436 results about "Computer image" patented technology

The invention relates to an infrared and visible light image fusion method based on a cross-domain Transform, and belongs to the field of computer image processing. The method comprises the following steps: respectively carrying out preprocessing operation on an infrared image and a visible light image to obtain a training data set; an end-to-end image generator network is designed, an encoder module is used for extracting deep semantic features of an infrared image and a visible light image, a fusion module introduces an axial attention mechanism to enhance the global modeling capability of the features, and feature fusion is carried out in combination with information of a spatial domain and a frequency domain; the fused features are gradually recovered to an image space through a decoder module, and a fused image is generated; constructing a fusion loss function module, and guiding the network to focus a significant feature difference between the source image and the fusion image based on a comparative learning idea; and finally, inputting the infrared and visible light image Y channel into the network model, generating a fusion image, completing a training process, and realizing unified optimization of fusion performance and visual quality.

The invention provides a processing method and system for reconstructing an LDI sub-pixel image through sparse compression and application, and the method comprises the steps: calculating a low-resolution to-be-exposed image corresponding to each phase structure in advance through a computer image compression algorithm; the low-resolution to-be-exposed images obtained through calculation are loaded to the digital micro-reflector according to the time sequence, and the display moment of each pattern is matched with the rotation position of the corresponding phase structure; coupling the pattern of the digital micro-mirror and the phase encoding mask in a frequency domain through a 4-f optical system, and reconstructing a high-resolution exposure pattern at a sub-pixel level by utilizing a diffraction effect; projecting a high-resolution light field of the reconstructed high-resolution exposure pattern to the surface of the photoresist, and forming a target circuit pattern by accumulating exposure dose; and the micro-nano structure with sub-pixel precision is obtained after development. The system comprises an image compression module, a pattern matching module and a pattern forming module. According to the invention, the manufacturing cost of the laser direct imaging equipment is reduced under the condition of the same precision.

The invention discloses a transformer substation defect detection model optimization and detection method based on YOLOv11, and relates to the technical field of computer image processing. The method comprises the following steps: on the basis of YOLOv11, replacing a down-sampling convolution module of a backbone network with an ADown module, replacing a C3K2 convolution block of a neck network with a ContextGuided module, introducing an AUX module in front of a detection head to obtain an improved model, and training the improved model through an equipment defect data set to obtain an optimized transformer substation defect detection model. The improved model significantly improves the detection efficiency and precision, and can be more suitable for defect detection tasks and safety standardization in the power industry.

The invention discloses a virtual fitting method and device, a storage medium and electronic equipment, and is applied to the technical field of computer image processing, and the method comprises the steps: processing a target video, and obtaining the human body posture data of each video frame in the target video; constructing a 3D human body model of each video frame by using the human body posture data of each video frame; for the 3D human body model of each video frame, acquiring clothing rendering information of the 3D human body model, and rendering virtual clothing on the 3D human body model based on a dynamic fitting algorithm and the clothing rendering information to obtain a virtual rendering video frame of the video frame; and generating an AR fitting video based on each virtual rendering video frame, and displaying the AR fitting video to the user. Therefore, the selected costume can be tried on in the AR form, the fitting effect of the costume can be displayed for the user, the problem that the size is improper due to the fact that the costume purchased online cannot be tried on is solved, the probability of refunding and changing goods is reduced, and good shopping experience is provided for the user.

The invention relates to the technical field of computer image processing, and particularly discloses a double-branch defogging method based on a Laplacian pyramid, and the method comprises the following steps: S1, building a training data set A containing an original clear image and a corresponding foggy image; s2, performing cutting and data enhancement processing on the input image; s3, a defogging network LPSDNet based on the Laplacian pyramid is constructed, the defogging network LPSDNet comprises a multi-feature fusion focusing branch and a Laplacian pyramid supplementary branch, and after feature extraction of the two branches is completed, features learned by the two branches are fused through a learnable fusion module; s4, training the defogging network LPSDNet by adopting the training data set A until a preset loss function is converged; obtaining a trained defogging network LPSDNet in combination with an objective function; and S5, inputting an image to be defogged into the trained defogging network LPSDNet for testing and verification to obtain a defogging result. The global structure information of the image can be fully captured, and the distortion phenomenon is prevented.

