258 results about "Context modelling" patented technology

The invention discloses a low-code development automatic generation method based on a large language model, which comprises the following steps: collecting natural language description information input by a user, and preprocessing; inputting the standardized demand corpus set into a large language model, and executing semantic understanding and context modeling; matching the low-code component library based on the structured semantic representation to generate component assembly description information; generating and verifying an engineering skeleton according to the assembly description information, and outputting an executable low-code application initial version; running the executable low-code application initial version, and monitoring and analyzing execution difference to generate an increment adjustment instruction; and inputting the increment adjustment instruction into the large language model, performing reconstruction and adaptive optimization, and outputting an executable low-code application final version. According to the method, large language model semantic understanding and adaptive optimization technologies are fused, automatic generation and continuous optimization of low-code applications are realized, and the method has the advantages of intelligence, high precision and engineering reliability.

The invention relates to the technical field of text feature fusion, in particular to an adaptive text guide fiber bundle feature fusion method and system based on a large language model. The method comprises the steps of performing multi-scale potential attention fusion based on features after fiber bundle feature fusion, including potential multi-scale feature embedding, inter-scale attention modeling, multi-scale fusion and residual enhancement, and attention regularization and scale consistency constraint; performing pseudo token sequence generation and aggregation based on features after multi-scale potential attention fusion, including pseudo token embedding representation, token serialization and slice mapping, semantic consistency and context modeling, and sequence aggregation and decodable interface generation; and processing the multi-modal fusion features subjected to pseudo token sequence generation and aggregation by using a large language model to generate natural language output. The invention provides an adaptive text-guided fiber bundle feature fusion method based on a large language model, which realizes efficient collaboration and semantic interpretable fusion of multi-modal information.

The invention discloses a dense small target detection method for an unmanned aerial vehicle aerial photography scene, and belongs to the technical field of computer vision and target detection. In order to solve the problems of small target scale dynamic change and feature expression weakening caused by flight height change, imaging resolution difference and scene complexity in aerial photography of an unmanned aerial vehicle, the invention provides a detection framework fusing an attention scale selection (AGSS) module and a dynamic local self-attention (DPSA) module. The method specifically comprises the following improvements: (1) an AGSS module enhances the significance and discrimination ability of small targets in multi-scale features through global context modeling and a dynamic weight distribution mechanism; and (2) a DPSA module introduces a sparse selection mechanism in a channel dimension, and focuses computing resources on a channel sensitive to a small target, so that efficient and lightweight attention modeling is realized. The above modules cooperate with each other, so that high reasoning efficiency is maintained, and small target detection precision and robustness in a complex background, low illumination and dense target scene are significantly improved. Experimental results show that on typical unmanned aerial vehicle aerial photography data sets such as VisDrone-DET2019 and the like, the method is superior to an existing mainstream method in multiple indexes such as the average precision (mAP), the accuracy rate and the recall rate, especially has obvious advantages in the aspects of integrity and stability of small target detection, and has good practical application value and popularization prospects.

The invention provides a geometric parameter measurement method based on underwater structure crack pixel-level semantic segmentation, and the method comprises the steps: constructing a multi-level feature extraction network based on ResNet-34, providing a structured feature pyramid for the subsequent context modeling and interference suppression, modeling illumination distribution features through a multi-head self-attention mechanism, and carrying out the modeling of the illumination distribution features through a multi-head self-attention mechanism. Regional scattering noise is extracted in combination with a cavity convolution pyramid, an environment context tensor is generated through fusion, a learnable gating mechanism is introduced for a decoding stage, dynamic modulation of jump connection characteristics and interference area response suppression are achieved, in addition, an edge perception auxiliary loss function is adopted, topological continuity of crack edges is enhanced, and finally, the method is applied to the field of jump connection. High-precision pixel-level segmentation of the underwater crack area can be realized without prior illumination correction, and the detection robustness and segmentation integrity in a complex environment are improved.

