A high-precision visual inspection system and method based on deep learning

Abstract

The application provides a high-precision visual detection system and method based on deep learning, and relates to the technical field of visual detection. The application comprises a perception decision layer, which transmits collected data to a feature engineering layer. The feature engineering layer transmits analyzed and processed data to a detection optimization layer. The detection optimization layer analyzes the data processed by the feature engineering layer, transmits the prediction results to the perception decision layer, and triggers an enhanced mode to drive the image acquisition device to perform high-frequency interception. The feature engineering layer generates high-quality images using multi-frame fusion technology and uses attention mechanisms to achieve noise suppression and target enhancement at the feature level. The detection optimization layer performs accurate identification based on the purified feature map and continuously monitors performance, digs out difficult cases, and optimizes parameters online to form a self-adaptive closed loop and improve the reliability of target detection in bad weather.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and in particular to a high-precision visual inspection system and method based on deep learning. Background Technology

[0002] High-precision visual inspection systems combine industrial cameras with intelligent algorithms to achieve rapid and automated inspection of product appearance, dimensions, and assembly structures. They can accurately identify micron-level defects, such as scratches, foreign objects, or dimensional deviations, on high-speed production lines. At the same time, they can complete complex tasks such as character recognition and positioning guidance. This system has largely replaced traditional manual visual inspection, effectively improving inspection efficiency and consistency, avoiding subjective misjudgments, and has become a core technology for quality control in fields such as electronics manufacturing, automotive industry, and pharmaceutical packaging. It significantly reduces defect rates and ensures process standardization.

[0003] Existing high-precision visual inspection systems face severe challenges under extreme weather conditions. Heavy rain, dense fog, and other conditions cause severe degradation in image quality, resulting in blurred, distorted, and reduced contrast images. This leads to severe contamination of feature maps extracted by deep neural networks, with a large amount of invalid features generated by weather noise. When making predictions based on these contaminated features, the system experiences a large number of false positives and false negatives, such as misidentifying rain lines or fog as targets, or failing to identify real objects that are obscured. Existing image enhancement-based preprocessing methods have obvious limitations. Their general design is not optimized for specific detection tasks. While removing noise, they may introduce new image artifacts and destroy useful features. Furthermore, their high computational complexity makes it difficult to meet the timeliness requirements of real-time monitoring. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a high-precision visual inspection system and method based on deep learning.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a high-precision visual detection system and method based on deep learning, comprising a perception decision layer, wherein the perception decision layer transmits the collected data to a feature engineering layer, the feature engineering layer analyzes and processes the obtained data and transmits it to a detection optimization layer, and the detection optimization layer analyzes the data processed by the feature engineering layer and transmits the prediction result to the perception decision layer.

[0006] The perception and decision-making layer includes a real-time weather perception module that performs real-time analysis of video streams and an intelligent strategy switching module with an internal decision threshold.

[0007] The feature engineering layer includes a high-frequency capture and data supplementation module for switching the image acquisition device from a normal frame rate to a high-frequency capture mode, a multi-frame information fusion module that utilizes temporal redundancy to highlight persistent targets, and an attention-driven feature cleanup module that receives the initial feature map extracted by the backbone network and performs operations.

[0008] The detection optimization layer includes a high-performance robust target detection module that takes the cleaned feature map as input and a system feedback module that takes the confidence level of the detection results and the consistency with historical data as feedback signals.

[0009] In a preferred embodiment, the perception decision layer includes a real-time weather perception module, which analyzes the video stream and outputs weather classification results and interference factors. The intelligent strategy switching module receives the weather classification results and interference factors, compares them with a preset decision threshold, and sends a working mode switching command to the image acquisition device and the feature engineering layer based on the comparison results.

[0010] In a preferred embodiment, the feature engineering layer includes a high-frequency interception and data supplementation module, which receives a working mode switching instruction and outputs a continuous image sequence. The multi-frame information fusion module receives the continuous image sequence and performs frame alignment and fusion processing to output a fused image. The feature purification module receives the fused image and performs feature extraction and purification processing to output the processed feature map to the detection optimization layer.

[0011] In a preferred embodiment, the detection optimization layer includes a robust target detection module. The robust target detection module receives the feature map output by the feature engineering layer and performs target detection, outputting the detection result. The system feedback module receives the detection result and generates a feedback signal, which is then transmitted to the perception decision layer and the external system.

