242 results about "Normalization (image processing)" patented technology

Apparatus and methods for learning in response to temporally-proximate features. In one implementation, an image processing apparatus utilizes bi-modal spike timing dependent plasticity in a spiking neuron network. Based on a response by the neuron to a frame of input, the bi-modal plasticity mechanism is used to depress synaptic connections delivering the present input frame and to potentiate synaptic connections delivering previous and / or subsequent frames of input. The depression of near-contemporaneous input prevents the creation of a positive feedback loop and provides a mechanism for network response normalization.

The invention belongs to the field of mode recognition, computer vision and image processing, particularly relates to an iris image segmentation and positioning method, system and device based on deeplearning, and aims to solve the problem of low iris recognition precision in a non-controllable scene. The method comprises the steps that a to-be-processed iris image is acquired; four mapping images are generated by adopting a multi-task neural network model, wherein the four mapping images respectively correspond to a pupil center, an iris inner boundary, an iris outer boundary and an iris segmentation mask; the iris segmentation mask mapping graph is processed by adopting threshold segmentation to complete iris segmentation; the pupil center position is predicted according to the geometrical relationship between the pupil center and the iris mask; the mapping graph is de-noised and calculated by utilizing a geometrical relationship among the pupil, the iris and the sclera to obtain iris inner and outer circle parameters and finish iris positioning. According to the method, the iris image acquired in the non-controllable environment can be effectively segmented and positioned, a good foundation is laid for subsequent normalization and recognition, and the iris recognition precision in the non-controllable environment is improved.

The invention relates to the technical field of computer image processing, specifically to a remote-sensing image super-resolution reconstruction method based on a generative adversarial network. Theremote-sensing image super-resolution reconstruction method comprises the following steps: constructing a remote-sensing image super-resolution reconstruction model consisting of a generator network and a discriminator network; introducing a scene constraint sub-network into the generator network to solve the problem of scene change, introducing an edge enhancement sub-network to solve the problemof edge transition smoothness of a generated image, introducing TV loss for noise suppression, and introducing a content fidelity to deal with the problems of instability and gradient disappearance in the training process; and introducing spectrum normalization into a discriminator network to control the performances of a discriminator so as to promoting better learning of the generator. The remote-sensing image super-resolution reconstruction method provided by the invention has the following advantages: a high-quality high-resolution remote-sensing image can be generated based on a low-resolution remote-sensing image; the precision of the low-resolution remote-sensing image in classification detection is effectively improved; the problems of edge transition smoothness and scene change in super-resolution of the remote-sensing image are solved; meanwhile, the problems of training instability and gradient disappearance under the GAN network are solved.

The invention discloses a SF6 gas detection method base on infrared image processing. The method comprises the steps of: S1, SF6 gas original infrared image is obtained with a passive infrared detection manner, and the original infrared image is subjected to non-uniformity correction, such that a corrected infrared image is obtained; S2, the corrected infrared image is subjected to normalization treatment, and the processed image is converted into a grayscale image; S3, the grayscale image is subjected to image filtering and grayscale linear transformation, such that a target image is obtained; and S4, the target image is subjected to real-time displaying. With the technical scheme provided by the invention, problems such as large infrared image noise and difficulty in real-time displaying in existing passive infrared SF6 gas leakage detection method can be solved.

The invention provides a mining electric locomotive passerby monitoring method based on image processing and an alarm system and belongs to the field of safe transportation of coal mines. The system comprises a video collection module, an image processing module and an acousto-optic alarm module, wherein an infrared video camera for video collection is used for collecting images at the front of an electric locomotive; the image processing module comprises image preprocessing, rail identifying and fitting and passerby identifying; the image preprocessing is based on a genetic algorithm, image self-adapting correction combined with a normalization non-complete Beta function and an image binarization method based on a pulse coupling neural network; the rail identifying and fitting is based on fuzzy edge detection fast algorithm with an improved threshold of a genetic algorithm and a heuristic connection method; the passerby identifying is based on a pulse coupling neural network image binarization method realized by virtue of an FPGA (Field Programmable Gate Array); and the acousto-optic alarm module comprises an acousto-optic alarm and a control circuit. The system can be used for effectively identifying passersby at the neighborhood of a rail, judging and pre-alarming the position of the passersby and eliminating the potential safety hazard of the electric locomotive in a transportation process.