The invention belongs to the field of photovoltaic panel hot plate image processing, and particularly relates to a photovoltaic panel defect detection method fusing multi-scale wavelets and a lightweight attention mechanism. According to the method, an RHDWT discrete wavelet transform module based on multi-scale decomposition is added in a model input stage to strengthen image edge and texture detail representation; a Mix Structure Block module is introduced into a backbone network of the YOLOv11, so that multi-scale features are fused, and the feature expression capability is improved; an LWGA lightweight global attention mechanism is introduced into a neural network connection layer to enhance the context modeling capability, and the detection effect on small target defects such as fine cracks and hot spots is improved. According to the model, through collaborative optimization in three aspects of input preprocessing, feature extraction and an attention mechanism, the precision and robustness of defect detection in a complex photovoltaic module infrared or visible light image are remarkably improved, and the model is suitable for scenes such as high-precision photovoltaic panel image detection and intelligent maintenance.

The invention discloses a dynamic multi-scale convolution and cross-level attention feature fusion method and device, and relates to the technical field of image semantic segmentation. The basic strategy of the method comprises the following steps: receiving a feature map output by a decoder and a residual feature output by an encoder at the same level, carrying out dual calibration through channel attention and space attention, and adding a channel enhancement feature and a space enhancement feature to obtain a preliminary fusion feature; then, after cross-layer features and preliminary fusion features are introduced to be spliced in a channel dimension, a new feature map after dynamic fusion is output by means of dynamic multi-scale convolution, finally, local and global feature extraction is conducted on the new feature map, and fusion output is conducted after fine screening is conducted on the local and global features and the dynamic fusion features through a gating mechanism. Therefore, the cross-level feature fusion can be realized by taking the dynamic calculation as a means and the cross-level context as a link, the context modeling capability of the segmentation model is enhanced, and the capability of the image segmentation task is further improved.

The invention relates to a traffic accident intelligent detection system and method based on a YOLOv12 improved architecture. The system comprises a YOLOv12 enhanced feature extraction network, a multi-scale detection head, a time sequence information fusion module, a real-time reasoning optimization engine and an intelligent decision fusion system, and realizes collaborative optimization of local feature enhancement and global context modeling by constructing six core technology modules and adopting collaborative learning of a C2f-Attention mechanism and deformable convolution. According to the method, a composite loss function special for traffic accidents is innovatively designed, and adaptive fusion of multi-scale features and difficult sample mining are realized through a multi-objective optimization mechanism of Enhanced Focus Loss, IoU-aware Loss and Severage-aware Loss. According to the method, the problems of low detection precision and false alarm and missing alarm caused by illumination variation, shielding and motion blur in a traffic monitoring scene are effectively solved, in the test of an AccidentsDesection YOLOv8 data set, the mAP at 0.5 reaches 91.27% and is improved by 8.6% compared with that of YOLOv8, the reasoning speed reaches 67 FPS, experimental results show that the system has excellent performance in the aspects of detection precision, real-time performance and model compression, and the method is suitable for popularization and application. The method achieves a remarkable effect in traffic accident intelligent identification, and has a remarkable technical effect and industrial application value.

The invention discloses an event causal relationship identification method based on iterative graph prompt learning, and belongs to the technical field of natural language processing. Comprising the following steps: constructing an initial definite causal graph; based on the initial definite causal graph, performing context modeling on a to-be-recognized text, extracting multi-path information of each event pair in the to-be-recognized text, and inputting the multi-path information into a generative language model to perform causal relationship generation to obtain an initial causal relationship result; based on the initial causal relationship result, adopting a graph structure constraint mechanism to carry out multiple rounds of iterative optimization on the initial definite causal graph, dynamically selecting an edge according to confidence, updating the initial definite causal graph, judging whether an iteration termination condition is met or not, and stopping iteration when the iteration termination condition is met to obtain an optimized definite causal graph; and obtaining an event causal relationship result of the to-be-identified text based on the optimized definite causal graph. According to the method, the event causal relationship identification accuracy is improved.

