74 results about "Feature compensation" patented technology

The invention provides a remote sensing image semantic segmentation method and system fusing a convolutional neural network and a visual state space model, and the method comprises the steps: firstly extracting the multi-scale semantic features of a remote sensing image based on a lightweight ResNet18 encoder; secondly, a decoder based on a visual state space module is used for modeling a long-distance dependency relationship in an image and recovering spatial resolution; the local feature compensation module is used for enhancing the perception capability of fine-grained semantic information and improving the segmentation precision of a small target area; and finally, the multi-scale attention enhancement module is used for fusing deep and shallow layer features and realizing collaborative optimization of spatial details and semantic information. According to the method, the global semantic features and the local detail information of the remote sensing image can be extracted at the same time, and the method has high segmentation precision and a good remote sensing image semantic segmentation effect.

The invention provides an automatic focus identification and grading method for a medical image, and the method comprises the steps: carrying out the standardization of an obtained multi-modal original image based on anatomical constraint, and obtaining a standardized image; generating semantic enhancement features through a cross-modal feature compensation network based on the standardized image and associated radiological text description; performing dynamic feature adaptation processing on the semantic enhancement feature to generate a modal adaptive feature; performing context reasoning through a multi-scale feature interaction algorithm based on the modal adaptive features to generate context reasoning features; and lesion identification decoding processing is carried out on the context inference feature map, a lesion segmentation mask is generated, and the lesion segmentation mask is used for extracting lesion area feature parameters to carry out lesion classification. By adopting the method, the adaptability to the missing mode can be enhanced, and the focus identification and grading precision can be improved.

The invention provides a Yangtze river export deep and far sea environment data association analysis method based on multi-modal fusion, belongs to the technical field of far sea environment analysis, and achieves accurate data registration by establishing a multi-source data space-time standardization model and adopting an adaptive space-time Kriging interpolation algorithm. A sparse coding ocean signal separation algorithm is utilized to construct an over-complete dictionary separation mixed signal to extract pure features, a feature extraction network is constructed based on an attention mechanism to automatically learn deep feature representation of physicochemical biological parameters, and a matrix rank loss detection algorithm is adopted to automatically supplement feature compensation vectors to ensure feature integrity. A cross-scale attention mechanism is constructed based on wavelet transform to fuse multi-scale features, a Shapley value interpretability evaluation system is established to quantify correlation feature importance, and the technical problem that correlation feature extraction is inaccurate in the multi-source heterogeneous marine environment data fusion process is solved.

The invention belongs to the technical field of building construction safety control, and particularly discloses a building construction site dangerous behavior recognition method and system based on machine learning, and the method comprises the steps: obtaining a single-view continuous video stream of a construction site, segmenting the single-view continuous video stream into a time sequence video frame sequence, and extracting an initial spatial feature sequence through a spatial feature encoder; a pre-trained virtual visual angle projection and feature compensation module encodes the visual angle implicit vector and maps the visual angle implicit vector to a visual angle invariant feature space, and a sequential context is combined to compensate occlusion missing features to generate an enhanced feature sequence; a time sequence memory alignment module captures long-term and short-term time sequence dependence and outputs time sequence consistency characteristics; and finally, outputting a current dangerous behavior classification result through the classification prediction head, and outputting a future dangerous intention probability curve through the time sequence prediction head. The method does not need an additional camera, can accurately cope with a shielding scene, predicts the danger in advance, and improves the construction safety monitoring efficiency and reliability.

The invention provides a multi-modal data fusion diagnosis assistance method and device, a storage medium and equipment, and the method comprises the steps: obtaining multi-modal data sent by a data collection module, the multi-modal data comprising at least two of 2D fundus color photo data, 3D OCT body data, OCTA blood flow parameters and wide-area fundus spliced image parameters; inputting the multi-modal data into a trained layered feature extraction module to obtain a multi-modal data feature vector; inputting the multi-modal data feature vector into a trained feature compensation-fusion module to obtain a complete multi-modal data fusion feature; and inputting the complete multi-modal data fusion features into a trained dynamic expert hybrid system to obtain structured diagnosis and treatment data and a text diagnosis auxiliary report, and sending the structured diagnosis and treatment data and the text diagnosis auxiliary report to a data display module.