The invention discloses a method for enhancing and extracting high stars and other sliding fixed stars in the view field of a telescope. Aiming at the problems that high background stars and other sliding fixed stars are low in detection rate and not high in positioning precision in the celestial positioning process, the astronomy and an image processing technique are combined, the sliding tracks of the fixed stars are calculated by using astronomical information, and reliable enhancement and high-precision extraction of the high stars and other sliding fixed stars are realized by using a background equalization method, a quasi-randomized orbital determination method, a noise normalization method, a double-threshold orbital determination correlation method, a two-step centroid extraction method and other image processing techniques, so that technical support is provided for the celestial positioning. The method has the benefits that the false detection rate and the misdetection rate are extremely low; the influence from interference factors among fixed stars, of other celestial bodies and background noise and the like can be better overcome; weak high stars and other fixed stars are extracted; information about the number, the size, the position, the precision and the like of the fixed stars are synchronously obtained; besides, fewer parameters are set and the set parameters are simple, and real-time processing can be realized.

The invention belongs to digital image processing and chemical analyzing domain, which is characterized in that it uses digital camera or video camera to sight the lathe on the light condition with different wavelength; it dose lens distortion correction and noise reduction process to the image; it uses the character of point sampling time point equidistribution and the rule of vertical distance five equilibrium to ascertain the control point; it uses bracketing method to construct image background by the control point; it exorcizes the background image from the adjusted thin layer image and dose brightness normalization to each image element point of the constructed background image; it divides the thin layer image and dose monomer to the grey of each area by the dividing result.

An image processing apparatus includes: a normalization unit that generates a plurality of normalized images by correcting, based on different imaging conditions, brightness of a plurality of images at different brightness obtained by imaging a same subject with the different imaging conditions; and an image combining unit that generates a combined image by combining the plurality of normalized images. When the normalized image corresponding to a reference image that is a brightest image among the plurality of images is defined as a reference normalized image, the image combining unit combines the reference normalized image and a non-reference normalized image that is the normalized image other than the reference normalized image to bring gradation of pixel of the combined image close to gradation of pixel of the reference normalized image if the gradation of pixel of the combined image is smaller than the gradation of pixel of the reference normalized image.

The invention discloses a multi-parameter tumor blood vessel normalization detection system and detection method. The system comprises a light source module, an imaging module, a blood flow detectionand analysis module, a blood oxygen detection and analysis module, a blood vessel normalization judgment module and a display module. According to the method, a two-dimensional blood flow image is reconstructed through an algorithm, blood vessel morphology characteristics are extracted through an image processing method, a two-dimensional tumor blood oxygen image is calculated through a laser image and a narrow-band green light image, tumor tissue hypoxia degree information is provided, and through a deep neural network discrimination method and combination with blood flow image characteristics and blood oxygen image characteristics whether tumor vessels are normalized or not can be judged. The method has the advantages of being multi-parameter, non-invasive, non-radiative, intelligent indiagnosis and the like, whether tumor blood vessels are in a normal window period or not is intelligently judged by monitoring tumor blood supply, oxygen supply and blood vessel forms, and a basis isprovided for making a more reasonable anti-tumor combined treatment scheme.

The invention provides a low-dose CT image denoising method and system based on self-supervised learning, and belongs to the technical field of image processing, and the method comprises the steps ofcarrying out the preprocessing of an obtained CT image, and carrying out the normalization of all pixel values of the preprocessed CT image; replacing a part of pixels of the normalized CT image by adopting a preset mask with the same size as the CT image; inputting the CT image replaced by the preset mask into a trained denoising neural network model to obtain a corresponding denoised image. According to the invention, in a network model training denoising network process; compared with the prior art, the method does not need paired LDCT images and high-dose CT images, deduces the high-dose CT pixel values of the replaced pixels through the unreplaced pixels, can greatly reduce the data collection cost, and completes the denoising task of the LDCT images under the condition of no high-dose CT images.

