Rigid body ultrasonic flaw detection system and method based on image feature extraction

Abstract

The invention discloses a rigid body ultrasonic flaw detection system and method based on image feature extraction, and the system comprises a data preprocessing module which is used for decoding a real-time video stream into a time sequence frame, extracting multi-scale features, and segmenting a multi-angle regional rigid body flaw detection waveform image; the data enhancement module is used for carrying out super-resolution normalization on the rigid body flaw detection waveform images subjected to region segmentation to obtain rigid body flaw detection waveform images with uniform sizes; the flaw detection result detection module is used for performing waveform detection classification on the rigid body flaw detection waveform images with the uniform size to obtain a detection classification result and confidence; and the result display module is used for visually displaying the result. According to the method, frame skipping detection is realized by utilizing inter-frame feature similarity calculation, and full-process calculation is only performed on key change frames, so that the redundant calculation amount is greatly reduced, the system can still keep high frame rate and low time delay under the real-time flow condition, and the online flaw detection requirement of a rigid body high-speed production line is met.

Description

Technical Field

[0001] This invention belongs to the field of industrial scene quality inspection and intelligent visual analysis technology, specifically involving a rigid body ultrasonic flaw detection system and method based on image feature extraction. Background Technology

[0002] Rigid bodies are prone to internal or near-surface defects during processing such as rolling, welding, and heat treatment, such as inclusions, porosity, cracks, delamination, and incomplete penetration. These defects may lead to structural failure and fatigue crack propagation during use, seriously threatening equipment safety.

[0003] Traditional ultrasonic flaw detection methods determine the presence and location of defects by manually observing ultrasonic scanning waveform images. However, the above methods have the following problems:

[0004] 1. High dependence on manual labor: Testing personnel need long-term training to identify different waveform characteristics, resulting in large subjective errors;

[0005] 2. Low efficiency: In a continuous production line, hundreds of flaw detection waveforms may be generated every minute, and manual judgment cannot meet the real-time requirements;

[0006] 3. Data waste: Traditional algorithms are based on fixed thresholds or simple template matching, which cannot fully utilize image feature information;

[0007] 4. Significant noise interference: The complex internal structure of rigid bodies leads to uneven noise distribution in ultrasound images, making them prone to misinterpretation.

[0008] Deep learning-based visual inspection methods have brought new solutions to non-destructive testing. Although traditional one-stage detection algorithms have high real-time performance, their detection accuracy is not high under noise interference and it is difficult to model frame-level temporal information. While traditional two-stage detection architectures have higher accuracy, their computational complexity is high and they are not suitable for real-time processing of high-speed video streams.

[0009] Therefore, it is necessary to propose a rigid body ultrasonic flaw detection system and method based on image feature extraction, which combines real-time stream acquisition, inter-frame similarity detection, feature extraction and two-stage detection mechanism to achieve efficient, low-latency and accurate detection of rigid body ultrasonic flaw detection images. Summary of the Invention

[0010] In view of the shortcomings of the prior art, the purpose of this invention is to provide a rigid body ultrasonic flaw detection system and method based on image feature extraction, so as to solve the problems of strong manual dependence, low efficiency, data waste and obvious noise interference in the existing ultrasonic flaw detection methods.

[0011] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0012] The present invention provides a rigid body ultrasonic flaw detection system based on image feature extraction, comprising: a data preprocessing module, a data enhancement module, a flaw detection result detection module, and a result display module;

[0013] The data preprocessing module is used to decode the real-time video stream into time-series frames, extract multi-scale features, and segment the multi-angle regional rigid body flaw detection waveform images.

[0014] The data augmentation module is used to perform super-resolution normalization on the rigid body flaw detection waveform image after region segmentation to obtain a rigid body flaw detection waveform image of uniform size.

[0015] The flaw detection result detection module is used to perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels.

[0016] The results display module is used to visualize the detection and classification results and their corresponding confidence levels.

