30 results about "Wavelet fusion" patented technology

The invention relates to the technical field of medical image processing, and provides a medical image segmentation method based on wavelet enhancement and multi-scale feature fusion. According to the method, a CNN-Transform double-branch coding structure is combined, a multi-scale wavelet fusion module is provided, from the perspective of a frequency domain, Haar wavelet transform is adopted to extract an image high-frequency sub-band so as to enhance edge and texture detail expression, dynamic weighting is performed on different frequency band features through grouping convolution and a sub-band attention mechanism, and the discrimination capability is improved; meanwhile, a multi-scale cavity pyramid structure is fused in a spatial domain, and after cross attention dynamic fusion is introduced, a feature alignment mechanism of a wavelet domain and the spatial domain is established; and collaborative fusion of frequency domain and space domain features is realized. The method effectively improves the segmentation precision of the fuzzy boundary and the fine-grained structure under the complex background, has good universality and adaptability, and is suitable for various medical image segmentation tasks.

The invention discloses a thyroid intraoperative real-time navigation method and a thyroid intraoperative real-time navigation system based on multi-mode optical fusion. The thyroid intraoperative real-time navigation method comprises the following steps: outputting visible light through an endoscope and coupling light waves of a narrow-band multispectral light source to irradiate an operative field, exciting parathyroid glands to generate near-infrared fluorescence, and receiving reflected visible light, split light, near-infrared fluorescence and laser speckle signals through an endoscope probe. And separating the composite optical signal into four channels, and respectively generating an anatomical structure color image, a blood vessel spectroscopic image, a parathyroid gland near-infrared fluorescence image and a laser speckle image. Performing decorrelation processing on the laser speckle image to generate a blood flow dynamic pseudo-color decorrelation speckle image; an anatomical structure, a blood vessel center line, a parathyroid gland contour and blood flow dynamic feature points are extracted through a multi-modal registration technology, after affine transformation space alignment is conducted, a comprehensive imaging map containing the anatomical structure, blood vessel distribution and parathyroid gland function marking information is generated through a wavelet fusion algorithm, and intraoperative multi-dimensional real-time tissue navigation is achieved.

The invention provides a self-adaptive bearing fault diagnosis method based on multi-base wavelet fusion. The objective of the invention is to solve the problems of noise reduction, insufficient feature extraction and low diagnosis precision under noise conditions. A Kaisixi University bearing public data set is used as original data, and Gaussian noise with different SNRs is superposed to simulate various noise intensities. And uniformly carrying out length alignment, down-sampling, equal-length segmentation, division and normalization preprocessing. Then, wavelet bases such as sym4, db4, coif5 and the like are adopted for parallel multi-scale decomposition and reconstruction; and adaptively determining the number of decomposition layers and a threshold strategy according to the noise level, and generating a de-noising branch. And performing weighted fusion on the denoising results of the branches, and performing iterative denoising on the residual error. Signals subjected to noise reduction processing are sent to a double-branch convolution-cycle-attention network, a convolution layer extracts features, an LSTM and a self-attention module capture time sequence changes, and accurate recognition of various bearing faults is achieved. The training adopts a segmented attenuation learning rate and an early stop strategy, and the robustness and generalization ability of different SNR working conditions are improved.

The invention discloses an intelligent nursing body position adjusting method and system based on respiratory function protection, and relates to the technical field of artificial intelligence and medical nursing. The method comprises the steps that body pressure distribution signals and respiratory thoracic impedance signals of a patient are collected in real time, a physiological data sequence is generated through Kalman filtering and wavelet fusion noise reduction, the physiological data sequence is input into a pre-trained deep learning network, and fusion feature vectors and body position performance indexes are output. And if the body position performance index is not lower than a preset threshold value, inputting the fusion feature vector into a pre-trained classification model, and outputting a classification label of the breathing state. A preset parameter optimization mapping table is inquired according to the classification labels, candidate parameter combinations sorted according to priorities are obtained, the parameter group with the highest priority is selected, a multi-motor cooperation instruction is calculated through a feedback control algorithm, and dynamic regulation and control of the bed surface are achieved. The invention aims to realize respiratory function protection and individualized nursing by sensing the respiratory state in real time, intelligently deciding and adjusting parameters, cooperatively executing control and adopting a safety protection mechanism.

