161 results about "Quantization (image processing)" patented technology

The invention relates to the technical field of image processing, in particular to an unmanned aerial vehicle tracking method based on a twin neural network and an attention mechanism, which is applied to continuously tracking a visual single-target unmanned aerial vehicle. According to the method, weight redistribution of channel attention and space attention is realized by using two attention mechanisms, and the representation capability of the model on an unmanned aerial vehicle target appearance model is enhanced by using an attention model for template branches of a twin network; and search images are preprocessed in a multi-scale zooming mode, response graph calculations are separately carried out, inverse transformation of scale changes of the unmanned aerial vehicle in a picture issimulated in the mode, search factors capable of generating larger response values serve as scale inverse transformation of the unmanned aerial vehicle so as to correct the size of a frame used for marking a target, and the transverse-longitudinal proportion of the frame is not changed. According to the method, the tracking precision of 0.513 is obtained through testing (the average coincidence rate is used as a quantization precision standard), and compared with other leading-edge tracking methods, the method has the advantage that the performance is obviously improved.

The invention provides an image coding method which is applied to the field of computer image processing. The method comprises the following steps of: firstly, marking the region of interest in an image; performing wavelet transform of the original image; finding the wavelet coefficient corresponding to the pixel of the region of interest after the transform in a wavelet coefficient matrix; then,quantizing the background region and the region of interest according to different quantizing standards so that the coefficient of the region of interest is located in a higher plane; performing threshold quantization of the wavelet coefficient matrix nearby; and coding by a method based on SPIHT (set partitioning in hierarchical trees). The method provided by the invention guarantees the high compression ratio as well as the display quality of the region of interest of a user; and when transmitting or browsing an image in a wireless network with a relatively low rate, the method can enable the user to get the most important information in the image very soon.

The invention relates to a method and a device for increasing a video encoding speed, and belongs to the field of image processing. The method and the device are characterized in that a CPU (central processing unit) of a video compression encoder performs logical operation for infra-frame prediction or inter-frame prediction mode selection, loop filter and entropy encoding, a GPU (graphics processing unit) of the video compression encoder performs numerical operation for movement estimation, movement compensation, transformation and quantization and inverse transformation and inverse quantization, and the device comprises a CPU framework flow line and a GPU framework flow line which are parallel to each other. The method and the device have the advantages that a heterogenous system with a combination of a CPU framework and a GPU framework is comprehensively utilized for increasing video compressing and encoding speeds, so that videos are effectively compressed on the premise that the quality is guaranteed, the encoding elapsed time is greatly shortened, the integral performance of the encoder is greatly improved, the integral complexity of the encoder is assuredly unchanged basically, the instantaneity of the video encoder is greatly improved, and the method and the device can be applied to real-time encoding and decoding places.

Inputted image data is converted to M number of multi-value data having a lower resolution than the inputted image data, and after quantization processing has been performed for each of the M number of multi-value data, an image is printed by M number of relative movements (M-pass printing) that corresponds to the M number of quantized data. By doing so, when compared with the case in which a resolution reduction process is not performed, it is possible to suppress the number of pixels that become the object of quantization processing, and it becomes possible to output an image with no fluctuation in image density or density unevenness without a decrease in the processing speed.

The invention is suitable for the technical field of image processing and provides an image coding method, and an image decoding method and device. The image coding method comprises the following steps: decomposing an original image into a plurality of MCUs (Minimum Coding Units); obtaining a quantization DCT (Discrete Cosine Transform) coefficient of each MCU according to a DCT and quantization algorithm; when any MCU is judged to contain a sensitive area, encrypting the quantization DCT coefficient of the MCU and adding a default area identification into the encrypted quantization DCT coefficient to update the quantization DCT coefficient of the MCU as an identified and encrypted quantization DCT coefficient; and carrying out coding compression on the quantization DCT coefficient of each MCU and generating an encrypted and compressed image. The image coding method, and the image decoding method and device have the benefit that only the quantization DCT coefficient of each MCU containing the sensitive area is encrypted and is added with the area identification, so that the quality of an image in a non-sensitive area is not affected, and the encrypted image can keep higher compression ratio, so that the encryption security is also improved.

Provided is quantization processing that can reduce color development defect due to dot overlapping and can output an image with reduced granularity when the image is printed by using multiple kinds of colorants. To this end, dot arrangement information for a colorant for which dot arrangement is already determined among multiple kinds of colorants is acquired for a predetermined region of the image, and an evaluation value of each pixel included in the predetermined region is derived based on the arrangement information. In addition, for the predetermined region, a target value for a predetermined colorant for which dot arrangement is yet to be determined is derived based on the image. Then, whether or not to arrange a dot of the predetermined colorant in the predetermined region is determined based on the target value and the evaluation value.