[0012] In a preferred embodiment, the real-time weather perception module includes a frame buffer for storing continuous frames, an image preprocessing unit for standardizing the buffered frames, and a lightweight deep neural network engine for extracting features from the preprocessed image and outputting weather classification results and interference factors.

[0013] The intelligent strategy switching module includes an input parser that parses data from the weather sensing module, a threshold register that stores decision thresholds, a decision logic comparator that compares interference factors with thresholds, a mode state machine that controls the system's operating state, and an instruction encoder that generates control instructions based on the state machine's state.

[0014] In a preferred embodiment, the high-frequency interception and data supplementation module includes an instruction decoder for parsing mode switching instructions, a frame rate controller for sending frame rate control commands to the image acquisition device, and a high-speed image buffer for storing high frame rate image sequences.

[0015] The multi-frame information fusion module includes a frame sequence selector that sequentially extracts a fixed-length continuous frame sequence from a high-speed image buffer; a motion estimation engine that calculates inter-frame motion vectors; a motion compensator for frame alignment processing; a weight calculation unit that calculates pixel fusion weights; and a pixel-level fusion core that generates a fused image.

[0016] The attention-driven feature cleansing module includes a backbone network that extracts initial feature maps from the fused image, an interference suppression branch that generates a weather interference mask, a target enhancement branch that generates a key target attention map, a feature cleansing operator that performs feature cleansing calculations, and a cleaned feature map cache that stores the output feature maps.

[0017] In a preferred embodiment, the robust object detection module includes a feature input interface for receiving and caching feature maps, a region proposal network for generating candidate regions that may contain objects on the feature maps, a classification and regression head for object classification and bounding box regression, and a post-processing engine for filtering and optimizing detection results.

[0018] The system feedback module includes a performance monitor that analyzes detection results and calculates performance indicators; a sample information mining tool that identifies and collects sample information that is difficult to analyze and process; an online learning engine that performs online adjustment of model parameters; and a sample database that stores sample information and annotation information that are difficult to analyze and process.

[0019] A high-precision visual detection method based on deep learning includes the following steps:

[0020] S1: Environmental Perception and Intelligent Decision-Making

[0021] The perception and decision layer performs real-time analysis of the input video stream. Through the real-time weather perception module, it caches continuous frames, performs image preprocessing and lightweight deep neural network analysis, completes multi-frame temporal analysis, and outputs weather classification results and quantified interference factors. The intelligent strategy switching module parses the results, compares the interference factors with the preset decision threshold, and generates a working mode switching command based on this, which is sent to the image acquisition device and feature engineering layer to realize the dynamic switching of the system's working mode.

[0022] S2: High-frequency image acquisition and data buffering:

[0023] The high-frequency capture and data supplementation module in the feature engineering layer starts after receiving the switching command. It parses the content through the command decoder, and the frame rate controller controls the image acquisition device to switch to the high-frequency capture mode. It also uses the ring buffer structure of the high-speed image buffer to temporarily store the high frame rate image sequence to solve the problem of matching the data production and consumption rate, and to provide a stable and continuous image sequence for subsequent processing.

[0024] S3: Multi-frame fusion and feature cleanup:

[0025] The multi-frame information fusion module receives a continuous image sequence, performs frame sequence selection, motion estimation and compensation, weight calculation and pixel-level fusion to generate a high signal-to-noise ratio fused image, suppresses instantaneous interference and highlights the real target. The feature purification module performs deep feature extraction on the fused image, obtains an initial feature map through the backbone network, and generates a weather mask and a target attention map by parallel interference suppression and target enhancement branches, respectively. Finally, the feature purification operator realizes noise suppression and target enhancement and outputs a purified feature map.

[0026] As a preferred implementation, S4: Target detection and result output:

[0027] The robust target detection module in the detection optimization layer receives the cleaned feature map, which is then cached by the feature interface. The region proposal network generates candidate regions, and the classification and regression heads process each candidate region in parallel to complete target classification and location regression. The post-processing engine refines the results by removing redundant boxes through non-maximum suppression and confidence filtering, and finally outputs high-precision detection boxes and category information.