A deep learning fan fault diagnosis method based on double wavelet fusion and CEEMDAN decomposition comprises the following steps: S1, collecting vibration signals of a fan motor driving end, a fan driving end and a non-driving end, and constructing a data set; s2, performing improvement on the basis of wavelet packet noise reduction to form a double-wavelet fusion noise reduction algorithm; s3, carrying out noise reduction processing on the acquired signals; s4, carrying out sample division on the noise reduction signals according to a fixed time window, and carrying out classification according to measurement points; s5, randomly dividing the samples into a training set, a verification set and a test set; s6, decomposing all the samples by using CEEMDAN, and extracting IMFs; s7, sending the sample into the time-frequency domain joint feature extraction model of the corresponding measuring point for modeling; s8, training to obtain a diagnosis model and exporting a weight file; and S9, loading the model detection test set, and comprehensively judging the operation state of the fan. According to the invention, the signal processing quality and the fault identification accuracy are improved.

The invention discloses an image super-resolution reconstruction method and device, equipment, a medium and a program product, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining a to-be-reconstructed image; inputting an image to be reconstructed into the image super-resolution model to obtain a reconstructed image output by the image super-resolution model; the image super-resolution model is based on a residual network model and comprises a plurality of context attention modules and a plurality of wavelet fusion modules; the plurality of context attention modules are used for extracting global features corresponding to the to-be-reconstructed image; each wavelet fusion module is used for fusing global contour information and local detail texture information corresponding to a plurality of feature maps associated with the to-be-reconstructed image; the reconstructed image is obtained by fusing the image super-resolution model based on the to-be-reconstructed image, the global features, the global contour information and the local detail texture information. According to the method, the problems that the to-be-reconstructed image is blurred, unclear and low in resolution are solved.

The invention relates to the technical field of image processing, and discloses an underwater image enhancement method based on sparse prior guidance and frequency domain information fusion, which comprises the following steps: constructing and training an underwater image enhancement network, inputting an original image into a red channel histogram equalization prior guidance network, and obtaining a prior weight map; a sparse priori fusion module is adopted to analyze the priori weight map and capture the context relation between global pixels of the image, priori flow branches are utilized to reinforce image color information, feature flow branches are utilized to expand channel dimensions to carry out image local refinement, and finally a feature map is extracted; and inputting the feature map into a wavelet edge fusion module, and cascading an edge enhancement network to obtain an enhanced underwater image. The red channel histogram equalization prior guidance module provided by the invention can effectively solve the problem of color distortion of an underwater image caused by insufficient brightness of a red channel, and the wavelet fusion module is utilized to refine image edge information and improve image detail information.

The invention relates to the technical field of citrus detection in agricultural automation, in particular to a citrus detection method based on a frequency domain aggregation attention mechanism and a multi-scale encoder. According to the method, an innovative detection framework is provided for the problem of small target and occlusion target detection, and the detection framework comprises a frequency aggregation attention network (FAN) and a multi-scale Transform encoder. The frequency aggregation attention network decomposes the feature map through two-dimensional discrete wavelet transform, and enhances the frequency domain features of the small target and the shielding target; the multi-scale Transform encoder is combined with convolution feature pyramid operation, and high-frequency detail information is reserved through a wavelet fusion module. In addition, an IoU is adopted to perceive query selection and optimize a loss function, so that the detection precision is remarkably improved. Through frequency domain feature enhancement and multi-scale feature fusion, high-precision detection of citrus fruits, especially small targets and sheltered targets, is realized, and the method is suitable for fruit grading, yield statistics and other scenes in the field of intelligent agriculture.