The invention discloses a two-stage unsupervised defogging method based on a pseudo haze image, and belongs to the technical field of computer image processing, and the method comprises the following steps: building a training data set containing a haze image and an unpaired clear image; constructing a two-stage unsupervised defogging network based on the pseudo haze image; training the defogging network by adopting the training data set until a preset loss function is converged; and inputting a to-be-defogged image into the trained defogging network to obtain a defogged image. According to the invention, learning training is carried out on the unpaired real haze-clear image pair, so that the defogging effect of the network on the haze image is improved, and the generalization ability of the network is improved; according to the method, traditional dark channel prior and decomposition and reconstruction of an atmospheric scattering model are introduced into a deep learning framework, so that the network defogging performance is improved, and the network interpretability is improved.

The invention discloses an industrial surface defect image classification method based on hybrid query strategy active learning, and belongs to the technical field of computer image processing and machine learning. The invention aims to solve the technical problems of high cost, long period and low rare defect recognition rate caused by category imbalance due to dependence on large-scale manual labeling. The core of the method is to execute a hybrid query strategy in an iterative loop: firstly, screening out a candidate sample set of model cognitive ambiguity through uncertainty measurement of prediction entropy; secondly, in the candidate set, a core set and hierarchical thought diversity sampling method is adopted to select a final to-be-labeled sample with both characteristic representativeness and category balance. And the query batch is manually annotated and then is used for carrying out iterative updating and optimization on the model. According to the method, the recognition precision and generalization ability of the classification model on various defects can be remarkably improved with extremely low manual labeling cost, and the model development period is greatly shortened.

The invention discloses a ginseng seed lossless classification method and a ginseng seed lossless classification system, relates to the technical field of computer image detection, and solves the problem that in the prior art, a ginseng seed classification method based on image and spectral characteristics of ginseng seeds is lacked. Respectively collecting image data of the ginseng seeds and hyperspectral data of the ginseng seeds; respectively preprocessing the image data of the ginseng seeds and the hyperspectral data of the ginseng seeds; respectively carrying out feature screening on the preprocessed image data of the ginseng seeds and the preprocessed hyperspectral data of the ginseng seeds; fusing the image data of the ginseng seeds after feature screening and the hyperspectral data of the ginseng seeds; after the RBMO algorithm is improved, the RBMO algorithm is combined with the RF model to construct an ORBMO-RF model; and respectively inputting the fused image data of the ginseng seeds and the fused hyperspectral data of the ginseng seeds into an ORBMO-RF model for processing, thereby completing classification of the ginseng seeds.

The invention discloses an interactive image segmentation algorithm based on an algorithm computer, and belongs to the technical field of computer image segmentation. The interactive image segmentation algorithm comprises the following steps: S1, image uploading and preprocessing, S2, construction of a bidirectional interactive propagation mechanism, S3, adaptive image segmentation, and S4, image detection after segmentation. The interactive image segmentation algorithm is combined with lightweight semantic propagation, adaptive segmentation and a user feedback closed loop, the user interaction frequency is remarkably reduced while the precision is ensured, and compared with a pure deep learning scheme, the interactive image segmentation algorithm is more suitable for small sample scenes and is suitable for scenes with double requirements on precision and efficiency, such as medical images and industrial quality inspection; the algorithm can quickly adapt to a new field, a new category or an unknown object, reduces dependence on a large amount of training data in a specific field, minimizes user interaction times and complexity on the premise of ensuring high segmentation precision, can improve user experience, and has the characteristics of extremely simple interaction, ultrahigh precision, superstrong robustness and real-time response.

The invention discloses a sparse matrix processing method and device in an image recognition process, equipment and a medium, and relates to the field of computer image processing, and the method comprises the steps: inputting a feature image into a convolutional neural network; if the convolution kernel meets a preset sparse matrix condition, performing structured sparseness on the convolution kernel to obtain a sparse matrix, compressing the sparse matrix by using an improved compression sparse row format to obtain a compression matrix, and constructing an index matrix of the compression matrix; storing the feature image, the compression matrix, the index matrix and the preset bias matrix in a global memory, activating a target thread bundle in the computing core, and distributing the target thread bundle to a target data processing channel; loading the compression matrix and the index matrix to a shared memory corresponding to the target data processing channel by using the target thread bundle, and loading the feature image and a preset bias matrix to a register file; and executing the operation instruction to read the current matrix operation data from the shared memory and the register file to the matrix calculation unit to complete the matrix multiplication and addition operation.