[0028] S5: Performance Monitoring and Closed-Loop Feedback

[0029] The system feedback module continuously monitors the detection effect, the performance monitor analyzes the confidence level and inter-frame consistency to calculate key indicators, the sample information mining machine for difficult-to-analyze samples automatically identifies and collects samples and builds a database, the online learning engine starts fine-tuning when conditions are triggered to optimize model performance, and the feedback signal is sent back to the perception decision layer to realize closed-loop optimization and adaptive adjustment of system parameters, ensuring continuous evolution capability.

[0030] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0031] This invention intelligently identifies weather interference and triggers an enhancement mode through a perception and decision-making layer, driving the image acquisition device to perform high-frequency interception to obtain rich temporal data. The feature engineering layer then uses multi-frame fusion technology to generate high signal-to-noise ratio images and employs an attention mechanism to achieve noise suppression and target enhancement at the feature level, effectively overcoming the lack of information in a single frame and feature distortion. The detection optimization layer performs accurate target recognition based on the purified feature map, while continuously monitoring performance, mining difficult cases, and optimizing model parameters online through a feedback mechanism to form an adaptive closed loop. This significantly improves the reliability of target detection in security monitoring under adverse weather conditions and effectively suppresses false detections and missed detections. Attached Figure Description

[0032] Figure 1 This invention provides a block diagram of a high-precision visual inspection system based on deep learning.

[0033] Figure 2 The flowchart illustrates the operation of a high-precision visual inspection method based on deep learning, as provided in this invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Example 1

[0036] like Figure 1 - Figure 2 As shown, the present invention provides a technical solution: a high-precision visual detection system and method based on deep learning, the system comprising a perception decision layer, a feature engineering layer and a detection optimization layer connected in sequence;

[0037] The perception and decision layer includes a real-time weather perception module and an intelligent strategy switching module. The real-time weather perception module receives and stores continuous frames of the video stream through an internal frame buffer, performs standardization processing on these buffered frames using an image preprocessing unit, and then extracts features from the processed image through a lightweight deep neural network engine, finally outputting weather classification results and quantified interference factors. The intelligent strategy switching module receives and parses the weather classification results and interference factors from the real-time weather perception module through its input parser, compares the interference factors with the decision thresholds stored in the threshold register through a decision logic comparator, controls the system's working state based on the comparison results through a pattern state machine, generates control commands based on the determined working state through an instruction encoder, and sends these commands to the image acquisition device and the feature engineering layer.

[0038] The feature engineering layer comprises three main modules: a high-frequency interception and data supplementation module, a multi-frame information fusion module, and an attention-driven feature purification module. The high-frequency interception and data supplementation module first receives control commands from the perception decision layer via an instruction decoder, then sends frame rate control commands to the image acquisition device via a frame rate controller, and simultaneously stores high-frame-rate image sequences in a high-speed image buffer. The multi-frame information fusion module extracts continuous frame sequences from the high-speed image buffer via a frame sequence selector, then calculates inter-frame motion vectors using a motion estimation engine, performs frame alignment processing via a motion compensator, and finally calculates pixel fusion weights using a weight calculation unit. Finally, a fused image is generated by a pixel-level fusion core. In the attention-driven feature purification module, the backbone network first extracts an initial feature map from the fused image, then generates a weather interference mask via an interference suppression branch and a key target attention map via a target enhancement branch. The feature purification operator then performs feature purification calculations and stores the processed feature map in a purified feature map cache, finally outputting it to the detection optimization layer.

[0039] The detection optimization layer comprises two main modules: a robust target detection module and a system feedback module. The robust target detection module first receives the feature map output from the feature engineering layer through the feature input interface. Then, the region proposal network generates candidate regions. Next, the classification and regression head performs target classification and bounding box regression on the candidate regions. The post-processing engine filters and optimizes the results and outputs the final detection result. The system feedback module receives the above detection results through the performance monitor, analyzes and calculates its confidence and consistency with historical data, and generates feedback signals accordingly. The difficult-to-analyze sample information miner identifies and collects difficult-to-analyze sample information. The online learning engine performs real-time adjustment of model parameters. Finally, the difficult-to-analyze sample information and related annotation information are uniformly stored in the sample database.

