High-robustness target detection method and device for reflective transparent body

Abstract

The invention discloses a high-robustness target detection method for a reflective transparent body, and the method comprises the following steps: a polarization difference image collection and preprocessing step: collecting images of the reflective transparent body in at least two different polarization directions, and generating a polarization difference image for eliminating specular reflection interference based on the images; a mixed scale feature extraction and detection step: inputting the polarization difference image into a mixed scale convolutional neural network, extracting multi-dimensional features, and outputting contour positioning information of the reflective transparent body and a defect identification result containing suspected defect information based on the features; and a tactile feedback closed loop verification step: based on the suspected defect information, controlling an execution mechanism to contact a corresponding area of the reflective transparent body, collecting a physical signal in the contact process, and verifying the authenticity of the suspected defect based on the physical signal. According to the method, high-precision and high-robustness defect detection of the reflective transparent body is realized through polarization difference and mixed scale feature extraction in combination with tactile feedback verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of machine vision and automatic detection, in particular to a high-robustness target detection method and device for a reflective transparent body. BACKGROUND

[0002] In the production and quality inspection links of the medical, food and chemical industries, the appearance quality detection of glass test tubes, bottle bodies and other reflective transparent bodies is crucial. However, the glass test tubes, bottle bodies and other reflective transparent bodies pose special challenges to machine vision detection due to their optical properties. First, their smooth surfaces can produce strong specular reflection, forming a high-brightness overexposure area in the image, which can easily obscure the imaging features of key defects such as micro-cracks and scratches, resulting in a high detection system miss rate. Second, the low contrast between the edge profile and the background of the transparent material, as well as the weak internal texture features such as bubbles, makes it difficult for traditional neural networks using fixed-scale convolution kernels to simultaneously and effectively extract the macro profile and micro details, resulting in incomplete feature extraction and limited detection accuracy.

[0003] In addition, the existing detection schemes have obvious limitations in applicability. On the one hand, most methods rely on data sets collected in specific scenarios for training, and when faced with new application scenarios of different sizes, materials or lighting conditions, the model's generalization ability is insufficient, and the detection performance decreases significantly. In addition, the functions of existing devices are often single, making it difficult to balance profile positioning, surface stain detection and accurate identification of internal micro-cracks and other deep defects, while micro-cracks and other defects directly affect the safety and sealing of products, which is a shortcoming in current detection capabilities. Therefore, there is an urgent need for a comprehensive detection scheme that can effectively overcome reflection interference, achieve multi-dimensional feature extraction, and have good cross-scene adaptability. SUMMARY

[0004] To solve the above technical problems in the related art, the present application provides a high-robustness target detection method and device for a reflective transparent body, which can overcome the above shortcomings of the prior art.

[0005] To achieve the above technical purposes, the technical solution of the present application is as follows: A high-robustness target detection method for a reflective transparent body; The high-robustness target detection method for a reflective transparent body includes the following steps: Polarization difference image acquisition and preprocessing step: acquiring images of the reflective transparent body under at least two different polarization directions, and generating polarization difference images that eliminate specular reflection interference based on the images; The mixed-scale feature extraction and detection step: inputting the polarization difference image into a mixed-scale convolutional neural network, extracting multi-dimensional features, and outputting the contour positioning information of the light-reflecting transparent body and the defect recognition result containing suspected defect information based on the features; The tactile feedback closed-loop verification step: based on the suspected defect information, controlling an actuator to contact the corresponding area of the light-reflecting transparent body, collecting physical signals during the contact process, and verifying the authenticity of the suspected defect based on the physical signals.

[0006] Further, the polarization difference image acquisition and preprocessing step includes: Based on the Stokes vector method, calculating a polarization degree image according to the images under the at least two different polarization directions; using an adaptive threshold segmentation method to process the polarization degree image to separate the specular reflection area and the target area, and generate the polarization difference image; Wherein, the adaptive threshold segmentation method is: taking the average value of the maximum and minimum gray values in the predetermined neighborhood window of the pixel point in the polarization degree image as the dynamic threshold value of the pixel point for segmentation.

[0007] Further, the polarization difference image acquisition and preprocessing step further includes: Before calculating the polarization degree image, based on image entropy, screening the optimal polarization direction combination for calculating the polarization difference image from a plurality of candidate polarization direction combinations.

