736 results about "High resolution image" patented technology

The present application relates to the field of computer vision. Provided are a visual encoding method and apparatus, and a visual encoding model training method and apparatus, which are used for using the same visual encoding model to encode images of different resolutions, and are applied to encoding scenarios for images of more sizes. The visual encoding method comprises: first, acquiring an input image, wherein the input image may be a high-resolution image and may also be a low-resolution image; and then inputting the input image into a visual encoding model, so as to output visual encoding data, wherein the visual encoding model is used for dividing the input image into a plurality of image blocks according to positional embedding, extracting features from each image block, and outputting visual encoding data on the basis of the features of each image block and corresponding positional encoding, the positional embedding is obtained by means of adjusting initial positional embedding on the basis of the difference between the input image and a preset resolution, and the positional embedding may specifically comprise a matrix corresponding to the division of the input image

The invention belongs to the field of computer vision and industrial defect detection, and discloses a steel coil end face defect detection method and system based on CBAM-BiFPN and multi-loss optimization, and the method comprises the steps: image preprocessing and region extraction: extracting a steel coil end face region through OpenCV and other image processing methods; dividing the high-resolution image into a plurality of small blocks and performing data enhancement; the method comprises the following steps: constructing a YOLOv11 network based on CBAM-BiFPN, introducing a CBAM attention mechanism and BiFPN feature fusion structure, and constructing an improved YOLOv11 detection network; multi-loss function joint optimization training is carried out, and model training is carried out in combination with loss functions such as Focal Loss and WIoU; and fusion of detection results and defect reconstruction output: reconstructing original image defects of all tile image detection results in a space coordinate mapping and redundant region fusion mode, and realizing high-precision overall detection output. Compared with a traditional method, the method still has the high recognition capability in a complex background and low-contrast scene, the detection precision and stability are obviously improved, and higher industrial adaptability and practical value are achieved.

The invention discloses an image super-resolution method and system based on semantic perception tokens, and relates to the technical field of computer vision, and the method comprises the steps: generating semantic confidence and grouping information through the aggregation of content perception tokens, and decoupling a basic residual error into a texture enhancement and degradation inhibition guidance graph; in combination with a static semantic constraint mask and a sparse matrix multiplication mechanism, progressive focusing of attention is realized; a diffusion time step embedding and cooperative modulator is introduced, semantic guidance information is dynamically injected into a multi-step denoising process, adaptive attention features and diffusion reconstruction features are fused, and finally a high-fidelity and high-resolution image is output. According to the method, content-adaptive high-resolution image reconstruction is realized through collaborative modulation of a sparse attention mechanism guided by semantic grouping and diffusion denoising guided by semantic decoupling.

The invention discloses an image enhancement method and system based on semantic constraint degradation modeling. The method comprises the steps that semantic masks and multi-scale degradation features are extracted based on a low-resolution image used for training; performing deep fusion on the extracted semantic masks and the multi-scale degradation features based on a double-flow parallel architecture to generate semantic-structure fusion features; forming a multi-modal guide condition, taking the multi-modal guide condition and the semantic-structure fusion feature as input together, and reconstructing a high-resolution prediction image through a diffusion generation model; constructing a structure consistency optimization total loss based on the high-resolution prediction image and the corresponding target image, and optimizing a diffusion generation model based on the structure consistency optimization total loss; and inputting a low-resolution image to be predicted into the optimized diffusion generation model to obtain a high-resolution image corresponding to the low-resolution image. According to the scheme of the invention, comprehensive and refined understanding of low-resolution images is realized through multi-module cooperation and deep fusion.

The invention discloses an efficient image super-resolution reconstruction method and system based on degradation region guidance, and the method comprises the following steps: S1, carrying out the region-level degradation type recognition and severity quantification of an input low-resolution image, and generating a global degradation distribution map with spatial consistency; s2, according to the global degradation distribution map and in combination with semantic-texture collaborative features, repairing a region which is judged to be seriously degraded by adopting a high-capacity branch, and repairing a region which is judged to be slightly degraded by adopting a light-weight branch; s3, fusing the output of the high-capacity branch, the output of the lightweight branch and the global detail enhanced image to generate a final high-resolution image; wherein the global detail enhanced image is obtained by enhancing the semantic-texture collaborative features.

