474 results about "Saliency map" patented technology

The invention discloses an optical remote sensing image salient target detection method based on progressive attention enhancement, and belongs to the technical field of computer vision. The method comprises the following steps: preprocessing an original data set; inputting the preprocessed image into a hierarchical progressive fusion encoder, capturing a global irregular topological structure and local fine-grained image details, and realizing cross-hierarchical feature fusion; inputting the output characteristics of the encoder into a global context enhancement module, and capturing multi-level context information by adopting a parallel multi-branch structure; and inputting the output features of the hierarchical progressive fusion encoder and the global context enhancement module into a multi-scale progressive attention enhancement decoder, carrying out hierarchical decoding on the input features by adopting a saliency-guided attention mechanism, and gradually aggregating deep semantic information and shallow detail features to realize coarse-to-fine progressive optimization, so as to improve the robustness of the multi-scale progressive attention enhancement decoder. And finally generating a saliency map. The method can effectively improve the processing performance of an irregular topological structure and a complex context relationship in the optical remote sensing image.

The invention relates to the technical field of industrial machine vision detection and image processing, in particular to an automobile part production mold surface smoothness detection system based on image enhancement, which comprises a data acquisition module used for acquiring an original grayscale image of the surface of an automobile part mold; performing low-pass filtering processing on the original grayscale image to eliminate imaging thermal noise; the manifold reconstruction module is used for constructing a structure tensor field; reversely deducing a pseudo-curvature field of the mold surface; the adaptive enhancement module is used for generating a corrected image; constructing a texture orthotropic diffusion model; generating a texture reconstruction reference image; the surface metering module is used for calculating the difference between the corrected image and the texture reconstruction reference image and generating a defect saliency image; calculating the surface roughness value of the mold surface; according to the method, the problem that design textures and abnormal scratches are difficult to distinguish in the prior art is effectively solved, and the technical bottleneck that micro defects are easily missed in a complex geometric structure in traditional visual detection is overcome.

The invention relates to the technical field of intelligent navigation, financial science and technology and medical health, and discloses a navigation method, device and equipment based on multi-modal knowledge enhancement and a medium, and the method comprises the following steps: obtaining a navigation instruction, and analyzing the navigation instruction to obtain a target constraint, a spatial constraint and an entity constraint; generating a structured target description text based on the target constraint, the spatial constraint and the entity constraint; an input image is acquired, a semantic saliency map is generated based on the input image and the structured target description text, and the semantic saliency map comprises a plurality of candidate areas; calculating a score of each candidate region of the semantic saliency map based on a preset scene knowledge graph to obtain a fusion score; and performing path planning according to the fusion score of each candidate region to obtain a target path. Navigation can be realized without navigation data training, and meanwhile, a potential target can be focused according to semantics, so that the navigation precision is improved.

The invention relates to the technical field of image processing, in particular to an image acquisition and analysis method and system, and provides the following scheme: obtaining a visible light and near-infrared multispectral image, generating a spectral difference image, and performing weighted fusion to obtain a first image; segmenting a target region based on the fused saliency map, and calculating a pixel reflectance ratio; solving a color mapping matrix according to the reflectance ratio, and carrying out color correction on the target region to obtain a standardized feature image; and extracting characteristic parameters such as spectrums, colors and textures, inputting the characteristic parameters to a multi-branch convolutional neural network, fusing the characteristic parameters through an attention mechanism, and outputting a state classification result and a quantitative index. The cross-spectral imaging difference can be adaptively compensated, and the fusion precision and the analysis stability are improved.

The invention discloses an electric power operation unmanned aerial vehicle safety supervision target identification and detection method, system and device, and a medium. The method comprises the following steps: carrying out electric power operation risk identification based on environment characteristics and task attributes of an electric power operation area; configuring a monitoring frequency, and performing data acquisition on a power operation area; constructing a target region segmentation device based on a predetermined operation monitoring target, performing target segmentation on the monitoring image, and outputting a region image set; traversing and comparing the regional image set acquired under the current monitoring node with the historical regional image set acquired under the previous adjacent monitoring node, selecting images with significant differences, and performing point cloud fitting to construct a three-dimensional target simulation model; and building an anomaly identification plug-in based on a predetermined monitoring index and the operation early warning area, carrying out anomaly detection on the three-dimensional target simulation model, and outputting an operation detection result. According to the scheme, global intelligent monitoring of electric power high-altitude operation is achieved, potential safety hazards are found in time, and the timeliness and accuracy of risk identification are remarkably improved.