The invention discloses an unmanned aerial vehicle small target detection method based on feature enhancement and selective attention and a related device, and relates to the technical field of computer image target detection, and the method comprises the steps: obtaining an aerial image of an unmanned aerial vehicle, adjusting the image to a preset resolution, and obtaining an adjusted image; inputting the image into a backbone network of a network architecture, extracting a multi-level feature representation through a plurality of CALBlock feature extraction modules, and enhancing features of the multi-level feature representation through an EMIT edge enhancement architecture to obtain an enhanced multi-level feature representation; inputting the enhanced multi-level feature representation into a neck network of the network architecture, and performing cross-scale feature fusion and enhancement through an FSAFPN structure to obtain a multi-scale enhanced feature map; and inputting the multi-scale enhanced feature maps into a detection network of a network architecture, and performing target detection by using a decoder to obtain target category probability distribution and bounding box coordinates.

The present invention relates to the technical field of computer image processing, and in particular to an agent-based image analysis and processing method and system, and a storage medium. Image quality analysis and estimation is performed by means of a pre-trained image quality analysis model, and a corresponding processing strategy and a corresponding image processing sequence are generated, thereby achieving automatic image quality analysis and evaluation and facilitating selection of different processing strategies based on different image content, so as to satisfy processing requirements of various application scenarios. A corresponding image processing model is retrieved on the basis of the result of matching between the processing strategy and basic information of each image processing model in a knowledge base, and then called according to the image processing sequence, and an automatic image processing flow control model automatically controls the image processing model to execute image processing according to the image processing sequence and performs monitoring, thereby achieving automatic image processing, reducing the requirement of manual intervention, and saving the time and cost.

The invention discloses a submerged chain conveyor liquid level measuring method and device, a storage medium and electronic equipment, and relates to the technical field of data processing.The method comprises the steps that image data, collected by an industrial camera, of a liquid level ruler in a submerged chain conveyor slag pool is obtained; preprocessing the image data based on a bilateral filter to obtain recognizable image data; the recognizable image data are input into a preset deep learning model for image recognition, point location information of the liquid level ruler is obtained, and the preset deep learning model is used for recognizing the point location information, contained in the image, of the liquid level ruler; and the liquid level reading of the liquid level ruler is determined according to the point location information. Compared with the prior art, the real-time image can be processed by utilizing a computer image recognition technology, the working safety is greatly improved while the measurement accuracy is improved, the environmental pollution and corrosion conditions of measurement components are greatly reduced, the labor and equipment cost can be reduced, and the measurement accuracy is improved. And the accuracy of the measured liquid level can be improved.

The invention provides an efficient image super-resolution method and device based on dual guidance semantic enhancement, and relates to the technical field of image super-resolution, visual content generation and the like in computer image process.The method comprises the steps that an image super-resolution model containing a pre-training diffusion model backbone structure is constructed, and a pre-training diffusion model is constructed; a trainable LoRA module is inserted into a U-Net structure and a VAE encoder of the image super-resolution model; calculating the perception image loss of the generated image and the real image by adopting a dual-guidance quality enhancement training strategy; based on a semantic alignment method of score matching, performing semantic feature constraint on the generated image through a condition vector of a pre-training diffusion model to form semantic consistency loss aligned with real image distribution; and carrying out end-to-end training on the image super-resolution model under joint optimization of perceptual image loss and semantic consistency loss, and after the training is finished, realizing high-resolution generation of a low-resolution image only through one-time U-Net reasoning.

The invention discloses a computer image processing system based on an intelligent identification technology, and the system comprises a real-time monitoring module which carries out the monitoring and learning of obtained image data information through a real-time image quality evaluation technology; an alarm system is arranged in the image recognition module, and the alarm system is triggered according to the change of the acquired image data information; the image optimization module receives the alarm signal, locks the damaged position of the image, intercepts image information before and after the image is damaged, and deeply learns the image information by adopting a degraded image restoration algorithm based on a neural network; according to the method, a degraded image restoration algorithm based on a neural network is utilized, image information is deeply learned, clear original information is extracted, fuzzy and de-noising processing is carried out, the damaged position of the image is effectively filled, the monitoring capability is enhanced, direct transformation from the damaged image to the restored image is realized by establishing an end-to-end neural network model, and the image quality is improved. And the image processing efficiency and accuracy are improved.