The invention discloses an insulation discharge defect detection method and system based on solar-blind ultraviolet and UVA dual wavebands, and belongs to the technical field of electrical equipment insulation state detection.The method comprises the steps that a solar-blind ultraviolet waveband optical signal, a UVA ultraviolet waveband optical signal, a discharge current pulse signal, a solar-blind ultraviolet waveband image and a UVA waveband image of detected equipment are synchronously collected; carrying out lock-in amplification noise suppression on the UVA ultraviolet light signal, and extracting a photon event rate feature; extracting a light intensity peak value characteristic from the solar blind ultraviolet light signal, and calibrating a discharge quantity value through a nonlinear regression model in combination with a photon event rate; carrying out dynamic range expansion on the UVA ultraviolet image, and carrying out wavelet fusion on the UVA ultraviolet image and a solar blind ultraviolet image; fractal dimension features of the discharge channel are extracted from the fused image; and inputting the photon event rate, the light intensity peak value, the discharge capacity value and the fractal dimension features into a deep learning model, and outputting defect types and severity.

The invention discloses a small target detection method based on wavelet fusion and variance guidance, and relates to the technical field of target detection, and the method comprises the steps: extracting input features through a backbone network; a wavelet fusion down-sampling module is used for processing multi-scale feature information, and high-frequency texture details and structure information are reserved; performing multi-scale feature fusion and channel weighting operation through a variance guiding spatial pyramid pooling module; carrying out feature fusion on the neck part of the detection network structure by adopting a PAN-FPN architecture, and transmitting features after feature fusion to a detection head; and the detection head executes positioning regression and category prediction tasks and outputs a detection result. According to the method, wavelet fusion down-sampling and variance guided spatial pyramid pooling are combined, the feature expression ability is collaboratively optimized from the two dimensions of frequency domain and statistical features, then more accurate and stable detection of a small target is achieved, the detection performance of a model in a complex background, partial shielding and target dense scenes is effectively improved, and the method is suitable for large-scale popularization and application. Good application prospects are realized.

The invention discloses a tree species identification method based on multi-modal spectrum and texture features of a wood cross section. The method comprises the following steps: acquiring hyperspectral image data of the wood cross section; constructing a comprehensive similarity matrix based on a plurality of spectrum similarity indexes, and selecting a representative wave band subset from the hyperspectral image data by adopting a multi-strategy wave band screening mechanism; performing multi-scale wavelet fusion on the representative wave band subset to generate a single-channel fusion image with consistent spatial resolution, and extracting points of interest and spectral features thereof from the single-channel fusion image; generating a gray-scale base map based on the hyperspectral image data, extracting various complementary texture feature maps from the gray-scale base map, and screening out significant points of interest from the various texture feature maps; and fusing the spectral features and the texture features, constructing feature vectors for representing wood tree species, and inputting the feature vectors to a classification model to complete tree species identification. According to the invention, rapid, accurate and intelligent identification of wood tree species can be realized.

The application relates to a SAR image change detection method based on fusion difference maps and morphological reconstruction, which comprises the following steps: S1, acquiring two SAR images of the same region at different times; S2, obtaining a logarithmic mean ratio difference map and a logarithmic ratio difference map according to the two SAR images; S3, performing SLIC superpixel segmentation on the logarithmic ratio difference map to obtain a superpixel segmentation difference map; S4, performing wavelet fusion processing on the logarithmic mean ratio difference map and the superpixel segmentation difference map to obtain a wavelet fusion difference map; S5, performing morphological reconstruction on the wavelet fusion difference map to obtain a final difference map; and S6, processing the final difference map by using a fuzzy C-means clustering algorithm to output a detection result map. The method can better utilize the information of neighboring pixels, has excellent detection effects on different types of SAR images, has very good robustness to noise, and can maintain the details of the image.