The invention relates to the technical field of image processing, in particular to a high-voltage electrical equipment surface defect identification method based on image processing. The method comprises the following steps: analyzing the gradient of pixel points in a to-be-analyzed image of a to-be-detected area on the surface of the high-voltage electrical equipment to obtain the weight of each pixel point; weighting the gray value of each pixel point in the to-be-analyzed image subjected to Laplacian filtering by using the weight of each pixel point of the to-be-analyzed image to obtain a first feature map; calculating a local standard deviation of each pixel point in the to-be-analyzed image so as to obtain a first parameter and a second parameter; constructing two Gaussian kernels based on the first parameter and the second parameter to filter the to-be-analyzed image to obtain a second feature map; fusing the first feature map and the second feature map of the to-be-analyzed image to obtain a defect saliency map of the to-be-analyzed image; and recognizing a surface defect area of the high-voltage electrical equipment based on the defect saliency map of each to-be-analyzed image. According to the invention, the accuracy of high-voltage electrical equipment surface defect identification can be improved.

The invention discloses a multi-modal image matching method and system based on saliency graph structure enhancement, and belongs to the field of image processing. The method comprises the following steps: firstly, innovatively constructing a pixel-level saliency confidence graph for measuring the matching potential of each region, and guiding an attention mechanism to be dynamically focused on a key region in a graph structure through the graph; secondly, multi-scale structure features and semantic segmentation information are fused, and the semantic perception ability of feature expression is enhanced; and finally, constructing two heterogeneous graph structures of an in-image structure graph and an inter-image semantic guidance graph, and realizing global-local information enhancement and cross-modal semantic alignment on the graph structures by introducing a self-attention and cross-attention mechanism of saliency modulation, so that the matching precision and stability are remarkably improved, and the matching accuracy is improved. And semi-dense matching of multi-modal images is realized.

The invention discloses an unmanned aerial vehicle outdoor inspection method based on multi-modal fusion, and the method comprises the following steps: obtaining the inspection data of an unmanned aerial vehicle, and carrying out the feature extraction and space-time alignment; calling a GNPDE algorithm, and constructing a multi-modal field mapping layer; constructing a structure-thermal field joint graph structure, and establishing a double-branch coupling solution structure; constructing a topological adaptive edge weight adjustment module, and dynamically modulating the edge weight by adopting a physical modulation function; introducing an energy conservation constraint layer, and executing constraint solution; a multi-scale PDE evolution algorithm subset is quoted and combined, and a scale weight sharing mechanism is adopted to complete cross-scale joint optimization; constructing an abnormal residual reasoning module, and generating an abnormal significance map; and performing spatial registration and superposition on the abnormal saliency map and unmanned aerial vehicle inspection data, and outputting an equipment-level inspection report and a risk level conclusion. According to the invention, the inspection abnormity identification precision and the structure-thermal field reasoning stability are improved.

The present disclosure describes methods, systems, apparatus, and media for object identification and classification, utilizing multi-feature and multi-modal data. This includes shape, material, brand, price, odor, taste, tactility, and sound. The system integrates a server space for data processing, a querying device for iterative searches, and a data interface module for refining results. It features AI-driven image optimization, feature extraction, and pattern recognition, employing novel techniques for fusing multi-feature and multi-modal embeddings utilizing multi-head attention. Additionally, a linker module powered by two active learning with feedback loops AI models consolidates scattered data into a unified object information database. The system also employs novel AI algorithms for isolating the object of interest through a saliency map and semantic analysis, as well as for enhancing raw images with a GAN-autoencoder.

The invention discloses a dynamic and static bidirectional interactive cooperative micro-expression recognition method, which comprises the following steps of: firstly, introducing a double-attention guided feature fusion module into a dynamic branch, and fusing optical flow and pixel difference between a micro-expression starting frame and a vertex frame to represent a micro-expression motion feature; a gating cross-layer feature transfer mechanism is proposed to screen key motion features and inject the key motion features into a deep network in a cross-layer manner, so that noise interference is effectively suppressed, and the problem of micro-expression signal attenuation caused by continuous down-sampling is relieved; the method comprises the following steps: firstly, designing a static branch, secondly, designing a mixed attention Transform block in the static branch to generate a micro-expression saliency map and a token redistribution Transform block guided by the saliency map, accurately extracting multi-granularity static characteristics of a key area in a vertex frame, and finally, adopting a bidirectional interactive attention module to deeply and interactively fuse dynamic and static characteristics to realize accurate classification of micro-expressions.