The invention provides a lung lobe segmentation method and device based on a UNet network and a computer readable storage medium, and relates to the field of lung lobe image processing. The lung lobesegmentation method comprises the following steps: acquiring lung CT image data from an image input device; carrying out normalization processing on the input lung CT image data; screening out an intra-pulmonary region and an extra-pulmonary region from the processed image data by utilizing a 2D UNet network, and taking the intra-pulmonary region as a lung region candidate region; dividing the lung region candidate region into five lung lobe mask regions by using a 3D UNet network to obtain regions of a left upper lobe, a left lower lobe, a right upper lobe, a right middle lobe and a right lower lobe; respectively carrying out morphological processing on the five lung lobe mask regions to obtain a final lung lobe segmentation result; and storing the lung lobe segmentation result in a memory or outputting and displaying the lung lobe segmentation result on a screen of a display. The lung lobe is quickly and accurately extracted through the UNet network, the lung cancer position is positioned, and guidance is provided for doctors to diagnose and treat lung cancer.

The invention belongs to the technical field of medical image processing, and particularly relates to a granulocyte picture fine-grained classification and recognition system based on deep learning. The system comprises a positioning module and a classification module; the positioning module performs feature extraction on an input granulocyte picture by using a Hourglass network model, positions all cells in the granulocyte picture, cuts out the positioned cells, leaves a single complete cell, and performs size normalization processing on all cut-out cells; the classification module is used for classifying the granulocytes positioned by the positioning module by adopting the constructed deep learning classification model; clinical doctors are assisted in accurately and efficiently completing granulocyte classification, identification and counting tasks, errors caused by subjectivity are reduced, the workload of the doctors is reduced, and the doctors are assisted in making disease judgment. According to the system, cell classification under unbalanced data and fine-grained classification among granulocytes can be effectively solved, and the network classification and recognition effect is improved.

The invention belongs to the field of light field image processing, and particularly discloses a micro-lens light field camera full-focus image generation algorithm based on a blocking effect. According to the algorithm, fusion is directly carried out based on 4D light field data; when the data fusion rule is established, the focusing degree evaluation is completed by calculating the blocking effect of light field macro pixels formed by light energy integration in different directions at the same spatial position, and finally, the fused 4D light field is integrated to obtain a full-focus image. The effectiveness of the algorithm is demonstrated through experiments. The fusion image obtained based on the angle information and the fusion image obtained based on the space information have thesame visual effect, the perception definition value of the image obtained through the algorithm is the highest from the average quality evaluation comparison result of the fusion image, and the phaseconsistency value and the normalization mutual information value are only second to those of the BF algorithm.

The invention discloses a non-intrusive load identification method based on a V-I trajectory diagram and a neural network. The method comprises steps of home-entry voltage and current data and activepower data being acquired in real time; judging whether a load switching event exists or not and whether a load operation state reaches a stable state or not according to the change of the active power; obtaining voltage and current data and power data of the load according to the steady-state voltage and current data before and after the event; a V-I track being converted into an RGB color imagecontaining information such as voltage and current phase difference and power by adopting a simple image processing technology. After an RGB color image is obtained, normalization processing is carried out; and load identification is carried out through a pre-trained convolutional neural network. Compared with the prior art, the method is advantaged in that the steady-state characteristics of theload are fully extracted through the convolutional neural network, and the neural network model can be directly operated on the embedded equipment without depending on the operation support of a server.

The invention belongs to the field of image processing and provides a reference-free tone mapping image quality evaluation method based on color statistics and information entropy. According to the method, the technical proposal adopted by the invention is that an image quality evaluation method without reference tone mapping based on color statistics and information entropy is provided. The method comprises the following steps: first, data preparation; 2, extract color statistical features, mainly comprise that following three steps: (1) decomposing color channel; 2) carry out division normalization operation on that channels in (1) to obtain respective color normalization coefficient; (3) parameter fitting; 3, feature extraction of information entropy; Step 4, quality prediction. Featureextraction of the test set, input to the trained model, predict the quality of the test chart score. The invention is mainly applied to image processing occasions.