[0040] The system feedback module transmits the generated feedback signal to the intelligent strategy switching module of the perception and decision layer. The intelligent strategy switching module adjusts the internal decision threshold according to the feedback signal, thereby completing the dynamic optimization of the system's working state and forming a self-optimization closed loop.

[0041] This embodiment includes a perception decision layer, a feature engineering layer, and a detection optimization layer connected in sequence. The perception decision layer first processes the input video stream through a real-time weather perception module: it receives and stores continuous video frames using a frame buffer, standardizes them using an image preprocessing unit, extracts image features using a lightweight deep neural network, and outputs weather classification results and corresponding interference factors. Subsequently, the intelligent strategy switching module parses the above output, compares the interference factors with preset thresholds using a decision logic comparator, dynamically adjusts the system's working state using a mode state machine, and finally generates control commands using an instruction encoder, which are sent to the image acquisition device and the feature engineering layer respectively. After receiving the commands, the feature engineering layer starts the corresponding working mode: the high-frequency interception and data supplementation module parses the commands, the frame rate controller adjusts the image acquisition frame rate, and simultaneously stores high frame rate image sequences through a high-speed image buffer. The multi-frame information fusion module... Continuous frames are selected from the sequence. First, frame alignment is achieved through motion estimation and compensation. Then, high-quality fused images are generated by combining weight calculation and pixel-level fusion technology. The attention-driven feature purification module extracts the initial feature map from the fused image and generates weather interference masks and key target attention maps in parallel through interference suppression and target enhancement branches. Finally, the feature purification operator completes the feature purification process and outputs the optimized feature map to the detection optimization layer. After receiving the feature map, the robust target detection module generates candidate regions, completes target classification and bounding box regression, and then optimizes and outputs the final detection result. The system feedback module analyzes the detection results, generates feedback signals, identifies sample information that is difficult to analyze and process, performs online learning, and updates the sample database. The feedback signals are also further transmitted back to the perception decision layer to dynamically adjust the decision threshold and continuously optimize the system's working state.

[0042] Example 2

[0043] like Figure 1 - Figure 2As shown, when the high-frequency capture and data supplementation module receives the working mode switching instruction from the perception decision layer, it immediately starts the workflow. The module first parses the mode switching instruction through the internal instruction decoder to determine that it needs to enter the high-frequency working mode. The frame rate controller sends a precise frame rate control command to the image acquisition device, triggering the acquisition device to switch from the normal frame rate of 25fps to the high-frequency capture mode of 60fps or higher. Subsequently, the generated high frame rate image sequence is temporarily stored through the high-speed image buffer. The buffer adopts the first-in-first-out (FIFO) data management strategy, which can effectively alleviate the problem of mismatch between data production and consumption rates, thereby ensuring that the subsequent processing unit obtains a stable and continuous image data stream. Finally, the module outputs a continuous image sequence to the multi-frame information fusion module for further processing.

[0044] After receiving a continuous image sequence, the multi-frame information fusion module performs precise frame alignment and fusion processing. First, the multi-frame information fusion module extracts a fixed-length continuous frame sequence from the high-speed image buffer in sequence through a frame sequence selector. Then, the motion estimation engine uses an optical flow algorithm to calculate the motion vector field between the continuous frames to accurately capture inter-frame motion information. The motion compensator aligns all frames in the sequence to the reference frame coordinate system based on the obtained motion vector and uses a bilinear interpolation algorithm to process pixel coordinate transformation, thereby effectively eliminating pixel-level offset caused by target motion or camera shake. The weight calculation unit assigns adaptive fusion weights to each pixel of each aligned frame according to the image quality evaluation index. In this process, areas with higher gradient magnitudes are given higher weights to ensure that high-quality image areas dominate the fusion. Finally, the pixel-level fusion core uses a weighted average fusion algorithm to calculate by combining the aligned frames and corresponding pixel weights to generate a fused image with significantly improved signal-to-noise ratio and clearer details. This image can effectively highlight the continuous real target while suppressing transient weather interference such as rain and snow.