The invention relates to the field of computer image processing, and provides a remote sensing multispectral image fusion method and system based on multi-stage feature correction, and the method comprises the steps: obtaining a spectrum mask through a low-resolution multispectral image and a panchromatic image, and covering the panchromatic image through the spectrum mask to obtain a corrected panchromatic image; performing scale feature enhancement on the low-resolution multispectral image, and performing down-sampling and convolution on the corrected panchromatic image to obtain image scale features; obtaining attention block input, and mapping the attention block input to obtain a mapping vector group; performing image correction by using the mapping vector group to obtain a feature correction vector, performing iteration on the feature correction vector to obtain a target feature correction vector, and obtaining a fusion feature matrix; and iterating the fusion feature matrix to obtain target fusion features, and carrying out feature refinement and convolution on the target fusion features layer by layer to obtain a high-resolution multispectral image. According to the invention, a high-fidelity and high-resolution multispectral image can be obtained.

The invention discloses a method for determining the shale content of tight sandstone, and relates to the technical field of exploration and research of tight rock oil and gas reservoirs. The method comprises the following steps: identifying a rock slice identification photo by using a computer image identification method, obtaining the area percentage of debris particles with the diameter greater than a set threshold value in the slice identification photo, and then calculating the area percentage of argillaceous particles in the slice identification photo; according to the area percentage of the argillaceous particles and the corresponding times of the polarizing microscope, displaying the argillaceous particles in a rectangular coordinate system in a scatter mode; fitting all scatter points by using a nonlinear least square method, and establishing a shale content volume percentage change relation curve under different objective lens multiples of the polarizing microscope; and solving the curvilinear equation, and calculating the shale content of the sample by using the curvilinear equation. According to the method, the error caused by the fact that the shale content is calculated by an eye estimation method singly through a slice identifier is avoided, and the analysis result is more consistent with the real content of the sample, so that the accuracy of the shale content determination by an image method is greatly improved.

The invention discloses a method for detecting the concentration of bacteria in water based on an image processing technology, and relates to the technical field of computer image processing. The method for detecting the concentration of bacteria in water based on the image processing technology comprises the following steps: acquiring preliminary detection accuracy; acquiring an area scanning evaluation value; generating a superpixel; and acquiring the concentration of the to-be-detected bacteria. According to the method, the to-be-detected sample data is preliminarily marked through the fluorescent marker to obtain the preliminary detection accuracy, meanwhile, the to-be-detected region division process in the laser layer-by-layer scanning operation process is monitored in real time to obtain the region scanning evaluation value, and then the obtained three-dimensional image is subjected to superpixel segmentation to generate superpixels; finally, the number of the to-be-detected bacteria in the superpixels is counted, and the concentration of the to-be-detected bacteria is obtained in combination with the preset experimental group water sample volume, so that the effect of improving the reliability of detecting the concentration of the bacteria in water based on data processing is achieved, and the problem that the reliability of detecting the concentration of the bacteria in water based on data processing is low in the prior art is solved.

The invention relates to the technical field of computer image processing, in particular to a farmland forest network extraction method and system based on a SegFormer model, and the method comprises the steps: obtaining a remote sensing image of a farmland forest network; the remote sensing image is preprocessed; performing image annotation on the preprocessed remote sensing image to obtain a training data set; a farmland forest network extraction model based on a Segform model is constructed; adopting the training data set to train the farmland forest network extraction model; and inputting a remote sensing image of a to-be-detected area into the trained farmland forest network extraction model to obtain a high-precision farmland forest network target area. According to the embodiment of the invention, the SegFormer model is optimized, the jump connection module is added, and the CBAM module is introduced, so that the perception capability of the model for space and channel information is enhanced, the segmentation precision is further improved through the jump connection and the up-sampling module, and the farmland forest network can be accurately extracted.

The invention belongs to the field of tunnel wall surface deformation monitoring, and particularly relates to a tunnel wall surface deformation monitoring method based on computer image recognition, which comprises the following steps: acquiring expandable data and accessing the expandable data to a subsystem; image stabilization operation is carried out on the edge side, and a standardized preprocessing subsystem is constructed to process data; constructing a multi-modal fusion architecture of a structural topological graph and a space-time diagram neural network; constructing a hierarchical alarm judgment mechanism of an uncertainty-driven risk score and a sectionalized adaptive threshold, and linking the BIM and work order circulation to form a closed loop; according to the invention, millimeter-level displacement resolution can be realized, false alarm and missing alarm can be obviously reduced, and long-term online traceable tunnel wall surface deformation monitoring and graded early warning closed loop can be supported.