The invention discloses an image processing method, device and equipment, and a readable storage medium. The method comprises the steps of obtaining a target image; inputting the target image into thequantized target deep neural network model for classification/detection to obtain an output result; processing the target image according to a strategy corresponding to the output result, wherein theprocess of obtaining the target deep neural network model through quantification comprises the steps of obtaining a floating point type deep neural network model obtained through pre-training; extracting weight features of the deep neural network model; determining a quantization strategy by utilizing the weight characteristics; and quantifying the deep neural network model according to a quantification strategy to obtain a target deep neural network model. According to the method, in the process of obtaining the target deep neural network model in a quantified mode, occupied resources are reduced, consumed time is shortened, and meanwhile model performance can be guaranteed, so that image classification/detection performance is guaranteed, and image classification processing performancecan be further improved.

A quantization method according to an aspect of the present invention includes the steps of quantizing a first image data by the use of a basic pattern and converting the first image data into a second image data that represents a binary or multi-level quantized pattern having a gray level smaller than that of the first image data. The basic pattern according to this aspect of the present invention presents high frequent occurrence of the basic tone patterns and a mostly uniform-distributed pattern of the clusters of different kinds of the basic tone patterns in the image with the long-distance autocorrelation (periodicity) of the basic tone patterns suppressed. Quantization by the use of this basic pattern provides a quantized pattern that reflects pattern characteristics of the basic pattern.

The invention discloses a method for identifying a metal solidification area with a texture. The method is characterized by comprising the following steps of: Step 100, correcting and segmenting a metal low-magnification sample image; Step 200, adopting a local self-adaptive threshold binaryzation processing algorithm based on grayscale mean square difference to extract binarization features of the texture in the sample image; Step 300, describing the direction characteristics of the texture by adopting a dot matrix type direction measurement algorithm, and carrying out direction filtering onthe texture features by adopting a direction filtering algorithm based on the levelness; Step 400, dividing the solidification area according to the texture direction information. According to the invention, a digital image processing technology is used to perform preprocessing, feature extraction, pattern recognition and the like on a solidification area image of a casting blank macrostructure soas to realize automatic recognition and quantization and improve the detection precision. Meanwhile, the labor intensity is reduced and the production efficiency is improved.

The invention relates to the technical field of image processing, and discloses a G-band chromosome HDR image reconstruction method, which comprises the training steps of a dequantization network model, a linearization network model and a detail reconstruction network model, and the step of reconstructing an HDR image by using the trained models. According to the method, three neural network models, namely a de-quantization network, a linearization network and a detail reconstruction network, are designed, and a training set is formed by utilizing an actual G-band chromosome HDR image and a corresponding real LDR image, so that the training of the three neural network models is completed; performing three steps of de-quantization, linearization and overexposure region detail reconstructionon a real LDR image by using the trained model, and finally completing a task of G-band chromosome HDR image reconstruction; according to the method, reconstruction of the HDR image is completed onlythrough one LDR image, so that G-banding chromosome characteristics are more obvious, subsequent analysis is facilitated, artifacts do not appear in the reconstructed HDR image, and the quality of the image is higher.

The invention discloses a self-adaptive bit network quantization method and system and an image processing method. The method comprises the following steps: acquiring a full-precision network model; obtaining a test data set under the applied classification task, and testing a classification result of the full-precision network model in the test data set; quantizing parameters of the full-precision network model by using a quantization function, and calculating standard errors of different parameters before and after quantization under a bit width condition to be selected; estimating the influence of the quantization of different parameters on the network performance, and obtaining the importance of the current parameter; solving a bit width allocation strategy with the highest accuracy under the target compression ratio; and quantizing the network according to a bit width distribution strategy to obtain a final network for image classification and target detection. According to the invention, the bit width and the quantization model of the network parameters under different compression rate requirements can be quickly given, meanwhile, high classification accuracy is ensured, and the universality of the quantization method is ensured.

The invention belongs to the technical field of digital image processing, particularly relates to an image compression method, system and device based on code rate control of sparse coding, and aims to solve the problems of low remote sensing image compression efficiency and low image reconstruction quality after compression due to the fact that the existing remote sensing image compression code rate is not easy to control and the code rate distribution is unreasonable. The method comprises the following steps: dividing a to-be-coded image into set sizes, and setting coding parameters; extracting an image block mean value and carrying out quantization and entropy coding; after the actual coding rate of the image is updated each time, comparing the actual coding rate with a set target coderate, and determining the next operation according to the comparison result; in each iterative encoding process, selecting image blocks with high complexity for sparse encoding, and jointly determining the number of the image blocks by the actual encoding rate of the current image, the set target code rate and the coefficient; and completing image coding at a set coding rate. The coding rate is accurate and controllable, distribution is reasonable, dynamic adjustment can be achieved, and efficient and high-quality compression of images can be achieved.