[0045] The attention-driven feature purification module receives the fused image and performs in-depth feature extraction and purification processing. The attention-driven feature purification module extracts rich multi-scale initial feature maps from the fused image through the backbone network, which contain multi-level visual features from low-level texture to high-level semantics. The interference suppression branch works in parallel through a decoder network consisting of 3 convolutional layers and 2 upsampling layers to learn and generate a mask for identifying weather interference areas. This mask limits the output value to the range [0,1] through an activation function, accurately identifying the noise areas that need to be suppressed in the feature map. The target enhancement branch generates an attention map for highlighting key targets through a similarly structured parallel decoder network. It also enhances the feature response of the area where the real target is located through activation function processing. The feature purification operator integrates the initial feature map, the weather interference mask, and the key target attention map to perform the core feature purification calculation, thereby suppressing noise areas and enhancing target areas. The purified feature map is stored in the purified feature map cache, waiting to be transmitted to the downstream detection optimization layer to provide high-quality feature representation for subsequent detection tasks.

[0046] In the detection optimization layer, the robust target detection module receives the cleaned feature map output by the feature engineering layer and performs accurate target detection analysis. The robust target detection module receives and caches the cleaned feature map from upstream through the feature input interface, and adopts a double buffering mechanism to ensure the stability of data supply. The region proposal network uses 3×3 convolution kernels to perform sliding window processing on the feature map, generating 9 anchor boxes of different scales at each spatial location to form candidate region proposals that may contain targets. The classification and regression heads process each candidate region in parallel: the classifier outputs the target category probability distribution through two fully connected layers and activation functions, accurately distinguishing different categories such as pedestrians and vehicles. The regressor fine-tunes the candidate region coordinates through the bounding box regression algorithm and outputs accurate target location information. The post-processing engine is responsible for refining the preliminary detection results. First, it removes overlapping redundant detection boxes through the non-maximum suppression algorithm, and then applies confidence threshold filtering to remove low-confidence false detections. Finally, it outputs high-precision detection boxes and category information.

[0047] The system feedback module receives the detection results and generates corresponding feedback signals. The performance monitor of the system feedback module continuously analyzes the detection results output by the robust target detection module. It uses a confidence distribution analyzer to statistically analyze the average detection confidence and its changing trend for each category. Simultaneously, it uses a temporal consistency checker to compare the detection results between consecutive frames. Based on the matching degree between the Kalman filter predicted trajectory and the actual detection position, it identifies abnormal fluctuation patterns and calculates the key performance indicators of the system. The sample information mining tool, which is difficult to analyze and process, automatically identifies and collects samples that the current model cannot detect correctly based on these performance indicators. These samples include low-confidence positive samples, high-confidence negative samples, and missed samples found by the background subtraction motion detection algorithm. The online learning engine is triggered under specific conditions. It calculates the update direction of the model parameters through the gradient calculation unit and uses the parameter updater to perform small-scale online fine-tuning of specific modules based on the stochastic gradient descent algorithm. All the mined sample information that is difficult to analyze and process, along with their annotation information, are stored in the sample database and managed using a relational database structure to provide sufficient data support for subsequent offline incremental training. The generated feedback signals are simultaneously transmitted to the perception decision layer and external systems, forming a self-optimizing system.

[0048] In this embodiment, the entire process begins with the high-frequency interception and data supplementation module of the feature engineering layer. By parsing instructions, the image acquisition frame rate is increased from 25fps to 60fps or higher, and a high-speed buffer is used to temporarily store the image sequence, providing stable and continuous high-quality data for subsequent processing. The multi-frame information fusion module performs fine processing on the continuous frame sequence: precise inter-frame alignment is achieved through motion estimation and compensation, followed by adaptive weight calculation and pixel-level fusion to generate a fused image with a significantly improved signal-to-noise ratio, effectively suppressing instantaneous interference such as rain and fog and highlighting the true target. The attention-driven feature purification module performs deep feature extraction and optimization on the fused image. After the backbone network extracts the initial feature map, the parallel interference suppression branch and target enhancement branch generate masks to identify noisy regions and highlight key targets, respectively. The attention map, through a feature cleansing operator, achieves noise suppression and target enhancement at the feature level, outputting a high-quality cleaned feature map. In the detection optimization layer, the robust target detection module receives the cleaned features, generates candidate target regions through a region proposal network, and completes accurate target classification and bounding box localization by a classification and regression head. Then, the post-processing engine filters redundant boxes and outputs high-precision detection results. The system feedback module forms a closed-loop optimization for the entire detection process: its performance monitor analyzes the detection confidence and temporal consistency, the sample information miner for difficult-to-analyze samples automatically collects incorrectly identified samples, and the online learning engine fine-tunes the model when conditions are triggered, storing all difficult examples in the database to support continuous learning, thereby ensuring that the system can continuously self-optimize and maintain excellent detection performance during long-term operation.

