151 results about "Residual Blocks" patented technology

A method for estimating 3D posture of a human body combining densely connecting attention pyramid residual network and equidistance restriction is composed of a discriminative human body 2D posture estimation part and a generative human body 3D posture estimation part. Firstly, a 2D human posture estimation model is constructed. The 2D human posture estimation model includes attention pyramid residual blocks and a funnel sub-network composed of several attention pyramid residuals. The attention pyramid residuals are used for multi-scale image feature extraction, and the funnel sub-network is used to generate human joint thermodynamic map. In order to solve the problem that the environmental context information is not fully utilized, the attention mechanism and multi-scale analysis are combined to capture the environmental context characteristics. In order to solve the problem of gradient disappearance / gradient explosion, the dense connection network is combined with the above attentionmechanism to improve the discrimination of a feature map. Then the loss function is constructed and the equidistant constraint term is introduced to fit the 3D posture of a human body by minimizing the loss function. The method of the invention has obvious advantages in the human body 3D posture estimation task.

Disclosed is a method for encoding a decoding a video signal. In the procedure of encoding the video signal, when a frame temporarily simultaneous with a frame including a macro block of an enhanced layer which will obtain a prediction video does not exist in a base layer, the macro block is encoded based on difference values of residual data using corresponding residual blocks in a past frame and a future frame of the base layer which are residual data corresponding to image difference values and using a residual block for the macro block of the enhanced layer. In another embodiment, the macro block is encoded based on difference values of residual data using corresponding residual blocks in a past frame and a future frame of the enhanced layer and the residual block for the macro block. Accordingly, a residual prediction mode is applied for a macro block of an enhanced layer even if a frame temporally simultaneous with a frame of the enhanced layer does not exist in a base layer, thereby improve coding efficiency.

There are provided a method of context adaptive binary arithmetic coding (CABAC) and an apparatus of the same. The CABAC method includes generating a plurality of syntax elements respectively corresponding to a plurality of predetermined sized residual blocks included in an input block; grouping flags of each of the residual blocks which indicate whether a non-zero transform coefficient exists among the syntax elements; and context adaptive binary arithmetic coding the grouped flags. The complexity of context adaptive binary arithmetic coding is reduced, thereby performing context adaptive binary arithmetic coding more quickly and easily.

Disclosed herein is a local dimming driving method and device of an LCD device, which is capable of compensating for a luminance deviation between blocks. The local dimming driving method of the LCD device includes divisionally driving all of a plurality of blocks of a backlight unit using a maximum luminance signal and measuring luminance per block, setting one of the plurality of blocks as a reference block, detecting luminance deviations between the reference block and the residual blocks, and setting an offset value per block for compensating for the detected luminance deviations per block, analyzing an input image in units of blocks corresponding to the plurality of blocks of a backlight unit respectively, detecting a representative value per block, and determining a dimming value per block according to the representative value per block, correcting the dimming value per block using the offset value per block, and controlling the luminance of the backlight unit on a block-by-block basis using the corrected dimming value per block.

Disclosed herein is a local dimming driving method and device of an LCD device, which is capable of compensating for a luminance deviation between blocks. The local dimming driving method of the LCD device includes divisionally driving all of a plurality of blocks of a backlight unit using a maximum luminance signal and measuring luminance per block, setting one of the plurality of blocks as a reference block, detecting luminance deviations between the reference block and the residual blocks, and setting an offset value per block for compensating for the detected luminance deviations per block, analyzing an input image in units of blocks corresponding to the plurality of blocks of a backlight unit respectively, detecting a representative value per block, and determining a dimming value per block according to the representative value per block, correcting the dimming value per block using the offset value per block, and controlling the luminance of the backlight unit on a block-by-block basis using the corrected dimming value per block.

The invention discloses a video super-resolution reconstruction method based on multi-memory and mixed loss, which comprises two parts of an optical flow network and an image reconstruction network. In the optical flow network, the optical flow between the current frame and the reference frame is calculated for the input multiple frames, and the optical flow is used as motion compensation to compensate the current frame to be similar to the reference frame as much as possible. In the image reconstruction network, the compensated multi-frames are sequentially input into the network, and the network adopts multi-memory residual blocks to extract image features, so that the following input frames can receive the feature map information of the previous frames. Finally, the output low-resolution feature image is amplified by sub-pixel, and added with the image amplified by bi-cubic interpolation to obtain the final high-resolution video frame. A hybrid loss function is used to train the optical flow network and the image reconstruction network simultaneously in the training process. The invention greatly enhances the feature expression ability of the inter-frame information fusion, andcan reconstruct the high-resolution video with rich details.