[0008] Further, in the mixed-scale feature extraction and detection step, the mixed-scale convolutional neural network extracts features and fuses them through at least two parallel convolution branches with different convolution kernel sizes; the channel attention mechanism is introduced in the fusion process.

[0009] Further, the at least one convolution branch uses deformable convolution for feature extraction.

[0010] Further, the tactile feedback closed-loop verification step includes: Dynamically determining the contact force according to the confidence included in the suspected defect information; collecting pressure data and vibration frequency data during the contact process; based on a dual-mode discriminant model that fuses the pressure data and vibration frequency data, determining whether the suspected defect is a real defect.

[0011] Further, the method further includes a model iterative optimization step: Adding the verified false detection samples and the defect samples synthesized by the generative adversarial network to the training data set, and using a transfer learning method combining the domain adversarial neural network and the elastic weight consolidation mechanism to update the mixed-scale convolutional neural network.

[0012] According to another aspect of the present application, there is provided a high-robustness target detection device for a light-reflecting transparent body, The high-robustness target detection device for a light-reflecting transparent body comprises: a polarization-differential imaging module, configured to perform the polarization-differential image acquisition and preprocessing step; a hybrid-scale detection module, connected to the polarization imaging module, configured to perform the hybrid-scale feature extraction and detection step; a haptic feedback closed-loop module, connected to the hybrid-scale detection module, configured to perform the haptic feedback closed-loop verification step.

[0013] Further, the polarization imaging module comprises an adjustable polarized light source, a polaroid carousel, and an industrial camera; the polaroid carousel is configured to provide the at least two different polarization directions.

[0014] Further, the haptic feedback closed-loop module comprises a mechanical arm, a pressure sensor arranged at the end of the mechanical arm, and a controller; the device further comprises a vibration sensor, and the pressure sensor and the vibration sensor are configured to acquire the physical signal.

[0015] The present application has the following beneficial effects: the polarization-differential imaging effectively suppresses the specular reflection interference on the target surface, so that the target details are clearly presented; the hybrid-scale feature extraction network is combined to synchronously acquire the contour and texture information, and the cross-domain adaptation strategy is used to enhance the model generalization, so as to achieve the purpose of high-precision positioning and comprehensive defect identification of the light-reflecting transparent body in a variable scene; further, the haptic feedback verification mechanism is introduced, the visual suspected defects are secondarily determined through multi-modal information fusion, the false detection rate of the system is significantly reduced, and finally the comprehensive improvement of detection reliability, accuracy, and environmental adaptability is realized. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0017] Fig. 1 is a workflow diagram of the polarization-differential imaging module according to the embodiments of the present application; Fig. 2 is a structural schematic diagram of the hybrid-scale convolutional neural network according to the embodiments of the present application; Fig. 3 is a whole flowchart of a high-robustness target detection method for a light-reflecting transparent body according to the embodiments of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0019] As shown in the figure, the high-robustness target detection method of the reflective transparent body according to the embodiment of the present application comprises the following steps: Figs. 1-3 Polarization difference image acquisition and preprocessing step: acquiring images of the reflective transparent body under at least two different polarization directions, and generating polarization difference images eliminating specular reflection interference based on the images; Hybrid scale feature extraction and detection step: inputting the polarization difference images into a hybrid scale convolutional neural network, extracting multi-dimensional features, and outputting the contour positioning information of the reflective transparent body and the defect recognition result containing suspected defect information based on the features; Tactile feedback closed-loop verification step: based on the suspected defect information, controlling an actuator to contact the corresponding area of the reflective transparent body, acquiring physical signals during the contact process, and verifying the authenticity of the suspected defect based on the physical signals. The high-robustness target detection method of the reflective transparent body according to the embodiment of the present application, in a specific embodiment, the polarization difference image acquisition and preprocessing step comprises:

[0020] Based on the Stokes vector method, calculating a polarization degree image according to the images under the at least two different polarization directions; using an adaptive threshold segmentation method to process the polarization degree image to separate the specular reflection area and the target area, and generating the polarization difference image; Wherein, the adaptive threshold segmentation method is: taking the average value of the maximum and minimum gray values in the predetermined neighborhood window of the pixel point in the polarization degree image as the dynamic threshold value of the pixel point for segmentation. The high-robustness target detection method of the reflective transparent body according to the embodiment of the present application, in a specific embodiment, the polarization difference image acquisition and preprocessing step further comprises:

[0021] Before calculating the polarization degree image, based on image entropy, screening the optimal polarization direction combination for calculating the polarization difference image from a plurality of candidate polarization direction combinations.