The invention belongs to the field of deep learning technology and remote sensing image segmentation, and particularly relates to a remote sensing image segmentation method based on Transsubnet edge information enhancement and multi-dimensional feature perception, and the method comprises the steps: S1, preparing a data set; s2, constructing remote sensing picture text description; s3, constructing and training a remote sensing image segmentation model; and S4, storing and testing the model. The invention designs a multi-modal feature extraction method based on parallelism of a sampling branch and a text feature extraction branch under edge feature compensation. The residual error mixing axial attention module is used for forming a transformer structure; and a text-picture multi-dimensional feature fusion enhancement module and a decoder part are embedded. According to the method, the ground feature identification capability can be improved through a text and picture multi-modal feature enhancement strategy, the segmentation boundary and small target object feature information is enhanced, more fine-grained features are reserved, the cross-regional long-distance dependency relationship is better captured, the common gradient disappearance problem in a deep network is relieved, and the method is suitable for large-scale popularization and application. And meanwhile, the small sample data set segmentation effect is improved.

The invention provides an underwater robot propeller type thruster cross-domain intelligent fault diagnosis method, and belongs to the field of underwater robot fault diagnosis. Energy characteristics and environmental parameters are introduced through a dynamic adversarial generative network, frequency domain physical constraints are constructed, pseudo fault data adaptive to a target domain are generated, and authenticity and rationality of generated signals are guaranteed by using time domain and frequency domain double discriminators. The environment residual error adaptation network fuses the residual error between the actual measurement characteristic of the target domain and the pseudo fault characteristic with the environment factor after the environment parameter coding by constructing a residual error dynamic compensation mechanism, constructs the dynamic mapping relation between the environment parameter and the characteristic residual error, adaptively generates the characteristic compensation amount, and finally performs the self-adaption on the characteristic compensation amount. Double-layer domain adaptation of global distribution alignment and local residual compensation is achieved, and the generalization ability and diagnosis accuracy of underwater robot propeller fault diagnosis in different water area environments can be effectively improved.

The invention discloses a radio frequency fingerprint hiding method based on feature compensation. The method comprises the following steps: at a transmitter end, carrying out carrier frequency offset compensation, I / Q imbalance compensation and PA nonlinear distortion compensation on a to-be-transmitted signal in sequence by using a preset joint compensation model to obtain a transmitted signal after radio frequency fingerprint hiding; the joint compensation model is constructed by the following steps: carrying out carrier frequency offset estimation and compensation, I / Q imbalance estimation and compensation and PA nonlinear distortion estimation on a received signal of a receiver in sequence, and constructing the joint compensation model according to a carrier frequency offset estimation result, an I / Q imbalance estimation result and a PA nonlinear distortion estimation result. According to the method, a scientific and reasonable parameter estimation and compensation sequence is designed, an illegal receiver is effectively prevented from extracting radio frequency fingerprints from received signals, the classification precision of the RFFI technology can be effectively reduced, and the privacy protection and recognition resistance of a transmitter are remarkably enhanced.

The invention provides a multi-scene image degradation integrated recovery method based on a visual language model, which realizes high-quality image reconstruction under various weather degradation conditions, and comprises the following steps: inputting a given degraded image into an encoder for encoding to obtain encoding features; querying the pre-trained visual language model through a cross-modal prompt generator to generate a multi-scale degradation perception cross-modal prompt; the pre-trained visual language model takes a problem and a degraded image as input; the degeneration perception cross-modal prompt is input into the restoration trunk through a guide attention alignment module to be aligned and fused with the coding features, and final output of the guide attention alignment module is obtained; refining and fusing the degraded image and the final output of the attention alignment guiding module through a double feature compensation module to obtain the final total output; and carrying out decoding processing and image reconstruction on the final total output through a decoder to obtain a high-quality output image.

The invention relates to the technical field of information engineering knowledge graph construction, in particular to a supervision knowledge graph dynamic modeling method based on multi-source heterogeneous data fusion, and the method comprises the steps: designing a four-dimensional knowledge graph modeling framework, and combining the entity alignment of a graph neural network, the relation extraction of remote supervision and the attribute processing of weight fusion. A dynamic evolution and cross-source consistent knowledge network is constructed, virtual features are generated through a virtual feature compensation mechanism (historical event retrieval and multi-dimensional feature weighted fusion) for a multi-modal data missing scene and mapped to a graph, data are effectively filled, the problem of incomplete knowledge coverage in a complex engineering scene is solved, a compliance entropy evaluation model is introduced, and the knowledge coverage is evaluated. The compliance state of the knowledge graph is quantitatively supervised from the three dimensions of rule coverage rate, data consistency and tracing integrity, and the hysteresis and subjectivity limitation of traditional manual compliance check is broken through.