The invention discloses a human posture estimation method based on deformation convolution, which solves the technical problem of estimating human posture from images. The method comprises the following steps: obtaining a training image; making a joint heat map; constructing a deformation convolution forward propagation module; constructing residual blocks and constructing multi-scale hourglass networks with deformed residual blocks; training and stacking of multi-scale hourglass networks with deformed residual block network structures; acquring that result of the human posture estimation. Theinvention uses deformation convolution and improves the internal connection mode of the hourglass network, a stacked multi-scale hourglass network with deformed residual blocks is constructed, For asingle natural color image, the human pose can be estimated more accurately by extracting and organizing the image features in the complex scene of distorted limbs or special posture, refraction or reflection of light, large variation of human scale, and occlusion. Used for human-computer interaction in multiple scenarios.

The invention discloses a CT lung nodule detection system based on 3D-Unet. The system includes a CT image input and preprocessing module, a candidate nodule detection module and a false-reporting elimination module which are sequentially connected. The CT image input and preprocessing module is used for reading a chest CT graph, obtaining spacing and origin information of the CT graph, and carrying out lung volume segmentation on the CT graph. The candidate nodule detection module is used for inputting the image, of which preprocesing is completed, into a Unet network, and obtaining a location of a candidate lung nodule. The Unet network includes sixteen residual blocks, two path units, four maximum-pooling units, one preparation unit, one probabilistic neuron failure unit and one outputunit. According to the system, automatic detection of the lung nodule is realized, problems of high doctor workloads and high misdiagnosis rates in lung nodule diagnosis are solved, an effect of detecting the lung nodule by the 3D-Unet is realized, more contextual semantic information in the CT image is utilized, and detection accuracy is higher.

The invention discloses a sign language recognition method based on a space-time attention mechanism. The method comprises the following steps: firstly, sampling a sign language video into a continuous sign language sequence with a uniform length as input of a model; and then, inputting the video frame sequence into a spatial attention network formed by 3D residual blocks, so that the network canautomatically pay attention to a salient region in a space. Analyzing the extracted convolution features through a ConvLSTM convolution long-short-term memory network, extracting long-time sequence features, and distributing time attention weights of different video frames to generate feature representation of the video; and finally, the generated feature representation passes through a Softmax classifier, and classification categories are output in a vector form. According to the invention, the interference of redundant information on recognition can be reduced, and the recognition accuracy is improved.

A circuit generally including a first module, a second module and a third module is disclosed. The first module may be configured to (i) generate a plurality of parsed residual blocks by parsing a plurality of 4×4 CAVLC (context-based adaptive variable length coding) residual blocks received in an input signal and (ii) generate a plurality of metric signals resulting from the parsing of the 4×4 CAVLC residual blocks. The second module configured to generate a plurality of scanning position signals based on the metric signals. The third module configured to generating an 8×8 CABAC (context-based adaptive binary arithmetic coding) residual block in an output signal by up-sampling the parsed residual blocks based on the scanning position signals.

The invention discloses a video image coding / decoding method based on geometric partitioning. The coding method comprises the following steps: firstly, a video image is partitioned into a plurality of rectangular coding blocks, and the rectangular coding blocks are geometrically partitioned; secondly, motion estimation is carried out for each geometrically partitioned irregular block, each irregular block obtains an own residual block, and the residual blocks corresponding to the irregular blocks are combined to a rectangular residual block; thirdly, the coordinates of pixels in the rectangular residual block are rearranged by utilizing geometrically partitioned boundary direction information; and lastly, two-dimensional orthogonal transformation is carried out for the rearranged rectangular residual block, a transformation coefficient is quantized and entropy coding is processed, in addition, geometrically partitioned information and rearranged information are incorporated into a code flow. The decoding process and the coding process are inverse. Since the geometrically partitioned boundary direction information is utilized and the pixels of the rectangular residual block are rearranged when the coding is processed, so that after orthogonal transformation, a high-frequency nonzero coefficient is reduced, and compression efficiency is improved.