[0022] ​According to the high-robustness target detection method of the reflective transparent body, in a specific embodiment, in the mixed-scale feature extraction and detection step, the mixed-scale convolutional neural network extracts features and performs fusion through at least two parallel convolution branches with different convolution kernel sizes; and a channel attention mechanism is introduced in the fusion process.

[0023] According to the high-robustness target detection method of the reflective transparent body, in a specific embodiment, the at least one convolution branch uses deformable convolution to extract features.

[0024] According to the high-robustness target detection method of the reflective transparent body, in a specific embodiment, the haptic feedback closed-loop verification step includes: The contact force is dynamically determined according to the confidence included in the suspected defect information; pressure data and vibration frequency data in the contact process are collected; and whether the suspected defect is a real defect is determined based on a bimodal discrimination model that fuses the pressure data and the vibration frequency data.

[0025] According to the high-robustness target detection method of the reflective transparent body, in a specific embodiment, the method further includes a model iterative optimization step: The verified false detection samples and defect samples synthesized by the generative adversarial network are added to a training data set, and a transfer learning method combining a domain adversarial neural network and an elastic weight consolidation mechanism is used to update the mixed-scale convolutional neural network.

[0026] In a second aspect, a high-robustness target detection device of a reflective transparent body is provided, The high-robustness target detection device of the reflective transparent body includes: A polarization differential imaging module is configured to perform the polarization differential image acquisition and preprocessing step. A mixed-scale detection module is connected to the polarization imaging module and is configured to perform the mixed-scale feature extraction and detection step. A haptic feedback closed-loop module is connected to the mixed-scale detection module and is configured to perform the haptic feedback closed-loop verification step.

[0027] According to the high-robustness target detection device of the reflective transparent body, in a specific embodiment, the polarization imaging module includes an adjustable polarized light source, a polarization plate turntable, and an industrial camera; and the polarization plate turntable is configured to provide the at least two different polarization directions.

[0028] In a specific embodiment, the tactile feedback closed loop module comprises a mechanical arm, a pressure sensor arranged at the end of the mechanical arm, and a controller; the device further comprises a vibration sensor, and the pressure sensor and the vibration sensor are used to collect the physical signal.

[0029] In order to facilitate the understanding of the above technical solutions of the present application, the above technical solutions of the present application are described in detail below through specific implementation details and principles.

[0030] In specific use, the high-robustness target detection device for a reflective transparent body according to the present application comprises a polarization imaging module, a hybrid scale detection module, a tactile feedback closed loop module, and a human-computer interaction module, and the specific structure is as follows: The polarization imaging module is composed of an adjustable polarized light source (wavelength 450-650 nm), a polarizer carousel (containing four polarization directions 0°, 45°, 90°, and 135°), an industrial camera (resolution ≥ 2048×1536), and an image acquisition card, and is used to collect images of the reflective transparent body under different polarization states and generate polarization difference images.

[0031] The hybrid scale detection module is based on an FPGA+GPU heterogeneous computing architecture and integrates a hybrid scale convolutional neural network (MS-CNN), which is used to extract the edge profile and internal texture features of the image and realize target detection and defect recognition.

[0032] The tactile feedback closed loop module is composed of a mechanical arm, a pressure sensor (accuracy ±0.01 N), and a displacement sensor, and when a suspected defect is detected visually, the mechanical arm is driven to lightly touch the target area, the defect authenticity is verified through tactile feedback, and the false detection rate is reduced.

[0033] The human-computer interaction module is configured with a touch screen and upper computer software, supports detection parameter setting, detection result visualization (defect position, type, and size labeling), and historical data query.

[0034] The detailed design of the key modules is as follows: (1) Polarization difference imaging module This module eliminates the interference of glass surface specular reflection by adjusting the polarization state of the light source, and the specific working process is as follows: 1. Multi-polarization state image acquisition: control the polarizer carousel to switch the four polarization directions (0°, 45°, 90°, and 135°) in turn, and the adjustable polarized light source emits light of the corresponding polarization state to irradiate the reflective transparent body (such as a test tube), and the industrial camera synchronously collects four images I0, I 45 , I 90 , and I 135 .