The invention belongs to the technical field of image fusion, and particularly relates to an infrared and low-light-level image fusion improvement method based on rolling guide filtering, which comprises the following steps of S1, acquiring a low-light-level image and an infrared image, and preprocessing the infrared image for denoising; s2, converting to an IHS space to extract an intensity component; s3, decomposing the dual-source I component into a small-scale layer, a large-scale layer and a base layer by using rolling guide filtering; s4, performing small-scale layer fusion, and combining an absolute value maximization strategy with non-local similarity weighted denoising; s5, performing large-scale layer fusion, and performing joint sparse low-rank representation by adopting principal component analysis; s6, fusion of the base layer adopts a visual saliency mapping method; and S7, after IHS inverse transformation reconstruction, realizing multi-level fusion in combination with bilateral filtering detail enhancement and a guide filtering saliency map. According to the invention, based on IHS space and rolling guide filtering multi-scale decomposition, the effect of cooperative utilization of multi-modal information is improved, and the method effectively improves the problems of insufficient target detection precision and poor scene reconstruction robustness in a dark environment.

The invention relates to the technical field of printed matter quality detection and computer vision and the technical field of artificial intelligence systems in the production field, in particular to a color box printed matter printed pattern visual detection algorithm. The invention discloses a visual detection algorithm for a printing pattern of a color box printed matter. The visual detection algorithm comprises the following steps of image acquisition, preprocessing, feature extraction, reference comparison, defect evaluation and defect classification. According to the scheme, the accuracy and the real-time performance of complex printing pattern defect detection are improved. A light-weight multi-scale attention convolutional neural network and a self-learning mechanism are combined, defects in various printing patterns such as block colors, fine characters and graph gradient are effectively detected, and human intervention and misjudgment are reduced. A defect saliency map is generated through color normalization, geometric correction preprocessing and reference map comparison, an online misjudgment cache and incremental learning strategy is introduced, self-adaptive optimization of a model along with time is kept, and real-time high-precision quality monitoring of an offset printing color box production line is achieved.

The invention discloses a building outer wall defect detection method based on infrared polarization imaging, belongs to the field of image processing, and aims to improve the accuracy and robustness of building outer wall defect detection. The method comprises the steps of firstly collecting building outer wall surface image data, and constructing an image data set; designing a channel disturbance intensity factor and a mutual information measurement strategy, constructing a multi-channel disturbance enhancement attention module, and extracting a multi-scale feature map; constructing a position sensing context fusion module, and enhancing a feature fusion effect through a position modulation coefficient; constructing a geometric guidance saliency modeling module, and generating a structural perception saliency map through the geometric guidance deviation factor; the above modules are integrated to construct a defect detection model, a to-be-detected image is input into the model, a defect area is output, and accurate detection and positioning of building outer wall defects such as cracks, falling and bumps are achieved.

The invention discloses an oil and gas pipeline magnetic flux leakage image defect identification method based on a deep attention mechanism, relates to the technical field of oil and gas pipeline detection, and is used for solving the problem of inaccurate identification of a magnetic flux leakage image of a pipeline elbow section. According to the method, the image frame sequence with the posture annotation is constructed through unified time reference and space coordinate mapping, and accurate alignment of the image and the pipeline position is achieved; a structural area marking graph and non-rigid normalization are introduced to compensate the distortion of the elbow section, and the image consistency is improved; constructing a structure perception embedded image on the distortion compensation image, fusing position, gradient and texture features to implement feature propagation and attention guidance, and generating a feature saliency map; a defect area is accurately extracted in combination with layered reconstruction and a two-stage judgment strategy; through continuous frame monitoring and recognition stability judgment, recognition parameters are adaptively updated, a closed loop from recognition to updating to verification is constructed, and the precision of oil and gas pipeline magnetic flux leakage image recognition under complex working conditions is enhanced.

The invention relates to the field of image processing, in particular to a box-type substation spraying quality detection method based on image processing, and the method comprises the steps: firstly obtaining a to-be-detected image, and calculating a pixel point structure tensor component after the to-be-detected image is preprocessed; extracting a gradient energy field and a gradient main direction angle field based on the component, and further calculating gradient radial consistency representing geometric characteristics of the defect; in combination with the to-be-detected image and the gradient energy field, respectively calculating saturation and flatness indexes representing optical reflection characteristics and intensity-gradient cross-correlation representing defect spatial forms; fusing the three features to generate a final saliency map; and finally, carrying out threshold segmentation and post-processing on the saliency map to obtain a detection result. According to the method, the physical spraying defect and the optical highlight reflection can be effectively distinguished, and the defect detection accuracy and robustness are remarkably improved.