[0017] Furthermore, the data preprocessing module specifically performs the following steps:

[0018] (11) Collect real-time video stream data from the ultrasonic flaw detector at the industrial site and convert it into structured image data that can be processed by the neural network. The video stream is decoded into time series frames, as shown in the following expression:

[0019] ;

[0020] in, These are the decoded time-series frames. For the time series Frame ultrasound image, size is ; The total number of frames for ultrasound images Perform grayscale normalization to reduce noise and eliminate brightness drift, resulting in the grayscale normalized [value]. Frame ultrasound images The expression is as follows:

[0021] ;

[0022] in, For the first Frame ultrasound images The horizontal and vertical axes are and The pixel value of the location; Indicates the first Mean grayscale value of a frame of ultrasound image. ; Indicates the first The standard deviation of grayscale values ​​in a frame of ultrasound images. ; To traverse each The summation function, It is a small constant;

[0023] (12) From the preprocessed grayscale normalized ultrasound image Extract multi-scale features; use the backbone feature extraction model. right Multi-scale features are extracted as follows:

[0024] ;

[0025] in, For multi-layer feature maps, each Corresponding to a feature layer, The global response weights for each layer are calculated using a channel attention mechanism. These weights are then multiplied channel by channel by the original feature map of each layer to enhance channels containing defect variations and suppress noise channels unrelated to defects, as follows:

[0026] ;

[0027] in, To enhance the defect sensitivity of the multi-layer feature map, This is the weight matrix. For global average pooling, for Activation function; multilayer feature map after enhancing defect sensitivity Each layer of features is multiplied by the feature weight coefficient of that layer to perform multi-layer feature fusion. The fused feature map is as follows:

[0028] ;

[0029] in, This is the fused global feature map. These are the feature weight coefficients for each layer;

[0030] (13) The fused global feature map The image was divided into 2×4, or 8, feature sub-regions containing ultrasound images. Each sub-region corresponds to the detection angle result of an ultrasonic flaw detector. A waveform segmentation model was used. Global feature map containing 8 ultrasound image feature sub-regions The region is segmented to obtain feature blocks of 8 corresponding sub-regions. The partitioning mapping is defined as follows:

[0031] ;

[0032] in, For the first Feature maps of each sub-region , For sub-regions in the global feature map The height and width range in the text.

[0033] Furthermore, the system also includes an acceleration module for improving detection speed through inter-frame similarity detection, specifically including:

[0034] Calculate the current time Compared to the previous moment global feature map If the similarity is greater than a certain threshold, the previous time step is used directly. The detection and classification results are used as the current moment The result, similarity is defined as:

[0035] ;

[0036] in, For inter-frame similarity; , These are the global feature maps for the current frame and the previous frame, respectively. This represents the dot product operation. This represents taking the modulus after expanding the vector; smoothing the similarity of consecutive frames ensures that waveform changes are judged based on the cumulative trend within the sliding window, thus avoiding misjudgments caused by single-frame noise. The expression is as follows:

[0037] ;

[0038] in, The similarity index is after smoothing. This is a similarity metric to the previous frame. The smoothing factor is (0.7–0.9). As the similarity threshold, when At that time, output the previous frame. The detection results are as follows: otherwise, super-resolution normalization is performed on the segmented ultrasonic image feature sub-regions, and waveform classification detection and corresponding confidence calculation are performed on the rigid body flaw detection waveform images of uniform size.

[0039] Furthermore, the data enhancement module specifically performs the following steps:

[0040] For each sub-region's feature block conduct Alignment to eliminate due to The quantization error caused by different sizes is as follows:

[0041] ;

[0042] in, For the first Alignment features of individual sub-regions This is a bilinear interpolation operation;

[0043] Align features Upsampling reconstruction is performed to recover higher resolution feature information, as follows:

[0044] ;

[0045] in, For the first Upsampling and reconstructing feature maps for each sub-region; It is a convolutional layer; For upsampling operators; Amplitude normalization is performed to eliminate noise caused by differences in characteristic amplitudes, as follows:

[0046] ;

[0047] in, It is a function with maximum value. Minimum value function For the first Sub-region amplitude normalization characteristics.

[0048] Furthermore, the flaw detection result detection module specifically performs the following steps:

[0049] (21) Using a classification model For the Sub-region amplitude normalization characteristics Perform binary classification to obtain the probabilities of positive and negative categories, as follows:

[0050] ;

[0051] ;

[0052] in, For the first The probability of each sub-region category Indicates the first The probability that each sub-region is a positive sample (defect-free) Indicates the first The probability that each sub-region is a negative sample (defective); for vector; This is an activation function used to... Mapped to a probability distribution;

[0053] (22) For each sub-region, the negative sample probability is taken as the defect confidence level, as follows:

[0054] ;

[0055] in, For the first The confidence that a sub-region is defective is defined as follows:

[0056] ;

[0057] in, Weights for each sub-region; The confidence level of the current frame being defective, if If the current frame is classified as defective, its corresponding confidence level is 1. Otherwise, the current frame detection and classification result is no defect, and its corresponding confidence level is 1. .