The invention provides a double-branch spectrum quantitative analysis method Mark-GLNet based on Pearson feature marking, and global dependency modeling and local feature extraction are collaboratively optimized. The Mark mark layer quantifies wavelength importance by using a Pearson's correlation coefficient, eliminates redundant wavelengths, and inhibits low correlation noise. Refining feature input GLNet: global branches utilize sequence continuity and cross-space correlation, dynamic weights highlight keys, and redundancy is suppressed; and local branches compensate channel correlation through multi-scale convolution, and reinforce details. Double branches are cooperated through feature compensation, redundant interference is reduced, and prediction precision is improved. Through verification of four open-source near-infrared data sets, the model is simple in structure, low in parameter quantity and calculation complexity and excellent in generalization ability, and has high practicability and expandability.

The invention discloses a three-dimensional occupancy perception method and system suitable for multiple tasks, and belongs to the field of robot perception, and the method comprises the steps: carrying out the manual labeling of point cloud data and image data, obtaining a point-by-point label and a pixel-by-pixel label, carrying out the data verification, converting the point cloud data with the labels into a 3D semantic occupancy label of a scene through 3D semantic reconstruction, and carrying out the recognition of the 3D semantic occupancy label. Forming sample data; training a three-dimensional occupancy perception model by using the sample data and performing three-dimensional occupancy perception prediction; sequentially performing feature enhancement extraction, feature dimension conversion, feature time sequence fusion and feature compensation correction on the image data based on a feature extraction module to obtain multi-scale voxel features; the feature information of the multi-scale voxel features is independently adjusted through a semantic segmentation head and a target detection head, then a 3D semantic occupancy result and an instance detection result are output, the two detection results are converted into an instance segmentation result and a target tracking result through a post-processing module, and low-cost and high-efficiency multi-task three-dimensional occupancy perception is achieved.

The invention relates to an image denoising method and device, equipment and a medium, and the method comprises the steps: obtaining a training data set which comprises a plurality of image sample pairs, and each image sample pair comprises a noise sample image and a noise-free supervision image corresponding to the noise sample image; dividing the image sample pair according to a preset size to determine image blocks corresponding to the noisy sample image and the noiseless supervision image; the method comprises the following steps of: embedding a plurality of improved Transform Blocks in each coding layer and each decoding layer of a preset U-Net trunk structure, and introducing a mixed feature compensation module at an input end and an output end in the U-Net trunk structure to construct an image denoising model; and inputting a to-be-denoised image into the image denoising model trained to the convergence state to determine a clean image corresponding to the to-be-denoised image so as to complete image denoising. According to the invention, the generalization ability and stability of image denoising in a complex real scene can be greatly improved.

The application discloses a gas extraction water ring vacuum pump fouling degree online detection system and method, relates to the technical field of vacuum pump fouling degree detection, and comprises a sound signal acquisition module, an excitation control module, a parameter sensing module, a signal processing module, a feature compensation module, a fouling inversion module and a model calibration module; the sound signal acquisition module is arranged on the outer wall of the pump body and is used for acquiring structural vibration signals generated in the operation process and outputting original vibration data; the excitation control module is connected with a piezoelectric excitation device and receives the operation state feedback of the sound signal acquisition module, is used for triggering the piezoelectric excitation device to apply high-frequency pulse signals to the pump shell when the suction negative pressure fluctuation is stable and continuously meets the steady-state condition, and synchronously starts the sound signal acquisition module to collect response signals; and the parameter sensing module is used for monitoring the cooling water inlet / outlet temperature and motor current waveform and outputting environmental and operation parameters.