The invention relates to the technical field of industrial vision, in particular to an aviation part crack detection and repair method which comprises the following steps: acquiring an aviation part optical image at a reference time point and an aviation part optical image at a to-be-detected time point; according to the method, logarithmic polar coordinate transformation is carried out on different time point images, rotation, scaling and translation parameters are extracted, an image registration relation is established, the structural consistency of time sequence images in a local area is enhanced, a structural tensor is constructed for each pixel neighborhood, the change characteristics of the dual-time-phase tensor are compared, a structural change saliency map is generated, and the structural change saliency map is obtained. Sensitive capture of a tiny deformation area is achieved, the responsiveness to an initial crack is improved, then a crack propagation interval is further refined into a main crack path in a self-adaptive threshold segmentation and skeleton extraction mode, the tip acutance of the crack is calculated in combination with tip contour information of a geometric boundary of the crack, and the initial crack is obtained. And quantitative support is provided for the crack danger degree.

The invention discloses an intelligent target detection method and device based on computer vision, and relates to the technical field of visual recognition. The method comprises the following specific implementation steps: step 1, detecting and positioning an area possibly containing a small target in an image by using saliency; step 2, carrying out saliency map fusion; step 3, carrying out super-separation treatment; step 4, adaptively cutting the image, and segmenting the large image by adopting a sliding window; step 5, combining repeated detection results; step 6, carrying out weighted fusion; according to the method, a set of complete processing flow from global saliency detection, candidate region screening, local super-resolution reconstruction, adaptive cutting to result fusion is formed through Step1 to Step6, and the accuracy and precision of small target detection are improved.

The invention provides an electric power line defect detection method and system based on visual identification, and relates to the technical field of line detection.The method comprises the steps that firstly, a visible light and infrared image dual-light registration and differential operation technology is adopted, and a fusion feature map capable of reflecting component thermal anomaly and material difference at the same time is generated; secondly, a black box type target detection model is abandoned in a part positioning link, but line segment screening and reconstruction are carried out by combining probability Hough transform with specific prior geometric knowledge of a power line, so that dependence on a large amount of labeled data is reduced, interpretability of a positioning process is enhanced, and the positioning accuracy is improved; according to the method, accurate areas of key components such as wire insulators can be extracted in a complex background, the concept of a probability saliency map is introduced in a defect identification core link, so that a defect area is enhanced and highlighted, and then a complete defect contour is determined by adopting an adaptive threshold segmentation and area growing algorithm; accurate mapping from pixel-level features to object-level defects is realized.

The invention relates to the technical field of artificial intelligence, in particular to a class activation mapping interpretability method fusing a spatial perturbation mechanism, which comprises the following steps: a first stage: capturing contribution of each channel to prediction, and obtaining a channel attention weight of a deep learning model through a gradient method; in the second stage, disturbance analysis is introduced, and space importance weights given to image samples by the model are obtained; and a third stage: fusing the interpretation information of the two view angles through weighted integration, and generating an interpretation saliency map of the region-level granularity. According to the invention, the dual positioning capability of an interpretable technology in a channel domain and a space domain can be improved.

The invention discloses a self-adaptive diffusion image editing method and system based on concept attention, and the method comprises the following steps: constructing a paired data set; analyzing the editing instruction, and extracting a key concept; a pre-trained T5 language model is utilized to convert the key concept into text embedding, and the text embedding is mapped to an image feature space; modifying a diffusion model based on a Transform architecture, embedding a concept attention module in an attention layer of a multi-modal diffusion converter, calculating an attention score between image features and concept embedding, and generating a concept saliency map; in the denoising process, the weight of the target area is adjusted by using the concept saliency map so as to realize accurate editing. According to the method, under the condition that the global image quality is not affected, the editing precision can be improved, interference to a non-target area is reduced, and meanwhile, reinforcement learning and real-time feedback are combined, so that the model can be adaptively optimized, and an editing result better meeting the user requirement is generated.