[0035] 2. Polarization difference calculation: The polarization degree image P is calculated using the Stokes vector method, and the formula is P = V [(I0-I 90 )²+(I 45 -I 135 )²] / (I0+I 90 +I 45 +I 135 ). An adaptive threshold mechanism is introduced, and an adaptive threshold segmentation method combining the Sobel operator and Bernsen algorithm is used. The average value of the maximum and minimum gray values in the 3x3 window is calculated as the dynamic threshold T(x,y), which replaces the fixed threshold, effectively suppressing the segmentation error caused by non-uniform illumination. At the same time, an image entropy-based polarization angle selection algorithm is added to calculate the information entropy of the image under each polarization direction combination, automatically selecting the polarization direction combination with the maximum information entropy (such as 0° and 90°, 45° and 135° optimal combination), further improving the target detail retention of the difference image. The specular reflection region (P≥T(x,y)) and the target region (P<T(x,y)) are separated by the dynamic threshold to generate the reflection-eliminated difference image.

[0036] 3. Image enhancement: The difference image is processed by adaptive histogram equalization to improve the contrast of target edges and background, laying the foundation for subsequent feature extraction.

[0037] (2) Mixed scale detection module Multi-dimensional feature extraction is realized based on MS-CNN, and the network structure is as follows: Input layer: receives the polarization difference image (size 2048x1536) and performs normalization processing (pixel value mapping to 0-1).

[0038] Mixed scale convolution layer: contains three parallel convolution branches, respectively using 1x1 (extracting detail texture), 3x3 (extracting local features), and 7x7 (extracting edge contour) convolution kernels, and the output feature maps of each branch enter the feature fusion layer. The fusion layer introduces a channel attention module to enhance the feature response of key defects such as micro-cracks and bubbles by adaptive weight distribution; at the same time, deformable convolution is used instead of traditional fixed convolution kernel, which improves the feature capture ability of irregular defects such as curved cracks and irregular bubbles by dynamically adjusting the convolution sampling position.

[0039] Feature pyramid layer: multi-scale down-sampling (scale factor 2, 4, 8) is performed on the fused feature map to construct a feature pyramid, realizing the detection adaptation of different sizes of reflective transparent bodies.

[0040] Detection head: Use dual-output detection head, output target contour coordinates (positioning error ≤0.5mm) and defect type (microcrack, bubble, stain) respectively, among which microcrack detection is realized by strengthening crack area feature response through exclusive attention branch to achieve accurate identification of cracks ≥0.1mm wide; In the output stage of the detection head, integrate Grad-CAM heat map generation function to visualize the key feature area of defect identification and provide explainability basis for the detection result.

[0041] (3) Cross-domain adaptive dataset construction To improve the generalization of the model, a cross-domain adaptive dataset and a transfer learning strategy are constructed: Dataset collection: Collect images of reflective transparent bodies in different scenarios, covering material differences (ordinary glass, quartz glass), size differences (diameter 5-50mm test tube), light differences (brightness 1000-5000lux), and defect types (microcrack, bubble, scratch), a total of 100,000+ samples, divided into training set, validation set and test set according to the ratio of 7:2:1.

[0042] Generative data augmentation: Construct a sample synthesis module based on GAN (Generative Adversarial Network), for extremely fine microcracks (width <0.1mm), special angle scratches and other difficult-to-naturally-collect sample types, generate high-fidelity defect samples through generative model to supplement the scarce categories of the dataset and improve the diversity and class balance of the dataset.

[0043] Domain adaptive training: Use domain adversarial neural network (DANN), add a domain classifier in MS-CNN, minimize the distribution difference between the source domain (fully labeled scenarios) and the target domain (rarely labeled scenarios) through adversarial training; At the same time, integrate EWC (Elastic Weight Consolidation) mechanism, estimate the importance of model parameters to old scene knowledge through Fisher information matrix, and impose a penalty on the update of key parameters to support the model not to forget existing knowledge when learning new scene data, realize incremental learning and online update.

[0044] (4) Haptic feedback closed-loop verification When visual detection detects suspected microcracks and other defects, start haptic feedback closed-loop verification: 1. Defect positioning: Get the defect coordinates (x, y) output by visual detection, drive the robot arm to move above the area, and control the robot arm to lower the height (1mm from the target surface) through displacement sensor.