The invention relates to the technical field of electromagnetic valve manufacturing and quality detection, and discloses an online detection system for the pressing force of a valve element of an electromagnetic valve, and the system comprises a composite magnetic field excitation module, a multi-physical field signal acquisition module, a multi-dimensional feature resolving module and a force value estimation output module. A composite magnetic field is applied to a valve element, and an induced voltage signal, a Hall signal and a temperature signal are synchronously collected; then, three characteristic quantities of harmonic amplitude, instantaneous magnetic field intensity and instantaneous temperature are calculated; and finally, inputting the three characteristic quantities into a preset multi-dimensional mapping model, and calculating to obtain a pressing force estimated value. The invention aims to solve the problem of low measurement accuracy caused by the fact that the existing single electromagnetic parameter detection is easily interfered by temperature and excitation field fluctuation. And through multi-physical field signal fusion and multi-dimensional feature compensation, the influence of disturbance factors is effectively eliminated, and the accuracy, stability and anti-interference capability of online detection are remarkably improved.

This invention provides a multimodal fusion-based method for correlation analysis of deep-sea environmental data outside the Yangtze River Estuary, belonging to the field of deep-sea environmental analysis technology. This invention achieves accurate data registration by establishing a spatiotemporal standardization model for multi-source data and employing an adaptive spatiotemporal kriging interpolation algorithm. It utilizes a sparse coding marine signal separation algorithm to construct an overcomplete dictionary to separate mixed signals and extract pure features. Based on an attention mechanism, a feature extraction network is constructed to automatically learn deep feature representations of physical, chemical, and biological parameters. A matrix rank deficiency detection algorithm is used to automatically supplement feature compensation vectors to ensure feature integrity. A cross-scale attention mechanism based on wavelet transform is constructed to fuse multi-scale features. A Shapley value interpretability evaluation system is established to quantify the importance of correlated features. This invention solves the technical problem of inaccurate correlated feature extraction during the fusion of multi-source heterogeneous marine environmental data.

The application discloses a hydrological remote sensing image target recognition method based on a deep semantic model, relates to the technical field of remote sensing image processing, and comprises the following steps: obtaining a primary semantic code of a hydrological remote sensing image through a deep semantic network, calculating a semantic attribution probability distribution according to the primary semantic code, generating an initial attention guide signal, dynamically adjusting the weight of a specific perception path in the network, realizing first feature re-extraction to obtain refined features, calculating a feature compensation vector based on the confidence deviation generated by matching the features with a standard feature library, correcting the refined features to obtain enhanced features, and inputting the enhanced features into the network again for analysis to obtain a final recognition result. Through the attention adjustment of the semantic guide and the feature compensation based on the confidence feedback, the method effectively improves the recognition accuracy of complex hydrological ground object targets and the model robustness.

The invention discloses a neural network-based vehicle axle intelligent identification optimization method and system, and belongs to the technical field of neural networks. Image acquisition is carried out in an ETC lane, a vehicle passing image sample set is constructed, scene identification parameters are established, region division is carried out on a vehicle image, and a shielding correction weight is calculated; multi-level features are extracted based on a convolutional neural network, feature compensation is carried out on the occlusion area through a neighborhood feature compensation module, and a complete axle candidate feature map is formed; inputting the candidate feature map into a lightweight neural network model, combining multi-scale fusion of shallow and deep features, dynamically adjusting the weight of a feature layer according to a shielding correction weight, and outputting an axle recognition result; and finally, according to a preset axle configuration rule base, matching a vehicle type identifier, and establishing a sample-parameter-classification mapping data set. According to the invention, axle features can be stably identified under complex illumination and shielding conditions, and the accuracy and robustness of ETC lane vehicle identification are improved.

The application discloses an end-to-end pest detection method with feature compensation and local information enhancement, and belongs to the field of deep learning. First, a Transform network framework based on feature representation compensation and cross-shaped window attention mechanism is established to extract multi-scale features of an image. Second, regional grid self-attention is used to replace even-numbered multi-head self-attention modules in a Transform model encoder to obtain feature information of the image. Finally, the first K encoder features in the last layer are selected to initialize the position query of a Transform decoder with contrastive denoising, and the content query remains a learnable parameter. Compared with a convolution-based method, the technical scheme of the application can achieve a higher average precision mean in a pest detection task.

The application discloses a kind of vehicle axle intelligent identification optimization method and system based on neural network, belong to neural network technical field;Image acquisition is carried out in ETC lane, vehicle passing image sample set is constructed, and scene identification parameter is established, vehicle image is divided into regions, and the weight of correction of obstruction is calculated;Based on convolutional neural network, multi-level features are extracted, features are compensated in the obstruction area by neighborhood feature compensation module, and complete axle candidate feature map is formed;Then the candidate feature map is input into the lightweight neural network model, the multiscale fusion of shallow and deep features is combined, and the feature layer weight is dynamically adjusted according to the obstruction correction weight, and the axle recognition result is output;Finally, according to the preset axle configuration rule base, the vehicle model identification is matched, and the sample-parameter-classification mapping dataset is established.The axle feature can be still stably identified under complex illumination and obstruction conditions, and the accuracy and robustness of ETC lane vehicle identification are improved.