The invention discloses a low-voltage series arc fault detection method, system and device, and belongs to the technical field of low-voltage series arc fault detection, and the method comprises the steps: obtaining an original current signal, and carrying out the preprocessing through sliding window segmentation and instance normalization; performing multi-scale feature fusion on the preprocessed analysis unit, and generating fusion features through parallel feature extraction and an attention mechanism; performing context modeling on the fused features through an encoder, inputting a self-adaptive bottleneck layer containing an expert hybrid network, and routing the features to the most appropriate expert network by context sensing gating according to global information; the decoder reconstructs the signal and calculates an error, and generates a dense abnormal fraction sequence; gaussian position weighted aggregation abnormal scores are adopted, and fault judgment is carried out in combination with a self-adaptive threshold decision mechanism based on K-Means clustering. The method can be trained without a fault sample, can dynamically adapt to complex current modes under different loads, gets rid of dependence on the fault sample, and accurately detects the arc fault.

The invention relates to a few-sample new-view-angle image synthesis method based on multi-scale mixed perception and state space collaboration, and solves the problems of depth ambiguity, excessive smooth texture and serious artifacts in a shielding region caused by insufficient geometric constraints in a generalized neural radiation field in a few-sample scene compared with the prior art. And the problems of high quadratic calculation complexity, large video memory occupation and low rendering efficiency caused by introducing 3D Transform to perform global modeling in the prior art are solved. And the defect of inaccurate geometric surface positioning caused by noise interference in the traditional sampling strategy based on attention weight is overcome. The method comprises the following steps of data set construction and sparse input acquisition; setting a hierarchical bidirectional feature aggregation network; multi-scale features are extracted; carrying out global context modeling; extracting and fusing local geometric features; and generating a new view angle image. According to the method, multi-scale features are extracted through the hierarchical bidirectional feature aggregation network, collaborative modeling of global context and local geometry is combined, depth ambiguity caused by insufficient geometric constraints under the condition of few samples is effectively solved, and the method is especially suitable for being used under the limited observation condition that only sparse source view angle images are provided. And performing high-quality new view rendering and image generation on an unknown scene.

The invention relates to the technical field of computer vision and video processing, and discloses a cross-domain multi-target automatic detection tracking association method based on an unmanned aerial vehicle platform, which comprises the following steps: firstly, obtaining an effective detection frame of a current frame through a target detector; meanwhile, a tracking bounding box is obtained through a tracker and a self-adaptive time sequence converter; then, executing three-level progressive data association: executing the first-level data association based on space association, and rapidly associating a target with continuous motion; the second stage performs re-recognition association on unassociated targets to process occlusion and reproduction; in the third stage, new track initialization is executed after multi-frame verification and quality evaluation are carried out on the detection frames which are still not associated; and finally, executing track life cycle management, and performing updating, state change or deletion on all tracks. According to the invention, through sequential context modeling and progressive fusion of depth features, the tracking precision, robustness and target identity retention capability in complex scenes (such as unmanned aerial vehicle maneuvering) are improved.

The invention discloses a remote sensing small target detection network and method based on frequency domain and space domain adaptive enhancement, and belongs to the technical field of artificial intelligence. The network is based on an FCOS lightweight detection framework, through cooperative extraction, adaptive enhancement and fusion of spatial domain and frequency domain features, the parameter quantity and the calculation complexity are greatly reduced, the small target detection precision and efficiency are significantly improved, and finally the balance of lightweight deployment and high-precision detection is realized. According to the invention, through the space-frequency feature adaptive enhancement network, space context modeling and frequency detail recovery can be completed in a single-stage detection framework at the same time, and an adaptive enhancement unit is introduced to dynamically adjust the enhancement strength according to the local signal-to-noise ratio, so that significant improvement and background suppression of a tiny target signal are realized.