[0045] 2. Multi-modal tactile sampling: Introduce reinforcement learning Q-learning decision mechanism, dynamically adjust the trigger threshold and contact force of tactile verification based on historical validation data - 0.08N gentle contact for high confidence (≥0.85) suspected defects, 0.1N standard contact for low confidence (0.6-0.85) suspected defects; The pressure sensor at the end of the robot arm synchronously collects the pressure distribution data and vibration frequency data of the contact area with the vibration sensor.

[0046] 3. Defect verification: Build a "pressure-vibration" dual-mode discriminant model, integrate pressure mutation features (such as micro-crack pressure drop ≥0.02N) and vibration frequency anomaly features (such as crack area vibration frequency offset ≥5Hz) for comprehensive judgment, if both dual-mode features meet the defect judgment conditions, it is judged as a real defect; Otherwise, it is a false detection. The model automatically annotates the false detection samples and adds them to the training set to dynamically optimize the detection threshold and verification decision strategy.

[0047] The high-robustness target detection method of the reflective transparent body described in the application comprises the following steps: Step 1: Polarization difference image acquisition and preprocessing: Collect multi-polarization state images through the polarization imaging module, select the optimal polarization direction combination based on image entropy, generate polarization difference images using Stokes vector method and adaptive threshold segmentation, and complete image enhancement through adaptive histogram equalization.

[0048] Step 2: Hybrid scale feature extraction: Input the preprocessed image into MS-CNN, extract multi-dimensional features through hybrid scale convolution layer (including attention mechanism and deformable convolution) and feature pyramid layer.

[0049] Step 3: Target detection and defect identification: The detection head outputs the reflective transparent body contour coordinates, defect type and feature heat map to obtain the preliminary detection result (including micro-cracks).

[0050] Step 4: Tactile feedback closed-loop verification: Start multi-modal tactile verification based on Q-learning strategy, determine the authenticity of the defect through the "pressure-vibration" dual-mode model, and output the final detection result (real defect / false detection).

[0051] Step 5: Model iterative optimization: Add false detection samples, new scene samples and GAN synthesized samples to the cross-domain adaptive dataset, update the MS-CNN model through DANN+EWC transfer learning, and continuously improve the detection robustness and generalization ability.

[0052] A preferred specific embodiment is as follows: 1. Device hardware deployment The polarization imaging module selects a Basler acA2040-90um industrial camera (resolution 2048x1536), and a Thorlabs adjustable polarized light source (wavelength 520 nm); the mixed scale detection module adopts an Xilinx Zynq UltraScale + FPGA and an NVIDIA Jetson AGX GPU; the tactile feedback module adopts an ABB IRB 120 mechanical arm, a FUTEK LCM300 pressure sensor (accuracy ±0.01N) and a PCB Piezotronics vibration sensor; and the human-computer interaction module is configured with a 10-inch touch screen and customized upper computer software.

[0053] 2. Test tube micro crack detection experiment example 1. Image acquisition: a glass test tube with a diameter of 15 mm is placed in the detection station, the polarizer disc is switched in four directions to collect images, the optimal polarization direction combination of 0° and 90° is selected through an image entropy algorithm, a polarization difference image is calculated, and the test tube surface reflection is effectively eliminated.

[0054] 2. Feature extraction and detection: the MS-CNN extracts the test tube edge contour (positioning error 0.3 mm) through mixed scale convolution and deformable convolution, the feature response is strengthened through a channel attention module, a suspected micro crack (size 0.12 mm x 2 mm) is detected on the pipe wall, and the upper computer synchronously generates a defect area heat map.

[0055] 3. Tactile verification: based on the Q-learning strategy, since the confidence of the suspected defect is 0.78, the mechanical arm moves to the target area with a standard contact force of 0.1N, the pressure sensor detects a pressure drop of 0.03N, the vibration sensor detects a vibration frequency offset of 7Hz, and the dual-mode model determines that it is a real micro crack.

[0056] Result output: the upper computer software labels the crack position, size and heat map, generates a detection report, and adds the sample to a cross-domain data set, and the model is updated through the EWC mechanism.