[0049] Example 3

[0050] The high-precision visual inspection system and method based on deep learning described in Embodiments 1 and 2 includes the following steps during operation:

[0051] S1: Environmental Perception and Intelligent Decision-Making

[0052] The perception and decision layer performs real-time analysis of the input video stream. The real-time weather perception module performs multi-frame temporal analysis through caching continuous frames, image preprocessing, and lightweight deep neural networks, and outputs specific weather classification results and quantified interference factors. The intelligent strategy switching module parses these results, compares the interference factors with internally preset decision thresholds, and generates working mode switching instructions based on the comparison results. These instructions are then sent to the image acquisition device and the downstream feature engineering layer, thereby realizing the dynamic switching of the system's working mode.

[0053] S2: High-frequency image acquisition and data buffering:

[0054] The high-frequency capture and data supplementation module in the feature engineering layer starts working after receiving the switching command. It parses the command content through the command decoder and drives the image acquisition device to switch from the normal frame rate to the high-frequency capture mode by the frame rate controller. The high frame rate image sequence generated is temporarily stored through the ring buffer structure of the high-speed image buffer, which solves the problem of matching the speed of data production and consumption, and provides a stable and high-quality continuous image sequence for subsequent processing.

[0055] S3: Multi-frame fusion and feature cleanup:

[0056] The multi-frame information fusion module receives a continuous image sequence and generates a fused image with a significantly improved signal-to-noise ratio through frame sequence selection, motion estimation, motion compensation, adaptive weight calculation, and pixel-level fusion. This effectively suppresses transient weather interference and highlights the real target. The attention-driven feature cleansing module performs deep feature extraction on the fused image. After the backbone network extracts the initial feature map, it generates a weather interference mask and a key target attention map through parallel interference suppression and target enhancement branches, respectively. Finally, the feature cleansing operator performs comprehensive processing to achieve noise suppression and target enhancement at the feature level, outputting a high-quality cleaned feature map.

[0057] S4: Target Detection and Result Output:

[0058] The robust target detection module in the detection optimization layer receives the cleaned feature map, caches the data through the feature input interface, and generates candidate target regions through the region proposal network. The classification and regression heads process each candidate region in parallel to complete the accurate classification and location regression of the target. The post-processing engine then refines the preliminary detection results by applying non-maximum suppression and confidence threshold filtering to remove redundant and low-confidence detection boxes, and finally outputs high-precision target detection boxes and category information.

[0059] S5: Performance Monitoring and Closed-Loop Feedback

[0060] The system feedback module continuously monitors the detection effect, the performance monitor analyzes the confidence distribution of the detection results and the temporal consistency between consecutive frames, calculates key performance indicators, and the sample information mining tool for difficult-to-analyze samples automatically identifies and collects sample information that is difficult to analyze and process based on these indicators, establishes a sample database, and when specific conditions are triggered, the online learning engine starts to perform small-scale online fine-tuning of the model to optimize system performance. The generated feedback signals are transmitted back to the perception decision layer to realize closed-loop optimization and adaptive adjustment of system parameters, ensuring that the system has the ability to continuously evolve.

[0061] Working principle:

[0062] like Figure 1 - Figure 2 As shown, the present invention operates as follows:

[0063] First, the perception and decision-making layer performs environmental perception and decision-making: the real-time weather perception module uses a lightweight deep neural network to perform multi-frame temporal analysis on the video stream, outputs weather classification results and quantifies interference factors, and the intelligent strategy switching module compares the interference factors with internal preset thresholds, generates working mode switching instructions and sends them to the image acquisition device and feature engineering layer to ensure that the enhanced processing mode is enabled in extreme weather.