The invention discloses a single image super-resolution reconstruction method and system based on wavelet transform and cross-domain feature fusion. According to the method, firstly, a low-resolution RGB image is mapped to a high-dimensional feature space through a shallow feature extraction module; performing up-sampling and discrete wavelet decomposition on the features by using a wavelet feature mixing module to obtain multi-band features; low-frequency and high-frequency depth features are respectively extracted through a double-branch structure, cross-domain fusion is realized by means of a deformable cross attention mechanism, and the feature expression ability is enhanced in combination with residual connection; and finally, reconstructing a high-resolution image through convolution, up-sampling and regularization processing. In the training process, a pixel-level loss function is adopted to optimize network parameters, the multi-frequency-domain feature sensitivity is effectively improved, texture and structure information is balanced, the image contrast, definition and structural integrity are improved, and high-quality real-time super-resolution reconstruction can be achieved.

The invention discloses a method for detecting surface defects of a blade of a wind driven generator. The method comprises the following steps: acquiring a high-resolution image of the blade of the wind driven generator acquired by an unmanned aerial vehicle; the method comprises the steps of pre-constructing a neural network model based on a deep learning target detection framework, and processing an image; the neural network model comprises a backbone network, a neck network and a detection head; a double-attention mechanism module is embedded in the backbone network and the neck network and is formed by connecting a channel-space attention mechanism and a parallel attention mechanism in series; a multi-scale feature fusion and enhancement operator is introduced into the backbone network, and a detection head operator of an optimization model is adopted to process a feature map output by the neck network; and training the neural network model by using the data set, deploying the neural network model on an unmanned aerial vehicle computing platform after training is completed, and performing real-time defect detection and recognition on an input blade image. According to the method, high-precision recognition of fine and small target defects and high-efficiency operation can be met at the same time, and real-time, accurate and automatic defect detection of the blade image is realized.

The invention discloses a real-time monitoring and early warning method for deformation of a foundation pit support based on oblique photography, and relates to the technical field of photogrammetry and deformation measurement, high-resolution images of a support structure are acquired in a multi-angle manner through a multi-stage route, and a high-density three-dimensional point cloud reference model is constructed after preprocessing; then periodically acquiring a monitoring point cloud, and realizing high-precision space alignment with the reference model through control point coarse registration and iterative nearest point fine registration; and finally, calculating the three-dimensional coordinate deviation of each point through point cloud matching, extracting the full-surface deformation, and generating a deformation cloud picture and a statistical report. According to the method, the technical spanning from discrete point monitoring to full-field continuous monitoring and from low-dimensional data to true three-dimensional vectorization deformation analysis is realized, and the comprehensiveness and accuracy of deformation monitoring are improved.

The invention relates to the technical field of satellite hydrological data processing, and particularly discloses a GRACE data super-resolution network space downscaling method fusing geographic information and environmental variables, which comprises the following steps: S1, acquiring original resolution GRACE data and original GLDAS data of a research area, and preprocessing the data; s2, dividing the data obtained by preprocessing in the step S1 into a training set and a test set, and training the GRACE data space downscaling model by using the training set to obtain a trained discriminator and a trained generator; and S3, inputting the GRACE low-resolution data in the test set and the high-resolution environment variable at the moment corresponding to the data into the generator trained in the step S2, and finally obtaining a downscaled high-resolution GRACE image. The method not only can be used for dynamic monitoring of regional scale underground water reserves and flood risk assessment, but also can be expanded and applied to scenes such as agricultural drought monitoring and ecological hydrological process simulation.

The invention relates to a visual guidance robot automatic control method, system and device and a medium. The method comprises the following steps: firstly, acquiring a target image sequence and extracting initial visual features, and processing the initial visual features through a deep learning algorithm to output a high-resolution image; calculating a distance switching value based on the high-resolution image, and if the distance switching value exceeds a preset threshold value, performing scale distortion correction on the high-resolution image to obtain a target image; multiple frames of time sequence data are fused through a Kalman filtering algorithm, image distortion real-time compensation and feature point track stability judgment are synchronously completed, and a target positioning precision estimated value is output; and finally, according to the target positioning precision estimated value, determining the position of a grabbing point through a posture mapping rule of mapping preset grabbing feature points in a target coordinate system to a robot-based coordinate system, and based on the position of the grabbing point, optimizing and fusing track parameters through a Kalman filtering algorithm to generate adaptive track parameters. The method improves the positioning precision and track following effect of automatic control of the robot.