The invention belongs to the field of hyperspectral image classification, and discloses a hyperspectral image classification method and system based on multi-scale spatial-spectral joint representation and dynamic context modeling, and the method comprises the steps: carrying out the feature dimension reduction processing of a hyperspectral image through principal component analysis; spatial spectrum collaborative information of hyperspectral data is deeply mined through a multi-scale spatial spectrum joint characterization module, and adaptive fusion and enhancement of spatial spectrum characteristics under different scales are realized; a dynamic context modeling strategy is introduced, and the perception ability of the model to context information is optimized by establishing a long-range dependency relationship between features; advanced feature integration and nonlinear transformation are carried out through a multi-layer perceptron, and precise classification of hyperspectral image ground objects is completed. According to the method, the performance superior to that of a current mainstream method is obtained on three public data sets, and the effectiveness and generalization ability of the method are verified.

The invention discloses a three-dimensional scene reconstruction method driven by monocular implicit depth features. The method comprises the following steps: extracting multi-view monocular depth features through a frozen monocular depth encoder; realizing cross-view feature alignment by using a geometry-vision similarity fusion module; generating a multi-scale monocular implicit depth feature through a multi-scale feature adaptive and processing module; a context modeling module and a feature pyramid network are adopted to extract multi-view context features; carrying out one-stage fusion with a monocular implicit depth feature to obtain a fusion feature; generating a depth map through the multi-view stereo depth estimation network; under the constraint of a depth map, generating spatial point features through multi-source feature projection sampling; performing two-stage fusion with the fusion features to generate point-level implicit features; optimizing the features through a frequency-space local-global gating mechanism; respectively inputting the optimization features into a Gaussian sputtering explicit rendering branch and a volume implicit rendering branch; and carrying out weighted fusion on the rendering result to generate a new view image.

The invention relates to the technical field of remote sensing images, and discloses a remote sensing image semantic segmentation method based on local feature enhancement, and the method comprises the steps: inputting a to-be-segmented remote sensing image into an improved remote sensing image semantic segmentation model, performing multi-scale feature extraction on the remote sensing image to be segmented through a local pyramid attention unit in a local feature enhancement module to obtain enhanced local features; inputting the enhanced local features into a global semantic capture module, and performing global semantic information extraction on the enhanced local features through a directional scanning module to obtain global semantic features; and performing cross-dimensional context information fusion processing on the global semantic features according to a context information sensing module to generate a semantic segmentation result of the remote sensing image to be segmented. Through the remote sensing image semantic segmentation model, the overall structure and local details of the image can be captured based on global and local context modeling, context information is effectively perceived and utilized, and the segmentation precision of the model is improved.

The invention discloses a target detection method and device based on deep learning and electronic equipment, and relates to the field of data processing. The method comprises the following steps: processing an input image to obtain a plurality of initial feature layers with different resolutions; performing global context modeling on each initial feature layer to obtain a plurality of feature layers, and calculating scale sensing foreground scores for a plurality of tokens in each feature layer; generating a confidence score of each feature layer, and screening out an optimal feature layer and a suboptimal feature layer; fusing a plurality of target feature layers, which are not screened out, in the feature layers to the optimal feature layer and the suboptimal feature layer respectively to obtain an enhanced optimal feature layer and an enhanced suboptimal feature layer; and performing sparse processing based on the scale perception foreground score to obtain a sparse token set, and inputting the sparse token set into a detector of a preset target detection model to obtain a target detection result of the input image. By implementing the technical scheme provided by the invention, the calculation redundancy and the memory overhead are reduced.

The invention provides a monocular depth estimation method and product based on convolution compensation dual-channel self-attention, and relates to the field of computer vision. Comprising the following steps: converting an event flow of a target scene into three-dimensional tensor representation; obtaining event image fusion multi-scale spatial features based on the image of the target scene and the three-dimensional tensor representation; modeling spatial context correlation in a spatial dimension by utilizing event image fusion multi-scale spatial features through a context modeling self-attention branch to obtain a context modeling self-attention result; through a modal fusion self-attention branch, using the event image to fuse the modal correlation of the multi-scale spatial features in the channel dimension modeling image and the event, and obtaining a modal fusion self-attention result; and pixel-level depth prediction is carried out by using a context modeling self-attention result and a modal fusion self-attention result to obtain a depth map, so that complementary characteristics between an event and an image are fully mined, fine-grained depth fusion expression is realized, and depth estimation precision and generalization ability are effectively improved.