The invention relates to the technical field of photoelectric detection and computer vision, and particularly discloses an infrared-based ground-to-air unmanned aerial vehicle detection and tracking method, which comprises the following steps: acquiring an infrared image sequence of a target airspace, dynamically adjusting the integral time and gain parameters of an infrared image by adopting a self-adaptive gain control algorithm based on a scene temperature histogram, and obtaining a target image sequence; generating an infrared image with enhanced thermal radiation characteristics; carrying out multi-level feature fusion detection on the infrared image by adopting a YOLOv5 improved multi-level global feature fusion network, and generating a saliency map containing the thermal features of the unmanned aerial vehicle rotor; according to the invention, by combining the infrared imaging technology, the improved YOLOv5 multi-level global feature fusion network and the improved BotSort algorithm, the detection precision and tracking stability of the low-altitude small unmanned aerial vehicle are effectively improved; the image quality is optimized through a self-adaptive gain control algorithm, the thermal radiation characteristic of the unmanned aerial vehicle is remarkably enhanced, and the failure problem caused by environmental interference in the traditional technology is avoided.

The invention discloses an electronic component packaging defect detection method and system based on image acquisition, and relates to the field of image analysis, and the method comprises the steps: collecting a multi-mode image of a to-be-detected electronic component package; pixel alignment is carried out on the multi-modal image, the multi-modal image after pixel alignment is used as an R channel, a G channel and a B channel to be stacked, and a three-channel pseudo-color image is generated; inputting the three-channel pseudo-color image into a pre-trained convolutional auto-encoder model to obtain a reconstructed three-channel image; analyzing the difference between the three-channel pseudo-color image and the reconstructed three-channel image to obtain an analysis result, and generating a three-channel residual image according to the analysis result; generating a single-channel defect saliency map based on the three-channel residual image; and when a connected region of which the pixel value exceeds a preset pixel threshold value exists in the single-channel defect saliency map, judging that the connected region is a real defect. The detection false alarm rate can be effectively reduced, and the production efficiency is improved.

The invention belongs to the technical field of image processing, and particularly relates to a casting surface defect detection method and system based on image segmentation, and the method comprises the steps: obtaining a gray-scale image and a brightness background image of a vortex casting to be detected, and calculating the illumination deviation of each position in the image; segmenting the grey-scale map of the vortex casting to be detected by using an Otsu method to obtain a binary image; constructing a circular structural element according to the width of the optical interference region and the resolution of the imaging equipment; performing corrosion operation on the binary image based on the structural elements to generate a structural saliency image; obtaining a defect saliency value by using a product of a gray value and an illumination deviation in the structure saliency map; and calculating a segmentation threshold value based on the statistical characteristics of the defect-free sample, and identifying the position of which the defect saliency value is greater than the segmentation threshold value as a defect point. Through physical constraint and illumination feature fusion, the problem that a dark vortex structure and a real defect are difficult to distinguish is effectively solved, and the detection precision is improved.

The invention relates to the technical field of visual detection of coating defects, in particular to a plastic base material surface coating defect detection method based on an LMC process, which comprises the following steps: acquiring a film layer morphology grayscale image of a plastic base material surface coating, constructing a pit direction difference degree of each pixel point, and obtaining a pit edge crack characteristic value of each pixel point; the method comprises the following steps of: obtaining a plurality of pixel points, further obtaining a pit characteristic value of each pixel point, extracting suspected pit characteristic points, obtaining pit dense saliency of each pixel point in combination with a position relationship between each pixel point and the suspected pit characteristic points, and obtaining a film pit saliency map by using a quaternary Fourier saliency detection algorithm. And carrying out image segmentation to obtain a coating pit defect area. According to the invention, the detection precision of coating defects can be improved.

The invention discloses a flat material surface quality detection method and device based on a lightweight significance detection model and a medium, and relates to the field of image recognition, and the detection method comprises the following steps: S1, marking a flat material surface damage data set, and dividing the flat material surface damage data set into a training set and a test set; s2, constructing a lightweight significance detection model; s3, constructing a loss function, adopting a deep supervision strategy to generate predictions for the output feature maps of each stage of the decoder, and calculating the loss; s4, training a detection model by using the training set; and S5, reasoning to obtain a saliency map of the surface defect. According to the method, the lightweight backbone network with the scale adaptive feature extraction module as the core is constructed, the model parameter quantity is only 2.29 M, the reasoning speed of 62 fps is achieved while the lightweight architecture is kept, the key features of the surface defects of the flat material can be efficiently extracted, the requirement of an industrial production line for high-speed detection is met, and the detection efficiency is improved. And a real-time detection target is achieved on the premise that the precision is not sacrificed.