The invention relates to the technical field of circuit board conductivity detection, and discloses a circuit board conductivity detection method and system based on deep learning, and the method comprises the steps: obtaining high-resolution image data and circuit design parameters of a to-be-detected circuit board, and generating a circuit board feature data set; calculating theoretical impedance distribution of each area of the circuit board by using a physical constraint neural network based on the circuit board feature data set, and outputting an impedance prediction parameter set; inputting the impedance prediction parameter set and the circuit symbol rule into a neural symbol inference device to generate a conductance fault diagnosis result; and obtaining diagnosis results of a plurality of detection sites, executing cross-site diagnosis fusion, and outputting a unified circuit board conductivity detection report. The method overcomes the limitation that a traditional visual detection method cannot identify electrical characteristic defects, and solves the detection problem of hidden faults such as discontinuous impedance of the high-frequency circuit board.

A method for greatly reducing data storage requirements of autonomous robots capable of product inventory is described. Instead using high resolution images for label identification and position tracking, a lower resolution map can be searched. Human or machines can identify position of labels, and via a reverse mapping, the corresponding position of a label on the high resolution image can be identified. Except for those portions of the high resolution image showing a label, most of the high resolution image can be discarded. Further processing on the limited image subset can be used to read the labels.

To employ low resolution, noisy tone mapping operations for high resolution images, at least one raw image frame is converted to a first image at a higher resolution and a second image at a lower resolution. Tone mapping is applied to the second image to derive a third image at the lower resolution. Histogram matching and regularization are performed to determine a lookup table approximating histogram matching of the second image to the third image. A global gain map is derived based on luma for the first image and the lookup table. A local gain map is derived by up-sampling and denoising residual differences between the third image and the lookup table applied to the second image. Based on the global gain map and the local gain map, a total gain map is determined for tone mapping the first image to produce a fourth image at the first resolution.

A method for training a prediction model to generate a high-resolution image representing defects on a substrate from a low-resolution image of the substrate. The method includes inputting a first image and a reference image of defects on a substrate, which are representative of images captured using different image capture conditions, to a neural network. The neural network is executed to generate a predicted image in response to the first image. A loss function that is indicative of a difference between a defect distribution in the predicted image and a defect distribution in the reference image is calculated and the neural network is modified based on the loss function. The neural network may be trained until the loss function is minimized.

The invention discloses an unmanned aerial vehicle bridge detection method based on machine vision. According to the method, an unmanned aerial vehicle is used as an autonomous mobile platform, and full-coverage and high-resolution image data acquisition of structural surfaces such as a bridge deck, a beam body and a bridge pier is realized through systematic task planning. The collected data is processed by a motion recovery structure and a multi-view three-dimensional algorithm to generate a high-precision three-dimensional live-action model and a digital orthophoto map. On the basis, a deep learning target detection algorithm is adopted to perform intelligent analysis on the image, apparent defects such as cracks, spalling and corrosion are automatically identified, positioned and classified, and accurate quantification of parameters such as crack width and spalling area is realized. According to the method, the problems of high risk, low efficiency, high subjectivity, existence of detection blind areas and the like in traditional manual detection are effectively solved, and key technical support is provided for digital and intelligent transformation of bridge operation and maintenance management.

The invention discloses an image super-resolution reconstruction method and system based on a deep learning model, and the method comprises the steps: collecting a preset number of high-resolution images, and obtaining a degradation phenomenon which may occur during the imaging of a target imaging device; degenerating the high-resolution image according to a preset rule on the basis of a degeneration phenomenon possibly occurring during imaging of the target imaging equipment to obtain a corresponding low-resolution image; establishing a training data set according to the high-resolution image and the corresponding low-resolution image, constructing a preset neural network and a loss function, and inputting the training data set into the preset neural network for training until the loss tends to be stable to obtain an image super-resolution reconstruction model; and obtaining a to-be-reconstructed low-resolution image, and inputting the to-be-reconstructed low-resolution image into the corresponding image super-resolution reconstruction model to obtain a corresponding reconstructed high-resolution image. The low-resolution image generated by the method is closer to real degradation distribution, so that the quality of super-resolution reconstruction of the image is improved.