The invention discloses a deep learning method for uniform background light and shadow based on a Transform model. The method comprises the steps that an input image and a corresponding portrait main body mask are acquired, and the input image comprises an area with uneven background light and shadow; preprocessing the input image and the portrait main body mask, including image standardization and feature fusion, to obtain a fused feature map; based on a patch segmentation method, converting the fusion feature map into sequence features; encoding the sequence features to output enhanced sequence features, modeling a light and shadow distribution dependency relationship of full image pixels through a global attention mechanism, and distinguishing portrait main body features and background features based on the portrait main body mask; and based on the inverse logic of patch segmentation, recovering the enhanced sequence features into a spatial feature graph and the like. According to the method, semantic mask prior, global context modeling and adaptive residual correction are organically integrated, and a solution is provided for solving the core problem in background light and shadow homogenization.

The invention provides a multi-agent collaborative cognitive calculation method based on point cloud feature marking, and belongs to the field of Internet of Vehicles. The method specifically comprises the following steps: firstly, inputting original point cloud data collected by a sensor of a vehicle intelligent agent into a point cloud feature mark generator, converting the original point cloud data into a one-dimensional point cloud feature mark sequence semantic perception dynamic encoder for encoding a mark sequence, and generating a feature sequence; packaging the feature sequence, the space coordinates and the pose information of the self-vehicle intelligent body into a message data packet; converting the feature mark coordinate space of the neighbor intelligent body into a coordinate system with the self-vehicle intelligent body as the center by a point cloud mark aggregation module, and generating a unified sequence; and the semantic perception dynamic fusion module carries out global context modeling and dynamic fusion on the unified sequence, corrects feature mark position deviation, generates a refined sequence, and inputs the refined sequence into a downstream task network for generating final prediction. According to the method, the perception robustness is obviously improved.

The invention provides a physical guidance-based sparse coding-driven underwater image enhancement method, a program, equipment and a storage medium. The method comprises the following steps: constructing a physical guidance preprocessing method based on a Jaffe-McGammy model, and adaptively estimating an underwater transmission image and background light to realize preliminary degradation recovery; elastic regularization sparse coding is adopted, and through sparsity and smoothness combined constraint, salient structure information such as edges and textures of the image is effectively extracted, so that the image definition and the detail recovery capability are enhanced; cross-channel attention fusion is adopted, a multi-scale context modeling and channel interaction mechanism is introduced, and a physical degradation graph is combined for guidance, so that color consistency and semantic expression integrity between channels are remarkably improved, and the problems of color shift and information loss in an existing method are solved. According to the method, through joint optimization of multiple loss functions, the constructed underwater image enhancement network keeps the illumination accuracy, texture continuity and structural integrity of the image in multiple dimensions at the same time.

The invention belongs to the technical field of image processing, and particularly relates to a passive domain adaptive eye fundus image segmentation method based on adaptive mask and curvature regularization, which comprises the following steps: constructing a teacher-student self-training framework, and generating a pseudo tag by using a weak enhancement teacher model to guide student model training, an adaptive mask consistency strategy is introduced in the training process, adaptive mask processing is carried out on a target domain image, the mask proportion and size are adaptively adjusted according to the sample difficulty and the size of a pseudo-label area, a model is guided to keep prediction consistency under the shielding condition, and thus the pseudo-label reliability and the context modeling capability are improved; meanwhile, by introducing an average negative curvature regularization constraint, an irregular boundary in a prediction result is suppressed, the smoothness and structural continuity of a segmentation result are enhanced, and finally the precision and generalization ability of the model in a cross-device and cross-dataset fundus image segmentation task are improved.