The invention relates to a real scene image defogging method based on disturbance defense and semantic guidance, and belongs to the field of computer image processing. The method comprises the steps that a disturbance defense module is constructed, defended features are obtained and input into a defogging network, feature extraction is conducted on the defended features in the forward propagation process, the defended features are input into a defogging decoder, and low-dimensional features are further recovered into defogged images; inputting the composite domain foggy image into a source defogging network to process and output an intermediate feature; the features extracted by the feature extraction module are input into a semantic decoder; the semantic decoder uses the learned semantic features to help the image restoration of the defogging network; the defogging features of the defogging decoder and the semantic features of the semantic decoder are fused; training a content disturbance defense module, a semantic decoder, a source defogging network, a multi-scale fusion module and a defogging network; and the trained defogging network is used for defogging the foggy image. According to the invention, a clearer and more real defogged image can be generated.

The invention discloses an agricultural greenhouse water and fertilizer management system and method based on video recognition, and relates to the technical field of computer image recognition, and the system comprises a camera which is used for collecting crop images; the environment sensor is used for collecting water and fertilizer environment data of the agricultural greenhouse; the local host is provided with a crop identification model, and the crop identification model is used for identifying the real-time growth state of crops in the crop image; the cloud server is deployed with a growth decision model and a growth regulation model, the growth decision model formulates a growth control strategy according to the types and growth stages of crops in the agricultural greenhouse and the water and fertilizer environment data, the growth regulation model calculates an income value, and a regulation scheme for the growth control strategy is solved by taking the maximum income value as a target; and forming an adjusted growth control strategy. According to the method, the growth state of the crops is considered on the whole instead of achieving the optimal state of some crops, and the benefit of the agricultural greenhouse is improved.

The invention discloses a computer vision image processing method based on artificial intelligence, which relates to the technical field of computer image processing and comprises the following steps: synchronously acquiring visible light, infrared thermal images, vibration spectrum and reference temperature data to ensure time-space consistency; a dominant frequency and a direction angle are extracted based on a vibration spectrum, a visible light image is directly driven to be deconvolved and deblurred, and edge details are recovered; performing dynamic weight fusion on the edge gradient map of the deblurred image and the normalized temperature difference map to generate a dual-channel high signal-to-noise ratio feature map; performing cross validation on the defect candidate region through fractal dimension and a heat conduction rule, and generating a high-confidence mask; and the direction angle is dynamically fed back and adjusted based on the mask edge fracture proportion, iterative optimization is performed until the defect edge continuity reaches the standard, and the method adapts to the complex vibration working condition.

The invention relates to the technical field of computer image processing and medical image analysis, in particular to a cervical cancer MRI (Magnetic Resonance Imaging) image segmentation method for improving a U-Net structure based on a feed-forward channel double-attention mechanism. In order to solve the problems that a traditional U-Net structure is weak in multi-scale information extraction capability, inaccurate in boundary fuzzy region recognition and the like in processing female abdominal cervical cancer MRI images, the method proposes that a feedforward connection mechanism and a double-attention mechanism are embedded into an encoder and a bottleneck module to form a novel U-Net segmentation model; the recognition and segmentation precision of the cervical cancer focus area is improved, and a high-quality image basis is provided for subsequent clinical diagnosis and treatment. A double-attention mechanism is integrated into a plurality of key nodes of the model, a joint channel-space attention module is added to the tail of each convolution module in an encoder, a cavity space pyramid pooling structure is introduced to a bottleneck position, a channel attention mechanism is embedded, and joint modeling of cross-scale, multi-channel and space context information is achieved.

The invention discloses a large-scale rotary equipment blade three-dimensional profile measurement and point cloud splicing method and system based on line laser, and belongs to the technical field of precision measurement and computer image processing. Reading a CAD model of a to-be-measured blade, extracting geometric features of the blade by a system, and analyzing the geometric features to generate a plurality of measurement path candidate sets; a measured blade is fixed on a rotary table, a multi-dimensional measurement movement mechanism is adopted to carry a line laser sensor to scan the surface of the blade step by step according to a planned final path, and accurate three-dimensional point cloud data is obtained; the method comprises the following steps: after acquiring point cloud data of a blade, initially aligning the scanned point cloud data in combination with position information of equipment; the optimized ICP algorithm is used, a transformation matrix between the point clouds is continuously adjusted through an iteration mode, the distance between the point clouds is minimized, mismatching and noise influences are further reduced, and therefore accurate registration is achieved. The invention aims to solve the problems of low precision and poor efficiency of aero-engine blade three-dimensional profile measurement.