The embodiment of the invention discloses a method, a device and a system for processing an image residual block, relating to the technical field of video coding and decoding and capable of increasing the video compression coding performance and reducing the complexity in coding and decoding calculation. The method for processing the image residual block, provided by the embodiment of the invention, comprises the steps of: performing transformation coding on residual blocks of blocks to be coded by adopting a transformation base to obtain transformation coefficients of the residual blocks; determining a scanning sequence of the transformation coefficients of the residual blocks according to the transformation base and the grain direction information of the blocks to be coded; and scanningthe transformation coefficients of the residual blocks according to the scanning sequence to obtain the scanned transformation coefficients.

A circuit generally having a first module, a second module and a third module is disclosed. The first module may be configured to (i) generate a plurality of parsed residual blocks by parsing an 8×8 CABAC (context-based adaptive binary arithmetic coding) residual block received in an input signal and (ii) generate a plurality of metric signals resulting from the parsing of the 8×8 CABAC residual block. The second module may be configured to generate a scanning position signal based on the metric signals. The third module may be configured to generating a plurality of 4×4 CAVLC (context-based adaptive variable length coding) residual blocks in an output signal by sub-sampling the parsed residual blocks based on the scanning position signal.

The disclosure relates to an image segmentation method and apparatus, and a computer program thereof. The image segmentation apparatus according to the disclosure may include an encoding module configured to obtain a plurality of feature maps having different resolutions with respect to an input image, using an artificial neural network including one or more residual blocks, and a decoding module configured to generate one prediction map from pairs of adjacent feature maps among the plurality of feature maps in bracket-shaped manner, wherein the decoding module performs one or more decoding rounds, each decoding round includes one or more ATF modules that generate a combined feature map using a high-resolution feature map having a higher resolution among an adjacent feature map pair constituting the feature map generated in the previous round and a low-resolution feature map having a lower resolution thereof, and the decoding round is repeatedly performed until the one prediction map is generated. According to the disclosure, it is possible to combine local information and global context in a balanced manner.

The invention provides a power system short-term load prediction method based on a time convolution network, and the method comprises the steps: collecting historical load data, and carrying out the preprocessing of the data; constructing causal expansion convolution models, and respectively inputting the preprocessed data into two different causal expansion convolution models for convolution processing; connecting the two results processed by the causal expansion convolution model to form a residual block; stacking the residual blocks to obtain a time convolution network; and carrying out full convolution layer calculation by using a time convolution network to predict a future power load demand. The invention provides a short-term load prediction method for a power system. A convolutionmodel is expanded through causality; causal convolution processing and extended convolution processing are carried out on the data, then residual convolution processing is carried out, learning objectives and difficulty are simplified, and finally full convolution layer calculation is carried out by using the time convolution network, so that the time and hardware requirements required in the prediction process are reduced, and the precision is equivalent to that of a mainstream recurrent neural network.

The invention discloses a video decoding method which comprises the steps that a converting unit and a prediction unit obtained after video compression bitstream decoding are converted into M * M residual blocks and M * M prediction blocks respectively, wherein M is equal to 16 or 8 or 4, a video compression bitstream comprises one of an HEVC bitstream, a VP9 bitstream and an AVS2.0 bitstream; according to the residual blocks and the prediction blocks, reestablished image blocks are obtained; and the reestablished image blocks are subjected to filtering, and decoding reestablished images are obtained. The invention further discloses a video decoder. Module development of the multi-protocol video decoder is lowered, and the area and the power consumption of the video decoder are lowered.

The invention discloses a high-resolution remote sensing image classification method based on a novel feature pyramid depth network. The method comprises the following steps: firstly, designing a novel deep convolutional neural network on the basis of a ResNet34 network model; secondly, inputting the high-resolution remote sensing image into the network for training, and taking the output of eachmain convolution layer of ResNet34 as a subsequent input feature; fusing the input features by using a feature pyramid network to form new features; then, fusing the new deep-layer features and the new shallow-layer features to serve as inputs of an upper branch and a lower branch, and designing two residual blocks and a global average pooling layer on each branch; and fusing the features of the upper and lower branches and then sending to a full connection layer, and classifying the remote sensing images through a SoftMax layer. According to the invention, feature extraction and fusion are carried out on the high-resolution remote sensing image based on the deep learning theory, so that each feature is enhanced. The new features are fused again and then sent to the upper branch and the lower branch to learn image-level features, and experiments prove that the proposed method can achieve a good classification effect.