The invention provides an entity enhancement and context-aware paragraph retrieval method for an RAG system for solving the problems of fuzzy query intention and insufficient paragraph context modeling in RAG retrieval. The method comprises the following steps: firstly, identifying and extracting a key entity by using a named entity, carrying out weighted fusion on the key entity and question representation obtained by a pre-training model to form an enhanced query vector, and accurately describing a core semantic intention of the question; secondly, performing semantic modeling on paragraphs in a document library, mining a potential semantic association relationship between the paragraphs, and constructing a context interaction model between the paragraphs based on a graph neural network and a gating loop unit mechanism, so as to obtain paragraph vector representation with complete semantics and clear hierarchy; and finally, calculating the similarity between the enhanced query and the paragraph vector, and completing high-precision paragraph-level retrieval. According to the method, the correlation and the recall rate are remarkably improved, insufficient entity utilization and weak context modeling are relieved, and good expansibility and cross-domain applicability are achieved.

The invention discloses a self-adaptive completion selection method for a dynamic knowledge graph, and belongs to the technical field of knowledge graphs. According to the method, candidate entity representations are generated respectively by utilizing a self-attention mechanism and a cross attention mechanism by simultaneously constructing an isolated modeling path based on self information query and a context modeling path based on graph structure neighbor information aiming at the problem of missing entity complementation in a dynamic knowledge graph. And further introducing a branch selection strategy, dynamically evaluating prediction results of different paths according to query features, context scales or prediction confidence, and adaptively selecting an optimal information modeling mode. According to the method, noise interference introduced by redundant contexts can be avoided, the reasoning calculation cost is remarkably reduced, meanwhile, it is ensured that graph structures and time sequence information are fully utilized when the contexts are effective, the accuracy and robustness of knowledge graph completion are improved, and the method is suitable for large-scale and dynamically evolved graph application scenes.

The invention relates to a high-concurrency real-time dialogue state tracking and context management method, and belongs to the technical field of natural language processing and human-computer interaction. The method comprises the following steps: firstly, unifying multi-modal input into standardized semantic representation; then fusing the historical context and external knowledge to perform semantic analysis; the state change triggered by analysis is packaged into an event and the event is published to a distributed log; an independent service asynchronous consumption event is used for atomic updating of a global state through version control and an optimistic lock mechanism; the read-write performance is improved through multi-level cache and incremental compression; meanwhile, streaming maintenance is based on a long context of a graph, and resources are optimized through life cycle management. Through unified semantic representation, event-driven state management, long context modeling and optimistic concurrency control, a closed-loop collaborative system is constructed, and the concurrent processing capacity, context coherence and response robustness of a dialogue system are remarkably improved.

The invention discloses a text-oriented violent word abbreviation detection method, device and equipment and a storage medium, and relates to the technical field of natural language process.The method comprises the steps that preprocessing, candidate abbreviation dynamic positioning and context modeling are conducted on an input text, and a semantic relation graph is constructed; fusing the text semantic feature and the graph structure feature of the graph to generate a joint feature vector; performing multi-dimensional alignment on the vector and a preset violent knowledge graph, and calculating cosine similarity between the vector and each graph node; if the maximum cosine similarity exceeds a preset similarity threshold value and the hazard level of the corresponding node is higher than a preset hazard level, marking the candidate abbreviation corresponding to the maximum value as a violent word abbreviation; and if the predefined security counter-example word exists in the dynamic context window, clearing the mark. According to the method, the novel abbreviation variant of the violent word in the text can be accurately identified, the violent intention in an ambiguous context is distinguished, and the zero sample scene adaptability and the attack and defense resisting capability are improved.

The invention belongs to the technical field of remote sensing image segmentation, and particularly relates to a self-adaptive multi-view lightweight collaborative image segmentation network method, which comprises the following steps of: 1, acquiring a multi-view remote sensing image data set; step 2, inputting the image data into a self-adaptive multi-view feature extraction module, fusing information of different views, and automatically screening key features through a gating mechanism to suppress redundancy; step 3, performing multi-scale context modeling and detail enhancement on the features by a double-domain adaptive lightweight attention module, and improving the expression ability of space and channel dimensions; and step 4, the channel adaptive fusion module performs weighted integration on the multi-source features to realize semantic consistency enhancement and robustness improvement. Aiming at a multi-view remote sensing image segmentation task, innovative improvement is carried out in view collaborative modeling, attention optimization design, a feature fusion mechanism and the like, the segmentation precision, the calculation efficiency and the generalization ability are remarkably improved while the lightweight of the model is guaranteed, and the method is suitable for remote sensing intelligent interpretation application in a resource-constrained environment.