The application discloses a low-illumination image enhancement method based on an improved Retinex algorithm, and applies to the technical field of image enhancement, and comprises the following steps: performing discrete wavelet decomposition on a low-illumination image to obtain low-frequency and high-frequency components of the image; converting the low-frequency component into HSV, separately performing brightness correction on a V channel, then converting back to RGB, performing bilateral filtering, and then converting to HSV to extract the V channel; performing image enhancement on the low-frequency component by using an improved Retinex algorithm based on a joint weighting of a bilateral filter and a Gaussian filter as a new center-surround function, and performing median filtering processing, then converting to HSV to extract the V channel; weighting and fusing the two V channels, retaining the H and S channels after algorithm enhancement, then converting back to RGB, performing discrete wavelet fusion with the high-frequency component after denoising, and stretching and outputting the enhanced low-illumination image. The application can effectively guarantee the color, edge details and avoid local distortion of the image.

This invention discloses a method, apparatus, and related components for generating defective images based on generative adversarial networks (GANs). The method includes: acquiring a fabric image and labeling the image category and defect location to obtain a realistic defective image; inputting the realistic defective image and a preset noise vector into the generator of a GAN to generate a simulated defective image; calculating the discriminator loss and generator loss of the GAN based on the realistic defective image and the simulated defective image, respectively, according to a preset loss function, to obtain the discriminator loss and generator loss; iterating the model parameters of the GAN based on the discriminator loss and generator loss to obtain a defective image generation network; and performing wavelet fusion processing on the simulated defective image output by the defective image generation network and the corresponding realistic defective image to obtain an ideal defective image. This invention improves the quality of the simulated defective image obtained through a GAN.

The invention relates to a runoff prediction method based on convolution long and short time memory network and wavelet fusion. The runoff prediction method comprises the following steps: (1) data preprocessing; (2) extracting spatio-temporal characteristics; (3) extracting time sequence features; (4) feature compression and alignment; (5) carrying out cross-modal feature fusion; and (6) runoff prediction. According to the method, collaborative modeling of a meteorological element space interaction relationship and runoff multi-scale time sequence characteristics can be realized, the defect of spatial-temporal characteristic splitting processing of a traditional model is overcome, the responsiveness of the model to runoff changes under complex meteorological conditions is enhanced through a dynamic characteristic fusion mechanism, and the generalization and accuracy of a prediction model are improved.

The invention discloses a pump cavitation state intelligent identification method based on multi-scale fusion DCNN, and the method comprises the steps: (1) collecting noise signals when a centrifugal pump is in different cavitation states, and segmenting the noise signals into a plurality of noise signal samples in a non-overlapping manner; (2) acquiring a wavelet time-frequency diagram of each noise signal sample, constructing a time-frequency diagram data set, and dividing the time-frequency diagram data set into a training set, a verification set and a test set; (3) constructing a time-frequency enhanced wavelet fusion multi-scale interpretable DCNN model, wherein the model comprises a time-frequency attention module, a fusion multi-scale module and a fusion class activation mapping module; (4) training the constructed model based on the training set, and storing optimal parameters of the model based on the verification set; and (5) testing the model based on the test set, and outputting the predicted cavitation state and the class activation diagram corresponding to each sample. According to the method, high classification accuracy can be obtained in the cavitation state recognition task of the centrifugal pump, and high interpretability is achieved.

This invention belongs to the field of video coding and image processing technology, specifically relating to a multi-level dual-stream pulse wavelet fusion panoramic video quality enhancement method, comprising the following steps: 1. Acquiring the original high-quality panoramic video image, performing video coding standard compression processing to generate a compressed low-quality panoramic video image; 2. Performing shallow feature extraction and mapping it to a high-dimensional feature space to obtain shallow features; 3. Performing feature splitting, inputting it into the parallel global and local stream branches of the pulse-driven biomimetic quality enhancement model SWFN for feature extraction; 4. Performing channel stitching, using fusion convolution for feature interaction and compensation, and outputting the fused features; 5. Performing residual connection and reconstruction, outputting a quality-enhanced panoramic video frame. This method can more accurately eliminate the block artifacts and quantization distortion caused by VVC coding compression, improving the reconstruction quality of ultra-high-definition panoramic video.