[0058] Furthermore, the result display module specifically performs the following steps:

[0059] The detection results are sent to the host computer via the Modbus protocol. The Modbus protocol data packet format is as follows:

[0060] ;

[0061] in, The entire data packet to be sent includes: the request address. The master station requests the slave station to perform the operation type Func, the current frame detection classification result and its corresponding confidence level encoding Data, and the cyclic redundancy check code CRC.

[0062] This invention also provides a rigid body ultrasonic flaw detection method based on image feature extraction. Based on the above system, the steps are as follows:

[0063] 1) Decode the real-time video stream into time-series frames and extract multi-scale features to segment the multi-angle regional rigid body flaw detection waveform images;

[0064] 2) Perform super-resolution normalization on the rigid body flaw detection waveform image after region segmentation to obtain a rigid body flaw detection waveform image of uniform size;

[0065] 3) Perform similarity detection on adjacent frames of multi-angle segmented rigid body flaw detection waveform images to determine whether the current frame uses the result of the previous frame;

[0066] 4) Perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels;

[0067] 5) Visualize the detection and classification results and their corresponding confidence levels.

[0068] The beneficial effects of this invention are:

[0069] 1. This invention targets the multi-angle partitioned structural features of rigid body flaw detection images. Through multi-partition block alignment, super-resolution enhancement, and parallel classification, it achieves high-sensitivity detection of local subtle waveform anomalies, and is particularly suitable for the identification of small-scale defects such as cracks, inclusions, and incomplete penetration.

[0070] 2. This invention utilizes inter-frame feature similarity calculation to achieve frame skipping detection, performing full-process calculations only on key changing frames, significantly reducing redundant computation, enabling the system to maintain a high frame rate and low latency even in real-time streaming, meeting the online flaw detection requirements of rigid body high-speed production lines;

[0071] 3. This invention obtains the overall flaw detection result and its corresponding confidence level by fusing the confidence scores of multi-zone classification results and combining them with regional weights. The result is then output to a PLC or host computer in real time via the Modbus protocol, enabling interconnected and low-cost industrial deployment. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the system of the present invention.

[0073] Figure 2 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0074] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0075] Reference Figure 1 As shown, the rigid body ultrasonic flaw detection system based on image feature extraction of the present invention includes: a data preprocessing module, a data enhancement module, a flaw detection result detection module and a result display module;

[0076] The data preprocessing module is used to decode the real-time video stream into time-series frames, extract multi-scale features, and segment the multi-angle, region-specific rigid body flaw detection waveform images; specifically, it performs the following steps:

[0077] (11) Collect real-time video stream data from the ultrasonic flaw detector at the industrial site and convert it into structured image data that can be processed by the neural network. The video stream is decoded into time series frames, as shown in the following expression:

[0078] ;

[0079] in, These are the decoded time-series frames. For the time series Frame ultrasound image, size is ; The total number of frames for ultrasound images Perform grayscale normalization to reduce noise and eliminate brightness drift, resulting in the grayscale normalized [value]. Frame ultrasound images The expression is as follows:

[0080] ;

[0081] in, For the first Frame ultrasound images The horizontal and vertical axes are and The pixel value of the location; Indicates the first Mean grayscale value of a frame of ultrasound image. ; Indicates the first The standard deviation of grayscale values ​​in a frame of ultrasound images. ; To traverse each The summation function, It is a small constant;

[0082] (12) From the preprocessed grayscale normalized ultrasound image Extract multi-scale features; use the backbone feature extraction model. right Multi-scale features are extracted as follows:

[0083] ;

[0084] in, For multi-layer feature maps, each Corresponding to a feature layer, The global response weights for each layer are calculated using a channel attention mechanism. These weights are then multiplied channel by channel by the original feature map of each layer to enhance channels containing defect variations and suppress noise channels unrelated to defects, as follows:

[0085] ;