The invention relates to the field of digital image processing, in particular to a super-resolution imaging method based on focal plane splicing and adaptive fusion. According to the method, sub-pixel offset among nine CCDs is preset through hardware, and nine frames of low-resolution image sequences with accurate displacement are obtained in push-broom. A central image is taken as a reference frame, high-precision mapping is realized based on hardware offset, motion estimation errors are avoided, effective pixels are screened by calculating robustness weight, an anisotropic Gaussian kernel function with a self-adaptive local structure is constructed so as to maintain image edge and detail features, and each frame is accumulated to a high-resolution grid in a weighting mode, so that a high-resolution image is obtained. And a sample compensation mechanism based on cumulative robustness is introduced, a fusion strategy is adaptively adjusted in an information insufficient area, and finally a high-resolution image is generated through normalization. The method significantly improves the imaging quality, suppresses artifacts and noise, and is suitable for the field of satellite remote sensing.

The invention discloses a motor insertion piece positioning detection method using optical recognition, and particularly relates to the technical field of motor manufacturing and detection. Collecting a high-resolution image containing the insert structure; carrying out image gray level enhancement and reflection interference suppression; extracting insert edge gradient information, and constructing an insert boundary point set; fitting the profile of the insert according to the boundary point set, and correcting the profile through a shape prior model; extracting barycentric coordinates and center positioning features of the inserts, calculating offsets delta X and delta Y and an angular deviation theta, and constructing a position offset vector V; the V is compared with a set tolerance model T, and whether the insertion piece is qualified or not is judged; an error distribution thermodynamic diagram is generated based on the offset data of the multiple punching sheets, and trend evaluation of the mold state and the material stability is achieved; according to the method, high-precision, full-automatic and visual detection of the sheet inserting position can be realized, and the sheet punching quality control level is improved.

The invention relates to the technical field of video image processing, and discloses a video image super-resolution enhancement method based on a generative adversarial network. The method comprises the following steps: segmenting a low-resolution video sequence, and analyzing frame timestamp information to generate a dynamic time axis containing key frame nodes so as to reflect video time change characteristics; then, in combination with the time axis and an image feature extraction system, feature mapping is carried out on the low-resolution frames, and a preliminary high-resolution image sequence with feature information is obtained; enhancing the preliminary sequence by using a generative adversarial network containing a generator and a discriminator in combination with a context-aware optimization technology to generate an enhanced sequence; and a machine learning algorithm is introduced to identify and classify different image quality region characteristic modes, and an intelligent identification rule base is established, so that an enhanced sequence is optimized. And finally, in response to user interaction, adjusting display parameters to realize personalized display, integrating a multi-dimensional analysis tool, calculating details of a specific region, and presenting a result in a visual interface, so that the video image quality can be improved.

The embodiment of the invention provides a degeneration perception prior and attention fused blind image super-resolution method. The method is applied to the technical field of image processing, and comprises the following steps: inputting a to-be-processed image into a trained degeneration perception priori and attention fused blind image super-resolution network for analysis and processing to obtain a reconstructed image; the network comprises a pixel embedding layer, a degradation feature extraction layer, a feature encoder and an image reconstruction layer. Performing feature extraction on the to-be-processed image through the pixel embedding layer to obtain first feature representation; performing feature extraction on the to-be-processed image through the degradation feature extraction layer to obtain a global degradation vector; inputting the first feature representation and the global degradation vector into a feature encoder for processing to obtain a second feature representation; and performing feature fusion on the first feature representation and the second feature representation, and inputting a fusion result into an image reconstruction layer for image reconstruction to obtain a restored high-resolution image, thereby improving the quality and authenticity of the generated image.

The invention discloses a remote sensing image super-resolution reconstruction method based on a diffusion bridge model. The method comprises the following steps: firstly, constructing a diffusion bridge model, and connecting low-resolution and high-resolution image endpoints; secondly, explicitly aligning states of an image recovery bridge process and a pre-training generation diffusion process by designing a state alignment mechanism so as to multiplex powerful generation prior of a pre-training model; meanwhile, a lightweight feature adapter is introduced into a decoding end of the pre-training model, and high-frequency detail features reserved by an encoder are injected into a decoding path through a gating residual mechanism so as to suppress an over-smoothing phenomenon caused by VAE decoding. According to the method, the problems of insufficient prior utilization and high-frequency detail loss of a traditional diffusion model in remote sensing image super-resolution reconstruction are effectively solved, the structural integrity and visual fidelity of the image are ensured while the reconstruction efficiency is remarkably improved, and the method is suitable for a multi-scene remote sensing image super-resolution task.