The invention provides an adversarial neural network-based watermark removal method, device and apparatus and a readable medium. The method comprises a training step of training an adversarial neuralnetwork on the basis of sample images so as to obtain an adversarial neural network, and a removal step of inputting images with watermarks into the adversarial neural network so as to obtain images without watermarks. According to the method, device and apparatus and the readable medium, the adversarial neural network with residual blocks is designed; the adversarial neural network comprises a generator network and a discriminator network; through the generator network in the adversarial neural network specific watermarks can be rapidly and efficiently removed in batches; and through the discriminator network, judgement is carried out determine whether the watermark removal effect achieves requirements or not, and if the judging result is positive, images are output.

The invention belongs to the technical field of computed tomography image processing. According to the specific technical scheme, the CT image super-resolution reconstruction method based on the generative adversarial network comprises the following specific steps: 1, establishing a dense connection relationship among different residual blocks based on a multi-stage dense residual block generatornetwork; 2, adding a bottleneck layer to the front end of each dense residual block; 3, optimizing the global network by adopting the Wasserstein distance loss and the VGG feature matching loss; 4, arranging a multi-path generator based on the sequence from thick to thin; 5, generating an image based on conditional expression generative adversarial learning; 6, reconstructing a CT image super-resolution reconstruction framework of the generative adversarial network based on multiple paths of conditions from coarse to fine; 7, reconstructing a loss function. According to the method, network redundancy is reduced, feature multiplexing among different residual blocks is realized, the maximum information transmission of the network is realized, the feature utilization rate is improved, and thereconstructed image quality is greatly improved.

The invention relates to a detection method and a detection system for foreign matters in an infusion bottle. The detection method comprises the following steps of: rotating a to-be-detected infusion bottle; photographing the bottle by utilizing a camera so as to obtain a multi-frame image; processing the image to obtain feature information of blocks in the bottle; removing the blocks formed by interferences of bubbles and lights; and comparing the feature information of the residual blocks with a detection standard, and judging whether the foreign matters exist, so as to judge whether the infusion bottle is qualified. The detection system comprises an intermittent rotating device and the camera, wherein upper compression devices and lower compression devices for fixing the relative positions of the to-be-detected infusion bottle and the intermittent rotating device are arranged on the intermittent rotating device, and friction wheels for driving the upper compression devices and the lower compression devices to rotate are arranged at the outer side of the intermittent rotating device. By utilizing the detection method and the detection system, the automatic detection on the foreign matters in the infusion bottle can be realized, and the problems of low efficiency, high labor intensity, easiness in generation of quality fluctuation and the like in the conventional detection method are solved; a reliable detection result can be obtained, and the export quality of products is effectively guaranteed.

The invention provides an underwater image denoising method based on a generative adversarial network, and the method comprises the steps: inputting an underwater image into a generation network composed of a plurality of residual blocks, and obtaining a feature mapping graph; mapping output feature mapping graph and clear water noiseless label image through VGG-19 network to obtain depth featurespace, calculating the perception cost of the feature mapping graph and a clear water noise-free label image in a depth feature space, calculating the perception cost, inputting the feature mapping graph output by a generation network into an adversarial network, and finally inputting a noise-containing underwater image into the generation network after training is completed, and outputting the noise-free image as a processed noise-free image. According to the method, the confrontation mechanism is introduced, the denoising effect is obvious, particularly, the edge texture information in the image can be effectively reserved and even enhanced through the method, and the better visual effect and imaging quality are achieved.