[0064] Then, the feature engineering layer initiates the enhancement process according to the instructions: the high-frequency interception and data supplementation module increases the image acquisition frame rate from the usual 25fps to 60fps or higher, and provides a stable high frame rate image sequence using a high-speed image buffer. The multi-frame information fusion module performs precise processing on the sequence, and completes inter-frame alignment using motion estimation and compensation techniques. Combined with adaptive weight calculation and pixel-level fusion operation, a fused image with significantly improved signal-to-noise ratio is generated, effectively suppressing instantaneous interference such as rain and snow. The attention-driven feature purification module performs deep feature extraction on the fused image. After obtaining the initial feature map with the help of the backbone network, the parallel interference suppression branch and target enhancement branch generate weather interference mask and key target attention map, respectively. Finally, the feature purification operator realizes noise suppression and target enhancement at the feature level simultaneously, and outputs a high-quality purified feature map.

[0065] Finally, the robust target detection module of the detection optimization layer receives the purified feature map and completes accurate target classification and localization based on the region proposal network, classification and regression head. The post-processing engine outputs high-precision detection results, while the system feedback module continuously monitors the detection performance, analyzes confidence and temporal consistency to identify sample information that is difficult to analyze and process. When the conditions are met, the online learning engine is launched to fine-tune the model, and difficult example data is stored in the database to support incremental learning. The feedback signal is transmitted to the perception decision layer to build a complete self-optimization closed loop. The precise cooperation of each layer maintains reliable detection performance in complex environments such as heavy rain and fog, and continuously improves detection accuracy.

[0066] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A high-precision visual inspection system based on deep learning, characterized in that, include: The perception and decision layer transmits the collected data to the feature engineering layer. The feature engineering layer analyzes and processes the data and then transmits it to the detection and optimization layer. The detection and optimization layer analyzes the data processed by the feature engineering layer and then transmits the prediction results to the perception and decision layer. The perception and decision-making layer includes a real-time weather perception module that performs real-time analysis of video streams and an intelligent strategy switching module with decision thresholds. The feature engineering layer includes a high-frequency interception and data supplementation module for switching the frequency interception mode of the image acquisition device, a multi-frame information fusion module for highlighting persistent targets, and an attention-driven feature cleanup module for receiving the initial feature map extracted by the backbone network. The detection optimization layer includes a high-performance robust target detection module that takes the cleaned feature map as input and a system feedback module that takes the confidence level of the detection results and historical data as feedback signals.

2. The high-precision visual inspection system based on deep learning according to claim 1, characterized in that: The perception and decision layer includes a real-time weather perception module, which analyzes the video stream and outputs weather classification results and interference factors. The intelligent strategy switching module receives the weather classification results and interference factors, compares them with a preset decision threshold, and sends a working mode switching command to the image acquisition device and the feature engineering layer based on the comparison results.

3. The high-precision visual inspection system based on deep learning according to claim 1, characterized in that: The feature engineering layer includes a high-frequency interception and data supplementation module, which receives a working mode switching instruction and outputs a continuous image sequence. The multi-frame information fusion module receives the continuous image sequence and performs frame alignment and fusion processing to output a fused image. The feature purification module receives the fused image and performs feature extraction and purification processing to output the processed feature map to the detection optimization layer.

4. The high-precision visual inspection system based on deep learning according to claim 1, characterized in that: The detection optimization layer includes a robust target detection module, which receives the feature map output by the feature engineering layer and performs target detection, outputting the detection result. The system feedback module receives the detection result and generates a feedback signal, which is then transmitted to the perception decision layer and the external system.

5. A high-precision visual inspection system based on deep learning according to claim 2, characterized in that: The real-time weather perception module includes a frame buffer that stores continuous frames, an image preprocessing unit that standardizes the buffered frames, and a lightweight deep neural network engine that extracts features from the preprocessed image and outputs weather classification results and interference factors. The intelligent strategy switching module includes an input parser that parses data from the weather sensing module, a threshold register that stores decision thresholds, a decision logic comparator that compares interference factors with thresholds, a mode state machine that controls the system's operating state, and an instruction encoder that generates control instructions based on the state machine's state. The real-time weather perception module receives and stores continuous frames of the video stream through a frame buffer, performs standardization processing on the continuous frames through an image preprocessing unit, extracts features from the standardized images through a lightweight deep neural network engine, and outputs weather classification results and interference factors. The intelligent strategy switching module receives and parses the weather classification results and interference factors through an input parser, compares the interference factors with the decision thresholds stored in the threshold register through a decision logic comparator, controls the system's working state according to the comparison results through a mode state machine, and generates control commands based on the working state through an instruction encoder and sends them to the image acquisition device and the feature engineering layer.