The invention provides a bandwidth adaptive feature compression method and device in cooperative computing, electronic equipment and a medium, and the method comprises the steps: carrying out the feature extraction of a to-be-processed image, and obtaining an intermediate feature corresponding to the to-be-processed image; inputting a current transmission bandwidth and the intermediate feature into a feature compression model to enable the feature compression model to compress the intermediate feature based on the current transmission bandwidth, and outputting the compressed intermediate feature; and sending the compressed intermediate features to a cloud device, so that the cloud device inputs the compressed intermediate features to a category feature compensation model for compensation, and performs category sensing processing based on the compensated intermediate features. According to the scheme, the stability of the data transmission efficiency can be improved, so that the transmission delay can be reduced.

This invention discloses a network security traffic feature compensation and identification method based on graph neural networks. The method includes: acquiring network traffic data to be tested and extracting key attributes; constructing a dynamic heterogeneous graph based on the key attributes; performing feature compensation processing on the dynamic heterogeneous graph, mapping nodes to a three-dimensional coordinate space and compensating for missing feature dimensions; inputting the compensated dynamic heterogeneous graph into a graph neural network model to generate node state information; comparing the node state information with a sparse baseline to output an anomaly score; and generating a virtual network function management strategy based on the anomaly score and deploying it to the target node. This invention achieves traffic feature compensation through three-dimensional coordinate mapping and neighbor feature aggregation, combines graph neural networks and a sparse baseline for anomaly detection, and forms a closed-loop response through inference acceleration and automatic VNF deployment, thereby improving the accuracy and real-time performance of network security detection.

The embodiment of the invention provides an audio processing method and system, computing equipment and a storage medium. The audio processing method comprises the following steps: acquiring various types of initial audio and reference audio features; based on the initial audio feature of the initial audio, determining an audio feature relationship between the initial audios of various types; constructing a feature compensation model based on a reference audio feature and an audio feature relationship by taking an auditory frequency response relationship as a constraint; obtaining a feature compensation value of the initial audio based on a feature compensation model; and adjusting the initial audio features based on the feature compensation value to obtain multiple types of target audios. According to the technical scheme of the invention, the loudness standards of various types of audios can meet the requirements, the spatial sense and the layering sense of sound are prevented from being damaged, and the unreasonable covering phenomenon of various types of audios in a sound mixing scene is avoided, so that the overall hearing sense and the spatial layering sense of various types of audios after being integrated into a final product are improved, and the sound quality is improved. The method can be widely applied to the fields of digital culture product manufacturing software and the like.

The application is suitable for the technical field of target detection, and provides a dead chicken detection method based on multi-scale reconstruction layer aggregation and feature compensation progressive fusion, including: acquiring a target detection image, the target detection image including dead chickens and live chickens, inputting the target detection image into a preset YOLO model, outputting a dead chicken detection result, the preset YOLO model including a backbone feature extraction network, a feature fusion module and a detection output head, the backbone feature extraction network being a preset multi-scale reconstruction efficient layer aggregation MSR-ELAN network, and the feature fusion module being a preset feature compensation progressive fusion pyramid FCPFN network. Thus, by segmenting and recombining the extracted features, using a multi-stage feature interaction and integration strategy, the overall perception ability of the target is optimized, thereby effectively improving the detection performance of the dead chicken target.

The present invention relates to the technical field of person re-identification. Disclosed are a cross-modal person re-identification method and system based on wavelet transform. The method comprises: constructing a two-stream ResNet-50 backbone network; by means of an ICB module, fusing shallow features of the two-stream ResNet-50 backbone network, so as to obtain fused features; by means of a WEB branch module, performing feature extraction on the fused features, so as to obtain modality-shared features of high-frequency and low-frequency portions; and by means of a dual-branch center-guided loss, optimizing the modality-shared features output by the WEB branch module and the fused features output by the ICB module, so as to obtain a final identification result. In the present invention, an information compensation module is combined with wavelet transform to aggregate shallow network features in different phases, and thus valuable information lost in network feature extraction is compensated for, thereby improving the quality of a final feature vector; and a wavelet enhancement module is provided, and a dual-branch center-guided loss is used to guide a network to mine modality-invariant information in wavelet subgraphs, thereby improving the model performance.