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 discloses a smiling face identification and encouragement system, and belongs to the technical field of image processing and artificial intelligence. The smiling face identification and encouragement system is implemented through the following steps: collecting basic expression samples; automatically starting videos at startup; collecting the expressions in real time; pre-processing the expression images; identifying the expression images; carrying out comprehensive statistics, analysis and judgment; summing up the phased expressions; and encouraging the staff to relax. According to the technical key point of the invention, a plurality of expression characteristics represented through vectors are matched with corresponding key characteristics in several pre-shot user typical expression samples to compare the similarity; the difference and geometric shape change between the Euclidean distance at a captured expression key characteristic and the Euclidean distance at a preset typical expression characteristic are judged; and after normalization, a statistic shape analysis method is used for judging whether the expression in the phase is anxiety or happiness according to the expression distribution probability of the staff in a certain time period. The system has the benefit of encouraging the staff to feel anxious less and smile more so as to improve the working efficiency and the living quality.

The invention discloses an infrared weak and small target detection method based on space-time joint local contrast, and relates to the field of infrared image processing and weak and small target detection. The method comprises the following steps: S1, constructing a sliding window with the size of 3*3, traversing a kth frame of image of an original sequence image, and obtaining a spatial domainlocal contrast response graph of the kth frame of image through spatial domain filtering; S2, calculating a variance value St of a continuous frame image, and obtaining a time domain local contrast response graph of the kth frame of image through time domain filtering in combination with variance value images of three adjacent frames of images; and S3, performing normalization processing on the time domain detection result and the space domain detection result, and combining the time domain detection result and the space domain detection result in a multiplicative fusion mode to obtain a space-time joint local contrast response of the kth frame of image. According to the method, spatial information and time information are fully utilized, the problems of low infrared weak and small targetdetection precision, scene robustness and the like caused by an existing method are solved, the detection performance and the low false alarm rate in infrared weak and small target detection under a complex background are improved, and the robustness of an algorithm is improved.

The present invention relates to the medical apparatus and instruments technology field, and aims to provide a full-automatic brain tumor image segmentation technology which is combined with the MRI to generate images, executes an FCM algorithm via a software and realizes the image processing. Meanwhile, the full-automatic brain tumor image segmentation technology is combined with the advantages of a collaborative fuzzy clustering algorithm, so that the FCM algorithm can be applied to the MRI brain tumor images more effectively. The technical scheme adopted in the present invention is that: the possibility FCM algorithm-based MRI tumor image segmentation method and system is composed of an MRI device and a computer, the MRI device generates images to input to the computer, and the computer is equipped with the following modules: a denoising module used for removing the noise and the brain tissues in the MRI brain tumor images and carrying out the normalization to prepare for the next step; a histogram statistics module; an FCM initial segmentation module; a matrix adjusting module; an FCM segmentation image module which uses the collaborative possibility FCM algorithm to obtain the segmentation images. The possibility FCM algorithm-based MRI tumor image segmentation method and the possibility FCM algorithm-based MRI tumor image segmentation system of the present invention are mainly applied to the image segmentation occasions.

The invention discloses a training method and device for a student model for image processing, and belongs to the technical field of knowledge distillation, and the method comprises the steps: obtaining the parameters of a classification layer in a teacher model which is obtained through the classification training of target objects in a plurality of image samples, initializing parameters of a classification layer in a to-be-trained student model by utilizing the obtained parameters, inputting at least part of the image samples into the student model for classification, and adjusting parameters of a target layer in front of the classification layer in the student model according to a classification loss value of the student model, and enabling the image features of each type of target objects learned by the target layer in the student model to approach the image features of the type of target objects learned by the target layer in the teacher model, and ending the training until it isdetermined that the classification error of the student model is smaller than a set error, wherein the teacher model and the student model each comprise a convolution layer, a classification layer anda normalization layer which are connected in sequence, and the normalization layers of the teacher model and the student model use the same normalization function.