The invention relates to a controllable Chinese character new font generation method and system based on a diffusion model, and belongs to the technical field of computer image processing. Based on the system, a user inputs a character image of a target stroke style and a character image of a target skeleton, then multi-scale skeleton features are extracted through a skeleton information extraction network, stroke style features are extracted through a stroke style extraction network, and the multi-scale skeleton features are extracted. And finally, denoising operation prediction of each step of the U-net network in the diffusion process is guided through the features, so that a new font character image conforming to expectation is generated. The method is inspired by a structure reference method, features are extracted from the input image through the neural network, a diffusion model is guided through skeleton features and stroke features to generate a new font style, and basic requirements of users are met.

The invention discloses a cast-in-place pile pouring process quality detection method capable of sensing ERT on the inner side and the outer side. The cast-in-place pile pouring process quality detection method comprises the steps that conductivity information of the inner side and the outer side of a cast-in-place pile is collected through an electrical resistance tomography electrode array arranged at the bottom of a cast-in-place pile feeding guide pipe; the two-dimensional position information of the electrical resistance tomography electrode array in the pile hole in the horizontal direction is positioned through an ultrasonic transducer array arranged at the bottom of the feeding guide pipe; the depth information of the electrical resistance tomography electrode array is acquired through an incremental encoder and a roller which move along with the feeding guide pipe in the vertical direction; transmitting the obtained conductivity information, the two-dimensional position information and the depth information to a computer imaging system to obtain a three-dimensional conductivity distribution image of the cast-in-place pile, and analyzing the obtained image to realize quality detection of the cast-in-place pile. Total cross-section imaging of concrete and a reinforcement cage can be achieved, and the positions and sizes of karst caves and collapse possibly existing in the pile hole wall can be accurately measured.

A camera tracking system receives patient reference tracking information indicating pose of a patient reference array tracked by a patient tracking camera relative to a patient reference frame. A local XR headset view pose transform is determined between a local XR headset reference frame and the patient reference frame. Remote reference tracking information is received indicating pose of a remote reference array tracked by a remote reference tracking camera. A remote XR headset view pose transform is determined between a remote XR headset reference frame of a remote XR headset and the remote reference array. A 3D computer image is transformed from a local pose determined using the local XR headset view pose transform to a remote pose determined using the remote XR headset view pose transform. The transformed 3D computer image is provided to the remote XR headset for display with the remote pose relative to the remote XR headset reference frame.

The invention relates to the technical field of tunnel detection, in particular to a tunnel wall surface deformation detection method based on computer image recognition, which comprises the following steps: S1, setting an image acquisition system; s2, image preprocessing; s3, extracting wall surface features; s4, deformation detection; and S5, deformation early warning and reporting. In an image acquisition link, image acquisition is not singly carried out, and three-dimensional point cloud data of a tunnel wall surface is acquired by utilizing a laser radar technology to assist image acquisition and is fused with image data. By making the features in the image correspond to the space coordinates in the three-dimensional point cloud data, the deformation condition of the tunnel wall surface can be analyzed more accurately, especially when the wall surface with a complex geometrical shape is processed, a more comprehensive space information preprocessing link can be provided, and the processing efficiency of the tunnel wall surface is improved. And in combination with an image registration technology in image preprocessing, the analysis efficiency among multiple images can be improved, and repeated processing caused by image differences is avoided, so that the deformation detection precision is improved.

The invention belongs to the technical field of computer images, and provides an object counting method and device for enhancing text guidance by utilizing frequency characteristics. The method comprises the following steps: inputting a query image to a visual encoder to obtain visual features; inputting the object description text to a text encoder to obtain text embedding; fusing the visual features and the text features by using a fusion network to obtain fusion features; the fusion features are input into a decoder to obtain a density estimation graph of the object, the decoder comprises K cascaded adaptive frequency selection modules and one output layer, and the kth adaptive frequency selection module comprises a first convolution unit, an adaptive frequency selector and a first up-sampling unit; the kth self-adaptive frequency selector carries out filtering processing on the amplitude spectrum and the phase spectrum of the output characteristic of the first convolution unit in the frequency domain, and converts a filtering processing result into a spatial domain; according to the method, related frequency components are dynamically emphasized in decoding, the space domain and frequency domain features are combined, and accurate object positioning and accurate counting are achieved.