The invention belongs to the technical field of image processing, and particularly relates to a stem cell morphological-mechanical characteristic automatic extraction system driven by a double-Swinin-Unet model, which comprises the double-Swinin-Unet model, the double-Swinin-Unet model integrally follows an encoder-decoder normal form, and a double-decoder and multi-module fusion design is introduced in combination with a self-attention mechanism of Transform and a jump connection thought of Unet, so that the stem cell morphological-mechanical characteristic automatic extraction system driven by the double-Swinin-Unet model driven by the double-Swinin-Unet model driven by the double-Swinin-Unet model driven by the double-Swinin-Unet model is obtained. The core module comprises an encoder, an ASSP module, a dual decoder, an MFFM multi-feature fusion module and a prediction head; the interior of the encoder comprises an initial processing layer, a multi-stage down-sampling layer, a CAM channel attention module and a jump connection. The double decoders are used for realizing layered feature recovery and refinement and comprise a decoder 1 and a decoder 2, and the two decoders are symmetrical in structure. According to the method, the limitation of a single decoder in a complex scene can be solved, meanwhile, high-precision feature recovery is realized, the problem that Transform is insufficient in large-scale context modeling is solved, and rich semantic features are provided for the decoder.

The invention belongs to the technical field of augmented reality, and particularly relates to a six-degree-of-freedom pose estimation system and method based on a Mama network, and the system comprises an image feature extraction and context modeling module, a scene coordinate regression and coarse pose estimation module, a pose sequence construction module, a pose sequence bidirectional Mama coding module, and a pose residual prediction and refining module. In the first stage, a coarse pose of the first stage is obtained, and in the second stage, the pose of the first stage is refined. The method is completely based on a monocular RGB image, does not depend on additional sensor data such as a depth map, a point cloud and an IMU, and has the advantages of being easy to implement, light in model weight, high in reasoning speed, suitable for indoor and outdoor scenes and the like.

The invention provides a cross-scale feature collaborative attention fusion method based on deep learning. The cross-scale feature collaborative attention fusion method comprises the following steps: step 1, preparing a remote sensing target detection data set; step 2, constructing a scale sensing context module ISCE; step 3, constructing a multi-level semantic texture feature fusion module MSTFF; step 4, constructing a cross-scale feature collaborative attention fusion method CFSANet; 5, inputting the remote sensing data set into a cross-scale feature collaborative attention fusion method CFSANet for training; step 6, reasoning the test remote sensing target detection image by using the trained model weight; and 7, finally outputting a test result of the model. According to the method, semantic alignment and fusion between adjacent layers are emphasized, so that detail loss caused by resolution reduction is relieved by effectively utilizing information of different feature levels, the context modeling capability of a detector is effectively enhanced, and the method is a key for solving the bottleneck of a current remote sensing image target detection technology.

The invention provides a non-local feature enhancement-based landslide image semantic segmentation method with dual-path feature extraction. According to the method, the global modeling capability of Transform and the local sensing advantage of CNN are fused, a main and auxiliary collaborative coding framework is constructed through enhanced Mix Transform (EimT) and lightweight CNN, the global semantics and edge detail features of a landslide area are effectively extracted, and the problem of small target feature sparsity is solved. According to the model, a multi-scale fusion and non-local context modeling (Non-Local Block) strategy is adopted to cope with challenges such as complex landslide form, large scale difference and discrete distribution, and a Gated Hybrid Attention decoder is designed to strengthen the recognition capability of a small-scale target and a complex boundary. Experimental results show that the method is superior to a reference model SegFormer-B2 in all evaluation indexes, the landslide IoU is improved by 5.34%, the mIoU reaches 94.77%, the mF1 score reaches 97.28%, and the recognition precision of the small-scale landslide in the complex landform is remarkably improved.