The embodiment of the invention provides an image processing method and device, equipment and a storage medium. The image processing method comprises the following steps: acquiring a first image and asecond image to be processed; splicing the first image and the second image to form a spliced third image; and extracting the content features of the third image and the style features of the secondimage through a pre-trained style conversion network model, performing normalization processing on the content features of the third image according to the style features of the second image to obtainprocessed target image features, and generating and outputting a corresponding target image according to the target image features. According to the embodiment of the invention, the first image and the second image can be spliced; and the content characteristics of the third image and the style characteristics of the second image can be deeply fused, so that the styles of the spliced third imageand the spliced second image are highly unified, and the method is more suitable for images of different styles, can be applied to the field of intelligent mapping, and realizes real-time and rapid intelligent mapping.

The invention discloses a lung parenchyma CT image segmentation method based on a weighted full convolutional neural network, and belongs to the field of medical image processing. The method comprisesthe following steps: selecting a public lung data set for preprocessing, and extracting a lung parenchyma boundary in a labeled image as a semantic category; designing an improved network structure based on a standard full convolutional neural network framework, and establishing an overall structure framework of the pulmonary parenchyma segmentation convolutional neural network by using a principle that a standard path structure for encoding and decoding simultaneously comprises jump connection, expansion convolution and batch normalization; adopting a weighted loss function layer; dividing the data set; carrying out offline model training out to acquire model weight parameters; inputting a test image and outputting a segmentation result by an output layer through layer-by-layer feedforward of a network. According to an existing lung parenchyma segmentation method, a segmentation missing phenomenon is prone to occurring in a focus area in lung parenchyma, and correct segmentation of the focus area in lung parenchyma segmentation can be effectively improved through enhancement processing on important pixels.

The invention discloses a traffic sign detection and identification method under the complex environment. According to the method, employing a color segmentation method of a normalization RGB space can reduce the image processing time, and algorithm timeliness can be guaranteed; employing a convex hull processing method can solve a problem that a traffic sign is partially shielded; contour analysis is carried out by employing a method extracting contour Fourier descriptors, and a problem that the traffic sign generates rotation, translation and scale change can be solved. Through adding the contour data with projection distortion to a standard database, a problem that the traffic sign generates projection distortion can be solved. Compared with the traditional method, the method has advantages of high robustness and good timeliness, and the method can be applied to intelligent automobiles or pilotless automobiles.

The invention discloses an SAR image target recognition method which mainly solves the problems that in the prior art, feature dimensions of a sample are too large, computation is complex, and consumed time is long. The SAR image target recognition method is implemented in the following steps: (1) an SAR image is initialized to obtain a training sample matrix and a test sample matrix; (2) the Johnson-Lindenstrauss theory and limited equidistance RIP conditions are used for constructing a sparse measurement matrix; (3) according to the sparse measurement matrix, the training sample matrix and the test sample matrix are subjected to dimensionality reduction and normalization to obtain a normalized sample matrix; (4) according to the normalized sample matrix, a nearest neighbor classifier is used for obtaining a category label of the tested sample. Compared with the prior art, the SAR image target recognition method reduces the feature dimensions of the sample and computation complexity, increases SAR target recognition precision and arithmetic speed, and can be used for image processing.

A computer implemented method is provided for processing a 3D fetal ultrasound image. A 3D fetal ultrasound image is obtained (either acquired or received from memory), and the spine is detected within the image. This enables a first reference axis to be defined. A second reference axis is defined perpendicular to the first reference axis, and the 3D fetal ultrasound image is updated (e.g. rotated in 3D space) using the first and second reference axes and an up / down (elevation) orientation detection. This provides a normalization of the orientation of the image, so that a machine learning approach is better able to identify landmarks within new images.