The invention relates to the technical field of semiconductor detection, and discloses a method and a system for detecting chips on the surface of a semiconductor wafer. The method comprises the following steps: acquiring a high-resolution image of the surface of a wafer, and separating a defect candidate region set from a reference background region through noise filtering and contrast equalization processing; extracting defect areas to be identified one by one, and accessing the defect knowledge graph to obtain potential defect types; performing multi-feature fusion on the potential defect type and the reference background region, generating a defect semantic feature vector through a context sensing encoder, and analyzing the vector to judge the actual defect type; and processing all the candidate areas and then outputting a defect detection report. According to the method, the distinction degree of defects and backgrounds is enhanced, the defect judgment range is narrowed, similar defects are accurately recognized, missing detection and false detection are reduced, the detection efficiency and accuracy are improved, and reliable technical support is provided for wafer production quality control.

The invention relates to the technical field of image data processing and detection, and discloses an image data processing and detection method of an OLED packaging structure, and the method comprises the steps: obtaining a high-resolution image of an OLED packaging region; extracting texture and edge features of the packaging layer through a multi-scale feature fusion network; the response of the tiny defect area is enhanced in combination with an attention mechanism; and classifying and positioning defects such as bubbles, cracks and foreign matters by using the trained deep learning model. In order to solve the problems of fuzziness and breakage of the bubble edge of the OLED packaging layer caused by gradual change of the refractive index of a material, a multi-scale Gaussian derivative filtering and path integral edge connection technology is innovatively adopted, gradient responses of different scales are dynamically weighted and fused, weak gradient edge signals are effectively enhanced, meanwhile, the broken edge is searched and repaired through a graph node path, and the defect that the bubble edge is broken is overcome. And complete capture of the bubble contour is realized.

The invention discloses a cable strength detection method and system, relates to the technical field of cable detection, and solves the technical problems that the metal cable detection process is not standard, the defect positioning and quantification precision is low, the contradiction between the non-metal cable detection efficiency and precision is prominent, and microscopic verification is lacked. A high-frequency arc-shaped phased array probe, a bicrystal oblique TOFD probe and an oil-based high-viscosity coupling agent are adopted, a layered focusing technology is combined, laser ultrasonic non-contact detection and ultrasonic microscope high-resolution imaging are adopted for a non-metal cable, and spiral scanning and circumferential fixed-point scanning are combined, so that the defect omission ratio in the circumferential direction is reduced; the abnormal region judgment takes a calibration test block as a reference, the misjudgment rate is reduced, the authenticity is verified through coarse scanning and locking of a core region, fine scanning and acquisition of a high-resolution image and multi-angle imaging, the defect type judgment basis is clear, and the misjudgment rate is reduced.

The invention discloses a rubber surface defect real-time detection system and method based on a knowledge graph, and relates to the technical field of rubber product quality detection. Comprising the following steps: S1, acquiring a high-resolution image of a rubber surface in real time; and S2, based on the high-resolution image, visual features of potential defects are extracted by using a deep learning model, and the visual features comprise texture, color, shape, size and position information. Preprocessed high-resolution rubber surface images are obtained in real time through the high-speed camera and the stable light source, and the problems that traditional manual detection is low in efficiency and prone to being influenced by subjective factors, and missed detection is caused are solved; multi-dimensional visual features such as textures and colors are extracted by using a deep learning model, mapping of visual features and semantic concepts is established through logical reasoning and semantic fusion in combination with a knowledge graph in which defect semantic knowledge is stored by a triple, and the defect that existing computer visual detection lacks deep semantic understanding is made up.

The invention relates to the technical field of remote sensing image processing, in particular to an image super-resolution system and method based on high and low frequency separation perception Mama, and the method comprises the steps: firstly carrying out the shallow convolution feature extraction of a low-resolution image; then entering a plurality of frequency sensing Mama groups, performing frequency separation and enhancement on each group through a high and low frequency feature adaptive enhancement module, and performing depth feature transformation through a plurality of frequency sensing Mama blocks; the extracted depth features are refined through a global channel-space attention module, and finally a high-resolution image is reconstructed through up-sampling. Through organic combination of the modules, the defects of insufficient frequency perception, low global modeling efficiency, insufficient feature optimization and the like are effectively overcome, and high-quality collaborative reconstruction of remote sensing image structures and textures is realized.