[0057] In summary, by means of the above technical scheme of the present application, the mirror reflection interference of the target surface is effectively suppressed through polarization difference imaging, so that the target details are clearly displayed; the profile and texture information are synchronously acquired by combining a mixed scale feature extraction network, and the model generalization is enhanced by using a cross-domain adaptation strategy, so as to achieve the purpose of high-precision positioning and comprehensive defect identification of the light-reflecting transparent body in a variable scene; further, a tactile feedback verification mechanism is introduced, the visual suspected defect is secondarily determined through multi-modal information fusion, the false detection rate of the system is significantly reduced, and finally the comprehensive improvement of detection reliability, accuracy and environmental adaptability is realized.

[0058] The above merely provides the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A highly robust object detection method for a retro-reflective transparency, characterized in that, Includes the following steps: Polarization difference image acquisition and preprocessing steps: Acquire images of a reflective transparent object in at least two different polarization directions, and generate a polarization difference image based on the images to eliminate specular reflection interference; Hybrid-scale feature extraction and detection steps: Input the polarization difference image into a hybrid-scale convolutional neural network to extract multi-dimensional features, and output the contour positioning information of the reflective transparent body and the defect identification result containing suspected defect information based on the features; Tactile feedback closed-loop verification steps: Based on the suspected defect information, control the actuator to contact the corresponding area of ​​the reflective transparent body, collect physical signals during the contact process, and verify the authenticity of the suspected defect based on the physical signals.

2. The highly robust target detection method for reflective transparent bodies according to claim 1, characterized in that, The polarization difference image acquisition and preprocessing steps include: Based on the Stokes vector method, a polarization degree image is calculated from the images under at least two different polarization directions; an adaptive threshold segmentation method is used to process the polarization degree image to separate the specular reflection region from the target region, thereby generating the polarization difference image; The adaptive threshold segmentation method is as follows: the average of the maximum and minimum gray values ​​within a predetermined neighborhood window of a pixel in the polarization image is used as the dynamic threshold for segmentation of that pixel.

3. The highly robust target detection method for reflective transparent bodies according to claim 1, characterized in that, The polarization difference image acquisition and preprocessing steps also include: Before calculating the polarization degree image, the optimal polarization direction combination for calculating the polarization difference image is selected from multiple candidate polarization direction combinations based on image entropy.

4. The highly robust target detection method for reflective transparent bodies according to claim 1, characterized in that, In the hybrid scale feature extraction and detection step, the hybrid scale convolutional neural network extracts features and fuses them through at least two parallel convolutional branches with different kernel sizes; a channel attention mechanism is introduced during the fusion process.

5. The highly robust target detection method for reflective transparent bodies according to claim 4, characterized in that, The at least one convolutional branch uses deformable convolution for feature extraction.

6. The highly robust target detection method for reflective transparent bodies according to claim 1, characterized in that, The haptic feedback closed-loop verification steps include: The contact force is dynamically determined based on the confidence level contained in the suspected defect information; pressure data and vibration frequency data are collected during the contact process; and a dual-modal discrimination model that integrates the pressure data and vibration frequency data is used to determine whether the suspected defect is a real defect.

7. The highly robust target detection method for reflective transparent bodies according to claim 1, characterized in that, The method also includes a model iterative optimization step: The verified false detection samples and the defective samples synthesized by the generative adversarial network are added to the training dataset, and the hybrid-scale convolutional neural network is updated using a transfer learning method that combines a domain adversarial neural network with an elastic weight consolidation mechanism.

8. A highly robust target detection device for reflective transparent bodies, characterized in that, The apparatus for implementing the method of any one of claims 1 to 7 comprises: The polarization difference imaging module is used to perform the polarization difference image acquisition and preprocessing steps. A hybrid scale detection module, connected to the polarization imaging module, is used to perform the hybrid scale feature extraction and detection steps; The tactile feedback closed-loop module is connected to the hybrid scale detection module and is used to perform the tactile feedback closed-loop verification step.

9. The highly robust target detection device for reflective transparent bodies according to claim 8, characterized in that, The polarization imaging module includes an adjustable polarization light source, a polarizer turntable, and an industrial camera; the polarizer turntable is used to provide the at least two different polarization directions.

10. A highly robust target detection device for reflective transparent bodies according to claim 8, characterized in that, The tactile feedback closed-loop module includes a robotic arm, a pressure sensor disposed at the end of the robotic arm, and a controller; the device also includes a vibration sensor, the pressure sensor and the vibration sensor being used to collect the physical signals.

Images