The invention discloses an online LiDAR moving target segmentation method based on Haar wavelet enhancement, and belongs to the technical field of target segmentation. The problem that an existing moving target online segmentation method is low in precision is solved. According to the method, a residual multi-layer perceptron structure is adopted, the residual multi-layer perceptron structure can reduce information loss caused by depth reduction by introducing a residual mechanism, the mapping quality from the point cloud to the BEV feature map is improved, and spatial information expression is enhanced. By designing the wavelet fusion block, multi-scale feature modeling can be enhanced based on Haar wavelet transform, information expression of a BEV feature map in the horizontal direction, the vertical direction and the diagonal direction is enhanced, and the capturing capacity of a moving target is improved. The optimized coding-decoding architecture can reduce the calculation overhead while improving the segmentation precision, and meets the real-time requirements of online tasks. Through point cloud feature extraction, feature fusion modeling and segmentation network design, segmentation precision and real-time performance are improved. The method can be applied to moving target segmentation.

The present invention discloses a target recognition method, system, device and medium for complex scenes, and relates to the field of computer vision technology. The method comprises: inputting a test image into a target recognition model for processing to obtain a target recognition result; the target recognition model is constructed based on a dual-path coding network and a weighted optimization loss function; the dual-path coding network includes a MobileNet path and a wavelet path; wherein the MobileNet path is used to extract multi-scale spatial features and is composed of a hierarchical encoder and a multi-feature prediction decoder connected in sequence; the wavelet path is used to extract frequency domain features and is composed of a global semantic wavelet coding module and a local wavelet fusion decoding module connected in sequence; the multi-feature prediction decoder is also connected to the global semantic wavelet coding module and the local wavelet fusion decoding module, respectively. The present invention can improve the precision and accuracy of target recognition for complex scenes.

This invention relates to an RGB-D salient target detection method based on visual liquid time constant and cross-modal wavelet fusion, belonging to the field of image processing technology. The method includes: using cascaded visual Mamba blocks to perform feature encoding on the acquired preprocessed RGB image and corresponding depth map, extracting multi-scale RGB features and multi-scale depth features; inputting the multi-scale RGB features and multi-scale depth features into a constructed visual liquid time constant module to obtain enhanced multi-scale RGB features and enhanced multi-scale depth features; inputting the enhanced multi-scale RGB features and enhanced multi-scale depth features into a constructed cross-modal wavelet fusion module to obtain cross-modal fused features; inputting the cross-modal fused features into a cascaded multi-scale decoder, and outputting an RGB-D salient target detection prediction map by constructing a total loss function. This aims to solve the technical problems of structural destruction and high-frequency detail degradation caused by existing Mamba architectures during global modeling.

The invention relates to the technical field of non-reflection signal processing, in particular to a processing method for shale oil non-reflection weak signal enhancement based on Fresnel zone constraint, which comprises the following steps: collecting well seismic geological data of a shale oil field, processing to obtain a CRP gather data vector, analyzing the frequency domain characteristics of each modal component of the CRP gather data vector, and obtaining a CRP gather signal; determining an amplitude attenuation stability coefficient and a non-reflection wave characteristic value of each modal component, analyzing the difference between each modal component and other modal components about the non-reflection wave characteristic value to obtain reflection signal identification of each modal component so as to extract a reflection signal, and performing wavelet fusion processing to obtain reflection wave field CRP gather data; obtaining subtraction CRP gather data by adopting a self-adaptive subtraction filtering algorithm; and through coherence consistency testing, mixed wave superposition processing and quality control processing, a result after non-reflection weak signal enhancement is obtained. According to the invention, the enhanced processing effect of the shale oil non-reflection signal can be provided.