The invention discloses a data noise suppressing method based on a residual block full convolutional neural network. A training set and a test set for suppressing earthquake noise by using a deep learning method are selected from the same data set, so that the generalization performance of a model is limited. Specific to the problem of generalization, a double residual block is fused on the basisof a Unet network according to the design principle of a network structure in order to enhance the capturing performance of the network for random noise. The data noise suppressing method is established on an end-to-end coding-decoding network structure, noise-containing earthquake data is taken as input, and the substitutive characteristics of the random noise are extracted by a plurality of convolution layers and residual blocks to construct coding; and then a plurality of deconvolution layers and residual blocks are used for constructing decoding, so that the output of the network is the noise-suppressed earthquake data. Compared with the existing earthquake data denoising method, the data noise suppressing method has the advantages that the double residual block is fused to perform secondary digestion and learning on the extracted random noise characteristics, so that the substitutive characteristics of the noise are learned more fully. Thus, the data noise suppressing method has remarkable advantages on the aspect of generalization, and can effectively suppress the random noise and protect effective signals.

The invention discloses a super-resolution reconstruction method based on a dirac residual deep neural network. The super-resolution reconstruction method comprises the steps: a network inputs a low-resolution image, learns image features through a plurality of dirac parameterized residual blocks, and employs sub-pixel convolution to reconstruct a high-resolution image; the network is divided intoan upper part and a lower part, and the upper part obtains high-frequency features of the LR through the deep dirac residual network, and then carries out reconstruction through sub-pixel convolution; the lower part reconstructs an image directly by sub-pixel convolution of low frequency features of the LR; and a reconstructed HR image is output by combining the two reconstruction structures. According to the super-resolution reconstruction method, the residual layer is improved through weight parameterization; the convolution feature dimension before the activation function Relu is reduced;and the convolution feature dimension after the activation function is increased. Meanwhile, the convolutional features of the input image and the features of residual network learning are combined for reconstruction, and a better SR effect is obtained under the condition that the network depths are the same.

The invention relates to a real-time fine granularity scalable encoding method based on H.264. According to the overall situation of current residual frames and the local distribution of current residual blocks, the invention finds singular points from the global analysis, makes the bit plane number wholly equilibrated through downward shift of the singular values, and makes corresponding upper shift at the decoding end; the invention determines all-zero blocks from the local analysis. The invention will use a 4*4 whole number transformation in the reinforced layer conversion fraction during encoding. Compared with the general FGS method, the complexity in the invention is obviously reduced, the video quality is higher, and the overall PSNR (peak signal noise ratio) change is smoother. Evidenced by experiments, compared with the FGS method of MPEG-4, when keeping a close code rate, the invention can increase the average brightness PSNR by 0.37dB, and increase the average encoding speed by 13.86fps, namely by 97 percent.

The application discloses a video compression encoding-and-decoding method on the basis of weighting quantification. The encoding-and-decoding process comprises that an attribute component of an image to be encoded is divided into multiple attribute blocks; the attribute blocks are predicted so that residual blocks are obtained, and a transformation coefficient of each frequency point in the residual blocks is obtained after transformation; one default matrix is selected, and weighting calculation is performed on initial quantization step length of the frequency points so that a weighting quantification step length matrix is obtained; and the transformation coefficient is quantified by using the weighting quantification step length matrix so that quantification blocks are obtained, then the quantification blocks are written-in a code stream and other information written-in the code stream is confirmed according to values of the quantification blocks. The application also discloses a video compression encoder-decoder on the basis of weighting quantification. A preset weighting quantification matrix set comprises multiple default matrixes so that transformation coefficient matrixes under different situations can be effectively weighted and quantified, and thus code rate required by encoding can be effectively reduced without reducing subjective quality.

The invention provides a license plate deblurring method for video detection. The method comprises the following steps: S1, collecting a plurality of sets of license plate data combinations includingblurred license plate images and corresponding clear license plate images, and dividing the combinations into training data sets, verification data sets and test data sets; S2, collecting a pluralityof license plate data combinations including blurred license plate images and corresponding clear license plate images. S2, designing a generation countermeasure network model of de-motion blur, the generation network includes two convolution blocks of step size, seven inversion residual blocks of Mobilenet V2 and two transposed convolution blocks; 3, training that generated antagonism network, and put the data set obtained in the step S1 into the generated antagonism network model for training; S4, on the basis of the deblurring generation countermeasure network model trained in the step S3,the license plate image blurred by the motion is input, and the output data is the generated clear license plate image. The invention can realize the license plate clarity, so as to quickly determinethe identity of the suspect, effectively help the criminal investigator to solve the case as soon as possible, and the generation network composed of Inverted Residual Block of Mobilenet V2 can betterextract the high-dimensional features.