The invention relates to the field of artificial intelligence, and discloses an image processing method, which comprises the following steps: acquiring a first image; carrying out feature extraction on the first image to obtain at least one first feature map, wherein the at least one first feature map comprises N first feature values, and N is a positive integer; acquiring a target compression code rate, which corresponds to M target gain values, wherein each target gain value corresponds to one first feature value, and M is a positive integer smaller than or equal to N; processing the corresponding first feature values according to the M target gain values to obtain M second feature values; and performing quantization and entropy coding on the processed at least one first feature map to obtain coded data, the processed at least one first feature map comprising the M second feature values. Compression code rate control can be realized in the same compression model.

The invention provides an image processing method, device and system, which relate to the technical field of image processing. A compact and efficient training base dictionary can be built, and the compression rate and the quality of reconstruction can be effectively improved. In the specific technical scheme, a training sample image is acquired; the training sample image is blocked to obtain at least one training sample image block; each training sample image block is subjected to gray scale quantization to obtain a first gray scale image block of each training sample image block; and all first gray scale image blocks are subjected to dictionary learning to obtain a training base dictionary. The method, the device and the system are used for image compression and decompression.

The invention discloses an image model training method, an image processing method, a chip, equipment and a medium. The method comprises the steps of obtaining a sample image; wherein one sample imageis associated with one original network output; inputting the sample image into a preset neural network model containing the initial parameters, and performing symmetric quantization on the sample image to obtain a first quantization network output; determining a first cosine similarity between the first quantization network output and the original network output; performing optimization on the preset neural network model according to the first cosine similarity, a preset similarity threshold and a preset mixing precision quantification method, and obtaining preset network output of the preset neural network model after optimization; and determining a loss value between a preset network output and the original network output, and recording a preset neural network model as an image processing model when a convergence condition is preset for the loss value. According to the invention, the parameters of the preset neural network model are reduced, and the calculation rate of the preset neural network model is improved.

The invention provides an image processing method and device, electronic equipment and a computer readable storage medium, and the method comprises the steps: carrying out the calculation and statistics of the texture gradient of a to-be-processed image, and determining an intermediate-frequency texture region and a high-frequency texture region in the to-be-processed image; and enhancing the intermediate-frequency texture region and the high-frequency texture region in different enhancement modes to obtain a processed image. According to the method, intermediate-frequency textures and high-frequency textures are enhanced to different degrees, and the original intention of texture enhancement is better met, so that the intermediate-frequency textures are not lost by a relatively large quantization step size at a low bit rate as much as possible, and a prediction mode is stabilized by enhancing the textures; the high-frequency strong texture enhancement is weak, the excessive phenomenon after enhancement is reduced, and the subjective quality of the image can be improved.

The invention discloses a deep hash method based on metric learning, relates to the field of computer vision and image processing, and solves the problems that a comparison loss function of an existing deep hash method can only enable feature vectors of images of the same category before quantization to be close as much as possible, but cannot encourage the same symbol; the values of different types of images before quantization are far away as far as possible, but the symbols cannot be encouraged to be opposite; finally, the quantized hash code is poor in discriminability, and misjudgment andother problems are caused. According to the invention, a hash comparison loss function is constructed; sign bit constraint is carried out on the real numerical value feature vector before quantization, so that the hash code of the representative image obtained after the real numerical value feature vector before quantization is quantized by a sign function is more accurate, and the sign is constrained through two control functions of fsim (fi.fj) and fdiff (fi.fj), other parts in the expression are used for enabling the feature values of the same category of images to be close and the featurevalues of different categories of images to be far. According to the method, the classification precision is effectively improved, and the misjudgment rate is reduced.

Formed is a dither pattern that makes it possible to output an image with excellent dot dispersibility and reduced graininess regardless of controls after quantization processing. To this end, an extended pattern in which multiple divided pixels correspond to each of multiple pixels included in quantization data and one or more of the multiple divided pixels are determined as dot-arrangeable-pixels in which dots can be printed is generated. Thereafter, the thresholds of the dither pattern are set based on the extended pattern in which the dots are arranged to obtain predetermined dispersibility.