The invention discloses an X-ray strain clamp defect detection method based on an improved YOLO11 model, and relates to the technical field of artificial intelligence, deep learning target detection and power system equipment nondestructive testing. According to the X-ray strain clamp defect detection method based on the improved YOLO11 model, a portable X-ray imaging system is carried by an unmanned aerial vehicle to collect a high-resolution image, and a professional data set of six types of defects is constructed by combining data cleaning, labeling and enhancing; an improved YOLO-CLAMP model fused with Res2Net-C3, IBN normalization, a C2PSA-CAA attention mechanism and other modules is designed, and the detection capacity for small targets, fuzzy boundaries and overlapped targets is remarkably improved; according to the method, optimization training is carried out by adopting SIoU Loss and Focal Loss, and the method is deployed on an edge computing platform through TensorRT, so that real-time reasoning is realized. According to the method, the average detection precision on a test set is high, the omission ratio is obviously reduced, the single-frame reasoning delay is low, an efficient and accurate technical solution is provided for intelligent inspection of the power transmission line, and the method has remarkable social and economic benefits.

The invention discloses a remote sensing rapid diffusion super-resolution reconstruction method based on adaptive scene perception, belongs to the technical field of mode recognition, and is used for super-resolution reconstruction of remote sensing images. The method comprises the following steps: preprocessing a high-resolution image, and calculating a residual error between the high-resolution image and a bicubic up-sampling result as a training target of a diffusion model; performing wavelet transformation on the bicubic up-sampling image through a frequency domain enhancement module, extracting frequency domain features, and fusing the frequency domain features with original spatial domain features to serve as condition input of a model; inputting the fused features into a conditional diffusion model, carrying out adaptive rapid sampling by adopting a linear-cosine hybrid scheduling strategy, and completing high-quality reconstruction within 20 steps; the feature expression capability is enhanced through an overlapping and crossing attention module integrated in the U-Net denoising network; and finally, fusing the residual error reconstruction module and the bicubic up-sampling image to obtain a final super-resolution reconstruction result. Experiments carried out on two common data sets (Potsdam and Toronto) show that the method of the invention significantly improves the sampling efficiency while maintaining the reconstruction quality, and is superior to an existing remote sensing image super-resolution method.

The invention provides a remote sensing image small target detection method. The method comprises the following steps: constructing a double-branch detection network comprising a shared backbone network, a super-resolution branch and a target detection branch, fusing deep and shallow feature maps by using the super-resolution branch in a training stage to reconstruct a high-resolution image, and constraining the texture and semantic feature extraction capability of the backbone network on a small target; meanwhile, the target detection branch generates a detection result based on multi-scale feature fusion, and the sensing ability of the network to a remote sensing small target is enhanced. In the application stage, the super-resolution branch is forbidden, the low-resolution image is processed only through the shared backbone network and the target detection branch, and the reasoning calculation overhead is remarkably reduced. Therefore, the framework not only optimizes the feature extraction capability of the backbone network through super-resolution task optimization and improves the small target detection precision through super-resolution task optimization, but also avoids extra resource consumption during reasoning through a branch dynamic switching mechanism, and realizes the balance between the detection performance and the calculation efficiency.

The invention relates to the technical field of computer vision and target detection, and discloses a small target detection network for adaptive fine-grained feature mining. The invention provides a network, and the network overcomes the contradiction between the calculation efficiency and the detection precision of a traditional method through collaborative design of adaptive fine-grained feature mining and RoI feature interaction. The network can detect a high-resolution image by using difficult areas in a conventional-level feature map and a high-resolution shallow-layer feature map at the same time, the difficult areas with dense information are automatically positioned through a foreground probability discriminator, background redundancy calculation is avoided by using an iterative mining strategy, and the small target detection speed is increased; meanwhile, the Inter-RoI feature interaction module realizes bidirectional complementary enhancement of deep semantics and shallow details in a key difficult area, and in combination with a high-resolution detection head and a result fusion mechanism, the feature characterization capability of a small target can be enhanced under a complex background, and finally, the feature characterization capability of the small target can be enhanced while the high reasoning efficiency is kept. And the detection precision and robustness of small targets which are non-uniformly distributed and have weak features in the aerial image are improved.