The invention relates to the technical field of digital image processing, and discloses a multi-modal image fusion method for spatial perception multi-scale feature learning, which comprises the following steps of: processing a registered multi-modal image pair by adopting a hierarchical expansion convolution structure so as to obtain multi-scale features with different receptive field ranges in parallel; processing the multi-scale features by using a two-layer four-way recurrent neural network, and introducing an attention mechanism to integrate local and global feature information of the image to obtain spatial direction sensing features; by introducing an attention coordination mechanism and jump connection, shallow details and deep semantic features are fused, and a fused image is reconstructed and generated through convolution operation. According to the method, through parallel multi-scale feature extraction, global spatial information perception and cross-level feature compensation, the performance of the fusion image in the aspects of information fidelity, structural integrity and detail definition is effectively improved, and a fusion result with a prominent target and rich textures can be generated.

The invention discloses a neural signal decoding method and system based on deviation prediction and feature compensation, and the method comprises the steps: obtaining a visual image and a neural signal, extracting a reference visual feature of the visual image based on a visual encoder, and extracting an initial neural signal feature of the neural signal based on a to-be-trained neural signal encoder; based on a deviation prediction module, deviation information is generated through the initial neural signal feature, and the deviation information is used for representing the difference between the initial neural signal feature and an ideal feature capable of being aligned with the reference visual feature; compensating the initial neural signal feature through the deviation information to generate a compensated neural signal feature; optimizing a neural signal encoder and a deviation prediction module; and obtaining a target neural signal, and decoding the target neural signal based on the optimized neural signal encoder and the deviation prediction module. According to the invention, the effect of highly aligning the decoded neural signal and visual signal can be realized.

The invention provides a construction road crack detection method and system, a readable storage medium and a computer. The method comprises the following steps: constructing a training data set according to each construction road image; constructing a feature extraction model, and introducing a wavelet convolution algorithm into the feature extraction model to generate a feature extraction optimization model; replacing a sampling module of the feature extraction optimization model with a dynamic sampling algorithm, and performing weighted fusion on the feature compensation algorithm and the backbone network features of the feature extraction optimization model to obtain a feature processing model; and constructing a weighted loss function, performing model optimization on the feature processing model by using the weighted loss function and the training data set to generate a crack detection model, and completing crack detection on the to-be-processed construction road image according to the crack detection model. According to the method, the weighted loss function fusing supervised learning and unsupervised learning is constructed, the potential of label-free data is fully mined, and the dependence on labeled data is greatly reduced while the precision is ensured.

This invention belongs to the field of network security technology and discloses a method for predicting industrial control system (ICS) time series based on multi-scale time-frequency joint sensing. It proposes an end-to-end learnable adaptive time-frequency feature extractor to replace the low-pass filtering method relying on fixed parameters in existing technologies. A data-driven dynamic multi-scale partitioning mechanism is proposed, which adaptively adjusts the scale partitioning method according to the actual distribution characteristics and fluctuation patterns of ICS time series data. A scale-adaptive residual projector is designed, and through multi-dimensional spatial projection and residual feature compensation, the problem of multi-scale feature dimension mismatch is solved, establishing a feature cascade enhancement relationship between coarse and fine scales, thus strengthening the effective fusion of multi-scale features and the ability to model long-term time series dependencies. A time-frequency cross-modal collaborative calibration mechanism is proposed, breaking through the limitation of traditional channel attention mechanisms that rely only on a single time dimension. It achieves joint learning of dual-domain information by fusing time-domain statistics and frequency-domain context.

The application discloses a rolling plate surface defect detection method based on improved YOLOv5, which contains three core improvements: (1) the backbone module introduces residual synchronized convolutional block (RSCB), which synchronously extracts complex defect features; (2) the neck module designs convolution-multilayer perceptron operation (Conv-MLP), which strengthens the interaction of global and local features of an image and improves the large-span defect detection capability; and (3) the fusion strategy is optimized, and details lost by deep layer features are compensated by shallow layer features. Through the multi-scale feature processing mechanism, the application improves the micro defect recognition precision and realizes higher-precision surface defect positioning and classification in an industrial scene.