[0086] in, To enhance the defect sensitivity of the multi-layer feature map, This is the weight matrix. For global average pooling, for Activation function; multilayer feature map after enhancing defect sensitivity Each layer of features is multiplied by the feature weight coefficient of that layer to perform multi-layer feature fusion. The fused feature map is as follows:

[0087] ;

[0088] in, This is the fused global feature map. These are the feature weight coefficients for each layer;

[0089] (13) The fused global feature map The image was divided into 2×4, or 8, feature sub-regions containing ultrasound images. Each sub-region corresponds to the detection angle result of an ultrasonic flaw detector. A waveform segmentation model was used. Global feature map containing 8 ultrasound image feature sub-regions The region is segmented to obtain feature blocks of 8 corresponding sub-regions. The partitioning mapping is defined as follows:

[0090] ;

[0091] in, For the first Feature maps of each sub-region , For sub-regions in the global feature map The height and width range in the text.

[0092] The data augmentation module performs super-resolution normalization on the rigid body flaw detection waveform images after region segmentation to obtain rigid body flaw detection waveform images of uniform size; specifically, it performs the following steps:

[0093] For each sub-region's feature block conduct Alignment to eliminate due to The quantization error caused by different sizes is as follows:

[0094] ;

[0095] in, For the first Alignment features of individual sub-regions This is a bilinear interpolation operation;

[0096] Align features Upsampling reconstruction is performed to recover higher resolution feature information, as follows:

[0097] ;

[0098] in, For the first Upsampling and reconstructing feature maps for each sub-region; It is a convolutional layer; For upsampling operators; Amplitude normalization is performed to eliminate noise caused by differences in characteristic amplitudes, as follows:

[0099] ;

[0100] in, It is a function with maximum value. Minimum value function For the first Sub-region amplitude normalization characteristics.

[0101] The flaw detection result detection module is used to perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels; the specific steps are as follows:

[0102] (21) Using a classification model For the Sub-region amplitude normalization characteristics Perform binary classification to obtain the probabilities of positive and negative categories, as follows:

[0103] ;

[0104] ;

[0105] in, For the first The probability of each sub-region category Indicates the first The probability that each sub-region is a positive sample (defect-free) Indicates the first The probability that each sub-region is a negative sample (defective); for vector; This is an activation function used to... Mapped to a probability distribution;

[0106] (22) For each sub-region, the negative sample probability is taken as the defect confidence level, as follows:

[0107] ;

[0108] in, For the first The confidence that a sub-region is defective is defined as follows:

[0109] ;

[0110] in, Weights for each sub-region; The confidence level of the current frame being defective, if If the current frame is classified as defective, its corresponding confidence level is 1. Otherwise, the current frame detection and classification result is no defect, and its corresponding confidence level is 1. .

[0111] The results display module is used to visualize the detection and classification results and their corresponding confidence scores; the specific steps are as follows:

[0112] The detection results are sent to the host computer via the Modbus protocol. The Modbus protocol data packet format is as follows:

[0113] ;

[0114] in, The entire data packet to be sent includes: the request address. The master station requests the slave station to perform the operation type Func, the current frame detection classification result and its corresponding confidence level encoding Data, and the cyclic redundancy check code CRC.

[0115] In a preferred embodiment, the system further includes an acceleration module for improving detection speed through inter-frame similarity detection, specifically including:

[0116] Calculate the current time Compared to the previous moment global feature map If the similarity is greater than a certain threshold, the previous time step is used directly. The detection and classification results are used as the current moment The result, similarity is defined as:

[0117] ;

[0118] in, For inter-frame similarity; , These are the global feature maps for the current frame and the previous frame, respectively. This represents the dot product operation. This represents taking the modulus after expanding the vector; smoothing the similarity of consecutive frames ensures that waveform changes are judged based on the cumulative trend within the sliding window, thus avoiding misjudgments caused by single-frame noise. The expression is as follows:

[0119] ;

[0120] in, The similarity index is after smoothing. This is a similarity metric to the previous frame. The smoothing factor is (0.7–0.9). As the similarity threshold, when At that time, output the previous frame. The detection results are as follows: otherwise, super-resolution normalization is performed on the segmented ultrasonic image feature sub-regions, and waveform classification detection and corresponding confidence calculation are performed on the rigid body flaw detection waveform images of uniform size.