Disclosed is a Gabor wavelet-fused multi-scale local level set ultrasonic image segmentation method. In the method, non-uniformity of the grayscale of an ultrasonic image is taken as a texture having cluttered directions, the multi-directional property of Gabor wavelets is used to process the image, and intermediate images in different filtering directions are fused by taking maximum values, so as to obtain an intermediate image having a weakened texture effect and an enhanced difference between a foreground and a background. For the feature of a weak edge of an ultrasonic image, a concept of multi-scale is used to improve the conventional LIC method, Gaussian convolution kernels having different variances are set, and a final edge is obtained by means of average fusion.

The present application relates to the technical field of forestry pest remote sensing monitoring, and discloses a cypress pine moth identification method based on image fusion, which combines a wavelet fusion algorithm dynamically adjusted according to insect state and canopy density, introduces a special feature model for undergrowth interference and tree age correction, and a fusion weight dynamically optimized based on the analytic hierarchy process and measured data, to build a highly adaptive and accurate intelligent analysis core, which can ensure that the present application can still achieve a high-precision index of a larva identification rate of 90% or above and an insect state and damage level determination accuracy of 95% or above when facing different tree ages, different tree species and complex canopy structures, fundamentally solving the problems of early cypress pine moth pest misjudgment, misjudgment and poor scene adaptability, and finally realizing the whole-process accurate identification of early cypress pine moth pests in complex habitats.

This invention discloses an electrocardiogram (ECG) synthesis method and system based on adaptive dual encoders and wavelet fusion. The method first calibrates the time offset between the cardiac impulse signal and the reference ECG through cross-correlation alignment; it then constructs an adaptive orthogonal wavelet decomposition module using learnable perturbation terms to separate multi-scale components; subsequently, it employs parallel feature extraction using dual encoders in the time and wavelet domains, and achieves dynamic fusion of heterogeneous features through a gated cross-attention mechanism; a bottleneck layer introduces bidirectional temporal convolution to enhance long-range dependency modeling and correct phase offset; in the decoding stage, multi-scale wavelet heuristic decomposition guides upsampling to reconstruct fine-grained morphology. This invention effectively solves the problems of weak individualized modeling capability and loss of high-frequency details in existing technologies, possessing advantages of high fidelity and low computational complexity, and is suitable for real-time health monitoring of edge devices.

The invention is suitable for the technical field of image processing, and provides a multi-level wavelet fusion Transformer low-dose CT noise reduction method, which comprises the following steps: S110, collecting LDCT imaging data of a patient, collecting low-dose CT image data and normal-dose CT image data, and storing the data in a DICOM format; s120, zero setting is carried out on all pixel values, exceeding the CT scanning effective area (namely in the boundary), in the LDCT image, normalization processing is carried out in the range of [-1024, 3072], and then the pixel values are stored in an npy format; s130, constructing a low-dose CT (Computed Tomography) noise reduction model of a multi-stage wavelet fusion Transform, and recording the low-dose CT noise reduction model as an MWFormer; s140, training is carried out on the constructed MWFormer; and S150, the constructed MWFormer is tested, and a low-dose CT noise reduction result is obtained. According to the low-dose CT noise reduction method, the problem that more high-frequency information is possibly lost in the processing process of an existing low-dose CT noise reduction method and the problem that rich local information and global information cannot be obtained at the same time are solved through combination of the two ways.

The invention provides an SAR (Synthetic Aperture Radar) image change detection method based on weighted fusion wavelet transform and an improved residual network. The specific implementation steps are as follows: (1) acquiring a weighted fusion difference chart; (2) discrete wavelet decomposition is carried out on the weighted fusion difference graph; (3) reconstructing and generating a discrete wavelet fusion difference chart; (4) identifying and extracting key feature information; (5) pre-classifying the discrete wavelet fusion difference chart; (6) selecting a training sample and a test sample; and (7) performing final classification on the test samples by using the trained SEPP-ResNet model. According to the method, through weighted fusion of wavelet transform and SEPP-ResNet, the understanding of SAR images can be enhanced, and multi-scale context information can be fully excavated, so that the network can realize more accurate and reliable change detection in a complex environment.

