global contrast to calculate to obtain a significance image of the image; combining a result of a significance detection and gradient information of the image to obtain an energy matrix; and by combining the energy matrix, using a line cutting algorithm to zoom the original image, thereby achieving adjustment of the image size. According to the invention, a significance region of the image can be detected and when the image size is freely changed, the main body features in the image are kept unchanged.

The invention relates to a marine target significance detection method based on spectrum singular value decomposition. The method comprises the following steps of utilizing the brightness and color channels of a marine image CIELab space to respectively carry out Fourier transform; according to set threshold values, selecting the non-main components of amplitude spectrum, and combining with original phase spectrum to carry out Fourier inverse transform, to obtain the significance map of each feature; and combining the color significance maps, and then combining with brightness significance maps to obtain a total significance map. The method has the advantages that a significance area in the marine scene is quickly extracted, so the target detection of a marine scene is favorably realized, the interference of marine clutters is better inhibited, the combination of significance maps with a plurality of dimensions is not needed, and the method can be realized on original image dimension; the method provides machine vision auxiliary means for target detection in marine peril searching and rescuing, marine monitoring, port video monitoring, detection of various ships in marine enforcement evidence collection and the like.

The invention belongs to the image processing field and relates to a significance detection method based on level-set super pixel and Bayesian frame to solve the detection problem of image significance. The method comprises the following steps: firstly, segmenting and combining the result of a level-set method to obtain new super pixels that meet the sizes of the different areas of the image; secondly, using the differences in colors and distances among the super pixels of the image inner part and the image edge part to construct a significance image; then, using the super pixels to represent the significance areas; putting forward three updated algorithms under the Bayesian framework; updating the significance image to obtain a significance result; increasing the currently available algorithm result to a similar level by the updated algorithms; and at least, using the detection algorithm based on human face recognition to process the image containing a person. The method of the invention is capable of recognizing the most significance part in an image and of increasing the result of a currently available algorithm to a higher level.

The invention discloses a significance detection method based on background priors. The method comprises the steps that a color image and a depth image are taken as input; superpixel segmentation anddepth map quality evaluation are carried out on the input images through a preprocessing operation; based on depth selective difference and background priors, the initial significance value of each superpixel region is calculated; and finally an initial significance map is optimized through the minimization of a cost function to acquire the final significance detection result. According to the invention, the problem that a traditional significance detection method based on color images cannot detect objects with similar visual characteristics with the background is solved; the problem of misdetection, which is caused by the fact that a bottom background region cannot be ignored when significance detection is carried out simply by relying on depth information; is solved; and the method provided by the invention is suitable for significance detection with color images and depth images, has the advantages of good overall effect and high accuracy, and can effectively detect significant objects.

The invention discloses an unsupervised oil tank target detection method based on a shape-guided significance model. The method comprises the steps of: inputting a remote-sensing image, calculating anedge response graph of the remote-sensing image, and carrying out clustering on all pixels in the remote-sensing image to form superpixels to obtain all the superpixels of the remote-sensing image; obtaining a plurality of clustering regions on the basis of all the superpixels and the edge response graph; utilizing the clustering regions to obtain round probability and a round probability graph;calculating a shape guidance-based significance graph according to all the superpixels and the round probability graph; obtaining a binary result graph through shape-guided significance graph; utilizing a binary result graph to mark oil tank regions in the remote-sensing image to obtain the target regions. Through the target detection method, oil tank targets in low-resolution remote-sensing images under different sizes and illumination conditions can be accurately detected, and the method has better robustness.

The invention provides a significance detection method based on selection of a regional candidate sample, and belongs to the technical field of artificial intelligence. On the basis of existing prior knowledge, depth characteristics and a classifier are introduced, and a selecting mechanism from rough to fine is used, the significance and targeted performance of the regional candidate sample are evaluated, a detection result is further optimized by utilizing super pixels, and a significant object in an image can be detected effectively. Compared with a traditional method, the detection result is more accurate. Especially for a multi-object image or an image in which the object is very similar to background, the detection result satisfies visual perception of humans more, and an obtained significance map is more accurate.