The invention provides an audio and visual fusion method for saliency prediction of a panoramic video, and the method comprises the steps: inputting the panoramic video, obtaining video frames and audio data which form the panoramic video, calculating a tangent image of each group of video frames to obtain K groups of tangent image sequences of the panoramic video, processing the tangent images to obtain visual modal features, and carrying out the saliency prediction of the panoramic video. An audio data sample is obtained after audio data processing, enhanced global time semantic information is further obtained, a final modal feature is obtained by adopting a two-stage fusion strategy, and the final modal feature is decoded by adopting dynamic convolution to obtain a high-precision saliency map. The accuracy and robustness of human attention prediction in the panoramic video are remarkably improved, and technical support is provided for immersive VR experience.

The present application relates to the technical field of image processing, and particularly relates to an image enhancement method and system for titanium metal surface detection. The method comprises the following steps: converting a titanium metal surface image to be processed from an RGB space to an HSI space, obtaining a luminance component, decomposing an initial reflection component into a plurality of high-frequency sub-band coefficients and a low-frequency sub-band coefficient, determining anisotropy degrees of each pixel point, generating an anisotropy saliency map, enhancing the high-frequency sub-band coefficients corresponding to the pixels with anisotropy degrees higher than a preset threshold, otherwise, inhibiting, reconstructing an enhanced reflection component by using the adjusted high-frequency sub-band coefficients and the low-frequency sub-band coefficient without processing, obtaining an enhanced luminance component, converting the enhanced luminance component back to the RGB space, and outputting a final enhanced image. That is, the scheme of the present application can obtain an enhanced image with a clear defect contour, few artifacts and preserved original color information, greatly improving the accuracy and reliability of subsequent automatic defect recognition.

The invention discloses an infrared super-resolution method based on visible light and infrared image fusion. The infrared super-resolution method comprises the following steps: acquiring and registering a high-resolution visible light image and a low-resolution infrared image; preprocessing the visible light image to extract a detail base and a texture saliency map; performing edge keeping smoothing processing on the infrared image by using the visible light image as guidance; the method comprises the following steps of: establishing a multi-factor decision model which depends on infrared thermal intensity, visible light texture significance and cross-modal edge consistency to calculate a pixel-level adaptive fusion weight; and finally, carrying out weighted fusion. Through an intelligent and content-adaptive fusion strategy, artifacts are effectively suppressed, the detail expressive force and the overall quality of the infrared image are remarkably improved while key thermal information is kept in a lossless manner, and the method is low in calculation cost and suitable for a real-time system.

The invention discloses an underwater target detection method and device. The method comprises the following steps: acquiring a first underwater image; performing target detection on the first underwater image by using an underwater target detection model to obtain a first saliency map corresponding to the first underwater image, and determining a first saliency target in the first underwater image according to the first saliency map, the predicted pixel value of each pixel point in the first saliency map is used for representing the probability that the pixel point belongs to a first saliency target, the underwater target detection model internally comprises a multi-stage nested U-shaped network and a saliency map generation module, and each stage of U-shaped network internally comprises a shallow feature extraction module and a deep feature extraction module; the shallow feature extraction module at least comprises a frequency domain channel attention layer. According to the method and the device, the technical problem that the existing target detection algorithm cannot accurately identify the saliency target in the underwater image due to the characteristics of edge blurring and high-frequency information loss of the underwater image is solved.

The invention discloses an automatic driving obstacle recognition method for a complex scenic spot road scene. The method comprises the following steps: acquiring an original image of a scenic spot road; inputting a multi-scale feature extraction network to generate a multi-scale feature map; inputting the image into a multi-branch decoder to respectively obtain an obstacle region segmentation result, an obstacle position bounding box coordinate and an obstacle type classification probability; in combination with surrounding environment information of road geometrical morphology, weather and signal lamp states, feature weighting is carried out on the surrounding environment information by adopting an attention mechanism, and an initial obstacle saliency map is generated; according to a preset obstacle type priority level, detecting and processing an identification result of spatial overlapping in the image to obtain an obstacle identification result after conflict resolution; and performing spatial consistency proofreading on the road path segmentation map and the road path segmentation map, eliminating inter-frame jump based on time sequence smoothing, and finally outputting an identification result. According to the invention, the problem of high-precision and real-time obstacle recognition for complex scenic spot road scenes is effectively solved.