6. The high-precision visual inspection system based on deep learning according to claim 3, characterized in that: The high-frequency interception and data supplementation module includes an instruction decoder for parsing mode switching instructions, a frame rate controller for sending frame rate control commands to the image acquisition device, and a high-speed image buffer for storing high frame rate image sequences. The multi-frame information fusion module includes a frame sequence selector that sequentially extracts a fixed-length continuous frame sequence from a high-speed image buffer; a motion estimation engine that calculates inter-frame motion vectors; a motion compensator for frame alignment processing; a weight calculation unit that calculates pixel fusion weights; and a pixel-level fusion core that generates a fused image. The attention-driven feature cleansing module includes a backbone network that extracts initial feature maps from the fused image, an interference suppression branch that generates a weather interference mask, a target enhancement branch that generates a key target attention map, a feature cleansing operator that performs feature cleansing calculations, and a cleaned feature map cache that stores the output feature maps.

7. A high-precision visual inspection system based on deep learning according to claim 4, characterized in that: The robust object detection module includes a feature input interface for receiving and caching feature maps, a region proposal network for generating candidate regions that may contain objects on the feature maps, a classification and regression head for object classification and bounding box regression, and a post-processing engine for filtering and optimizing detection results. The system feedback module includes a performance monitor that analyzes detection results and calculates performance indicators; a sample information mining tool that identifies and collects sample information that is difficult to analyze and process; an online learning engine that performs online adjustment of model parameters; and a sample database that stores sample information and annotation information that are difficult to analyze and process.

8. A high-precision visual inspection method based on deep learning, and a high-precision visual inspection system based on deep learning according to any one of claims 1-7, characterized in that: Includes the following steps: S1: Environmental Perception and Intelligent Decision-Making The perception and decision layer performs real-time analysis of the input video stream. Through the real-time weather perception module, it caches continuous frames, performs image preprocessing and lightweight deep neural network analysis, completes multi-frame temporal analysis, and outputs weather classification results and quantified interference factors. The intelligent strategy switching module parses the results, compares the interference factors with the preset decision threshold, and generates a working mode switching command based on this, which is sent to the image acquisition device and feature engineering layer to realize the dynamic switching of the system's working mode. S2: High-frequency image acquisition and data buffering: The high-frequency capture and data supplementation module in the feature engineering layer starts after receiving the switching command. It parses the content through the command decoder, and the frame rate controller controls the image acquisition device to switch to the high-frequency capture mode. It also uses the ring buffer structure of the high-speed image buffer to temporarily store the high frame rate image sequence to solve the problem of matching the data production and consumption rate, and to provide a stable and continuous image sequence for subsequent processing. S3: Multi-frame fusion and feature cleanup: The multi-frame information fusion module receives a continuous image sequence, performs frame sequence selection, motion estimation and compensation, weight calculation and pixel-level fusion to generate a high signal-to-noise ratio fused image, suppresses instantaneous interference and highlights the real target. The feature purification module performs deep feature extraction on the fused image, obtains an initial feature map through the backbone network, and generates a weather mask and a target attention map by parallel interference suppression and target enhancement branches, respectively. Finally, the feature purification operator realizes noise suppression and target enhancement and outputs a purified feature map.

9. The high-precision visual detection method based on deep learning according to claim 8, characterized in that: S4: Target Detection and Result Output: The robust target detection module in the detection optimization layer receives the cleaned feature map, which is then cached by the feature interface. The region proposal network generates candidate regions, and the classification and regression heads process each candidate region in parallel to complete target classification and location regression. The post-processing engine refines the results by removing redundant boxes through non-maximum suppression and confidence filtering, and finally outputs high-precision detection boxes and category information. S5: Performance Monitoring and Closed-Loop Feedback The system feedback module continuously monitors the detection effect, the performance monitor analyzes the confidence level and inter-frame consistency to calculate key indicators, the sample information mining machine for difficult-to-analyze samples automatically identifies and collects samples and builds a database, the online learning engine starts fine-tuning when conditions are triggered to optimize model performance, and the feedback signal is sent back to the perception decision layer to realize closed-loop optimization and adaptive adjustment of system parameters, ensuring continuous evolution capability.

Images