The invention discloses an HEVC (High Efficiency Video Coding) decoding complexity control method based on video significance and belongs to the field of video decoding. According to the method, firstly, the significance of each LCU (Largest Coding Unit) is predicted through a region significance detection algorithm of an HEVC compression domain; secondly, a plurality of training videos are selected to train function relationships between deblocking filters / simplification movement compensation and decoding complexity / video sensing distortion, and an HEVC decoding complexity control and sensing quality optimization equation is established according to a significance map and the function relationships; and finally the equation is solved, thereby precisely controlling HEVC decoding complexity and ensuring the sensing distortion to be least. The method has the advantages that when the decoding complexity is controlled, the sensing quality of a video is ensured to the greatest extent, a user can carry out decoding according to target complexity appointed by a set demand at a decoding end, so that HEVC decoding can be widely applied to terminal devices with different computing capabilities or different electric quantities, or video playing can be finished in appointed time under a certain electric quantity.

Methods of encoding and decoding for video data are described in which multi-level significance maps are used in the encoding and decoding processes. The significant-coefficient flags that form the significance map are grouped into contiguous groups, and a significant-coefficient-group flag signifies for each group whether that group contains no non-zero significant-coefficient flags. If there are no non-zero significant-coefficient flags in the group, then the significant-coefficient-group flag is set to zero. The set of significant-coefficient-group flags is encoded in the bitstream. Any significant-coefficient flags that fall within a group that has a significant-coefficient-group flag that is non-zero are encoded in the bitstream, whereas significant-coefficient flags that fall within a group that has a significant-coefficient-group flag that is zero are not encoded in the bitstream.

The invention discloses a significant object segmentation method based on adaptive three threshold values, and the method comprises the steps: firstly calculating the significance value of each pixel in a color image through a regional contrast algorithm, and generating significance maps in the same size; secondly carrying out statistics to obtain a significance histogram, finding a threshold value to preliminarily classify the significance maps into two classes through an adaptive three-threshold-value method, enabling the difference between the two classes to be maximum, finding the other two threshold values to finely classify the significance maps into four classes, and enabling the difference among the four classes to be maximum. The method enables the pixels of the significance maps to be classified into four types of seed points according to the obtained three threshold values, replaces manual interaction with the seed points to carry out the initialization of a GrabCut algorithm, and obtains a segmentation result. The method provided by the invention obtains the seed points from the significance maps through adaptive three threshold values, and effectively improves the significant object segmentation effect.

The invention discloses an RGB-D saliency object detection method based on a cascade improved network, and belongs to the technical field of image processing. Most existing RGB-D models directly aggregate the features from the CNN networks of different levels, and the noise and interference information contained in low-level features are easily introduced. The invention creatively provides a cascade improved structure, a saliency map generated by the features of a high-level part is used as a mask to improve the features of a low-level part, and then a final saliency map is generated by aggregating the improved low-level features. In addition, in order to eliminate the interference information of the depth map, the invention provides a depth enhancement module for preprocessing before thedepth features and the RGB features are mixed. According to the method, four evaluation indexes are used for carrying out experiments on seven data sets, and the results show that the method surpassesall current most advanced RGB-D saliency object detection methods.

The invention provides a rapid saliency detection method based on a multi-scale feature attention mechanism. The method comprises the following steps: firstly, processing an image through a deep convolutional network; obtaining features based on different convolution layers, including shallower layer features and advanced semantic features, then processing the obtained features based on differentconvolution layers, and finally inputting the processed shallower layer features and advanced semantic features into a decoder to generate a saliency detection graph. According to the invention, interference of most background features is eliminated; the calculation efficiency is improved; the background information is effectively inhibited; semantic information is better utilized for advanced semantic features by using pyramid expansion convolution, the features can be further refined by using a double-decoder junction, and the finally generated saliency map can completely highlight a saliency region in the image with a clear boundary and effectively suppress a background region.

The invention discloses a significance fusion method based on 3D fixation point predication of a picture. The method comprises saliency map generation and picture-based fusion. The saliency map generation comprises acquiring the saliency map of each original picture frame from an original video sequence. The picture-based fusion comprises the steps of constructing an energy function of an originalpicture according to the saliency map through objects of minimizing significance smooth restriction and minimizing significance difference between the original picture and the adjacent original picture; solving the energy function in the original picture, and obtaining a target significance map. According to the significance fusion method, the significance smooth restriction between a super-pixeland an adjacent super-pixel and a significance difference between the original picture and the adjacent original picture are considered so that the significance fusion method realizes relatively highsignificance in predicating different modal characteristics in a multi-modal characteristic fusion process.