The application relates to the technical field of deep learning, and provides a model construction method and device, an image processing method and device, a hardware platform and a storage medium. The model construction method comprises the steps that a neural network model used for image processing is trained, wherein the neural network model comprises at least one depth separable convolution module, and each depth separable convolution module comprises a layer-by-layer convolution layer, a point-by-point convolution layer, a batch normalization layer and an activation layer which are connected in sequence; and the trained neural network model is quantized to obtain a quantized neural network model. According to the method, firstly, model parameters are quantified, so that the data volume of the parameters is effectively reduced, and the model is suitable for being deployed in NPU equipment. Secondly, the depth separable convolution module in the method is different from the depth separable convolution module in the prior art, and a batch normalization layer and an activation layer are not arranged between a layer-by-layer convolution layer and a point-by-point convolution layer, so that values of model parameters are distributed in a reasonable range, and the model parameters can be quantized with high precision.

The invention relates to the technical field of image processing, and discloses a remote sensing scene classification method and system based on multi-scale feature fusion. The method comprises the steps: obtaining the image features of a collected remote sensing image under different scales, inputting the image features into a preset convolutional neural network model, and carrying out the multi-scale feature fusion, obtaining fusion features, obtaining bottom-layer features of a remote sensing image, carrying out normalization processing on the bottom-layer features and the fusion features to obtain target features, and inputting the target features into a preset classifier to carry out scene classification so as to obtain the category of the remote sensing image, inputting image features of an obtained remote sensing image under different scales into a preset convolutional neural network model to obtain fusion features, then performing normalization processing on the fusion features and bottom-layer features to obtain target features, and realizing remote sensing scene classification based on the target features. The phenomena of high inter-class similarity and large intra-class difference are avoided, and the remote sensing image classification precision and accuracy are improved.

The invention discloses a motor axle vibration anomaly detection and prediction system and method, and relates to the field of image processing, and the system comprises a visual collection system, a server, a motor element segmentation system, a multi-modal data construction system, and a machine vision system. The visual collection system acquires motor images from different directions; the motor element segmentation system executes a segmentation algorithm to generate a wheel shaft vibration sequence, a motor vibration sequence and a shell vibration sequence. The multi-modal data construction system performs data preprocessing and normalization to generate a dual-modal data set; and the machine vision system constructs a three-dimensional residual convolutional neural network by using the bimodal data set, and the machine vision system performs feature extraction and fusion on the bimodal data set to obtain detection and prediction results of external defects and vibration defects of the motor. According to the invention, a time sequence anomaly detection algorithm is integrated, detection can be carried out in the production process of a motor production line, and short-life products which may have problems in the future can be predicted.

The invention relates to a near-infrared vein image processing method for filtering hair noise, which comprises the following steps: 1) collecting a near-infrared vein image and a visible light imageof a part to be imaged, and carrying out size standardization; 2) graying the visible light image; 3) carrying out blocking, and gray normalization and first filtering on each sub-image after blocking; 4) sequentially performing morphological expansion operation and morphological corrosion operation on each sub-image; performing gray value subtraction operation on each subimage and the corresponding subimage; 5) performing binarization and second filtering on each sub-image; 6) obtaining a coordinate set of hair pixel points; and 7) filtering hair noise of the near-infrared vein image. Compared with the prior art, the invention has the advantages that the specific positions of the hair noise pixel points can be given, the hair noise can be processed more effectively and accurately, and meanwhile damage to vein information in the near-infrared image is small.

The invention provides a target detection method and device for CT image data, and electronic equipment, which relate to the technical field of image processing. The method comprises the steps of: obtaining the CT image data, the CT image data comprising target features, performing pixel normalization on the CT image data to obtain a plurality of grayscale images, extracting a grayscale image group from the plurality of grayscale images according to a preset method, and combining the grayscale images into a corresponding pseudo-color image, the grayscale image group being composed of grayscaleimages according to a specified number and a specified sequence, and inputting the pseudo-color image into a pre-trained neural network, so that the pre-trained neural network detects the pseudo-color image to obtain the category and the position of the target feature. Therefore, the accuracy of CT image data target detection can be effectively improved.