The invention discloses a rotary machine composite fault diagnosis system and method based on a multi-complex wavelet fusion interpretable time-frequency network, and relates to the technical field of composite fault diagnosis, and the method comprises the steps: constructing an interpretable time-frequency network, introducing an asymmetric joint loss function to establish label correlation, and explicitly capturing the joint probability of a composite fault. The network framework adopts a multi-complex wavelet kernel convolution layer, five different complex wavelets are utilized to generate interpretable time-frequency representation, and the weight of each wavelet is dynamically adjusted through a kernel weight module to obtain fusion time-frequency information. Mixed multi-scale convolution is introduced to enhance multi-scale feature interaction of a time-frequency domain, and a dynamic time-frequency fusion module adaptively balances time-frequency domain information through a learnable time-frequency factor. According to the invention, data verification strategies under different scenes and working conditions are used to show excellent fault detection and analysis capability, and the reliability performance under various scenes is ensured.

This invention belongs to the field of nondestructive testing technology and proposes a full-focus imaging method for axial defects in pipelines based on multi-frame, multi-mode wave fusion. The method employs a detection system consisting of a full-matrix data acquisition instrument, a phased array probe, and angle wedges. Multiple frames of full-matrix data are acquired from different circumferential positions on one side of the pipeline to be inspected. For each reconstruction point within the imaging area, six wave modes are used to perform time-delay superposition and amplitude weighting processing on each frame of acquired full-matrix data. The strongest response from each frame is extracted and fused to reconstruct the contour of the prior unknown axial defect within the pipeline. This method considers the influence of pipeline curvature on ultrasonic wave propagation. By correcting the propagation time calculation, time-delay superposition and amplitude weighting processing, and multi-frame data fusion, it achieves imaging characterization and quantitative detection of prior unknown axial defects within the pipeline.

This invention belongs to the field of computer vision and deep learning technology, and provides a method and system for small target detection in UAV aerial photography based on frequency domain-wavelet fusion. A small target detection head branch is added after the second layer output of the backbone network: after downsampling to a preset pixel resolution, it is adaptively weighted and fused with the deep features output by the detection head through a structure. A decoupled head structure is used to output classification prediction, bounding box regression prediction, and confidence prediction respectively, for detecting small targets with a resolution smaller than the preset pixels. The feature map output by the backbone network is input into a LOWTC-based feature extraction module, which performs two-level wavelet decomposition, recursively convolving the low-frequency subband to capture long-distance context, and preserving edge details in the high-frequency subband. After reconstruction by inverse wavelet transform, it is fused with the deep semantic feature structure at multiple scales. This solves the technical problems of low detection accuracy and susceptibility to noise interference in small target detection in UAV aerial images.

The invention discloses a complex-valued wavelet depth-of-field fusion method and device, a storage medium and a program product. The method comprises the steps of collecting an original image, converting the image into a complex value wavelet domain, fusing the image according to an amplitude maximum principle and recording a layer number, using a highest frequency sub-band layer number to back-source construct a texture image, fusing the texture image, and using a low frequency sub-band cascade of the fused texture image to replace a low frequency sub-band of a preliminary fused image. And inverse complex value wavelet transform is carried out on the corrected fusion image. According to the method, the low-frequency substrate is corrected on the premise that error-prone high frequency is not directly carried, low-frequency drift, halation and local contrast abnormity caused by the fact that perfect reconstruction conditions are not met in traditional complex value wavelet fusion are relieved, meanwhile, the detail sharpness is kept, and the method is suitable for microscopic imaging, macro shooting and industrial detection.