[0121] Reference Figure 2 As shown, the present invention also provides a rigid body ultrasonic flaw detection method based on image feature extraction. Based on the above system, the steps are as follows:

[0122] 1) Decode the real-time video stream into time-series frames and extract multi-scale features to segment the multi-angle regional rigid body flaw detection waveform images;

[0123] 2) Perform super-resolution normalization on the rigid body flaw detection waveform image after region segmentation to obtain a rigid body flaw detection waveform image of uniform size;

[0124] 3) Perform similarity detection on adjacent frames of multi-angle segmented rigid body flaw detection waveform images to determine whether the current frame uses the result of the previous frame;

[0125] 4) Perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels;

[0126] 5) Visualize the detection and classification results and their corresponding confidence levels.

[0127] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A rigid body ultrasonic flaw detection system based on image feature extraction, characterized in that, include: The module includes a data preprocessing module, a data augmentation module, a flaw detection result detection module, and a result display module. The data preprocessing module is used to decode the real-time video stream into time series frames, extract multi-scale features, and segment the multi-angle regional rigid body flaw detection waveform images. The data augmentation module is used to perform super-resolution normalization on the rigid body flaw detection waveform image after region segmentation to obtain a rigid body flaw detection waveform image of uniform size. The flaw detection result detection module is used to perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels; The results display module is used to visualize the detection and classification results and their corresponding confidence levels.

2. The rigid body ultrasonic flaw detection system based on image feature extraction according to claim 1, characterized in that, The data preprocessing module specifically performs the following steps: (11) Collect real-time video stream data from the ultrasonic flaw detector at the industrial site and convert it into structured image data that can be processed by the neural network. The video stream is decoded into time series frames, as shown in the following expression: ； in, These are the decoded time-series frames. For the time series Frame ultrasound image, size is ; The total number of frames for ultrasound images Perform grayscale normalization to reduce noise and eliminate brightness drift, resulting in the grayscale normalized [value]. Frame ultrasound images The expression is as follows: ； in, For the first Frame ultrasound images The horizontal and vertical axes are and The pixel value of the location; Indicates the first Mean grayscale value of a frame of ultrasound image. ; Indicates the first The standard deviation of grayscale values ​​in a frame of ultrasound images. ; To traverse each The summation function, It is a small constant; (12) From the preprocessed grayscale normalized ultrasound image Extract multi-scale features; use the backbone feature extraction model. right Multi-scale features are extracted as follows: ； in, For multi-layer feature maps, each Corresponding to a feature layer, The global response weights for each layer are calculated using a channel attention mechanism. These weights are then multiplied channel by channel by the original feature map of each layer to enhance channels containing defect variations and suppress noise channels unrelated to defects, as follows: ； in, To enhance the defect sensitivity of the multi-layer feature map, This is the weight matrix. For global average pooling, for Activation function; multilayer feature map after enhancing defect sensitivity Each layer of features is multiplied by the feature weight coefficient of that layer to perform multi-layer feature fusion. The fused feature map is as follows: ； in, This is the fused global feature map. These are the feature weight coefficients for each layer; (13) The fused global feature map The image was divided into 2×4, or 8, feature sub-regions containing the ultrasound image. Each sub-region corresponds to the detection angle result of an ultrasonic flaw detector. A waveform segmentation model was used. Global feature map containing 8 ultrasound image feature sub-regions The region is segmented to obtain feature blocks of 8 corresponding sub-regions. The partitioning mapping is defined as follows: ； in, For the first Feature maps of each sub-region , For sub-regions in the global feature map The height and width range in the text.

3. The rigid body ultrasonic flaw detection system based on image feature extraction according to claim 2, characterized in that, The data enhancement module specifically performs the following steps: For each sub-region's feature block conduct Alignment to eliminate due to The quantization error caused by different sizes is as follows: ； in, For the first Alignment features of individual sub-regions This is a bilinear interpolation operation; Align features Upsampling reconstruction is performed to recover higher resolution feature information, as follows: ； in, For the first Upsampling and reconstructing feature maps for each sub-region; It is a convolutional layer; For upsampling operators; Amplitude normalization is performed to eliminate noise caused by differences in characteristic amplitudes, as follows: ； in, It is a function with maximum value. Minimum value function For the first Sub-region amplitude normalization characteristics.

4. The rigid body ultrasonic flaw detection system based on image feature extraction according to claim 3, characterized in that, The flaw detection result detection module specifically performs the following steps: (21) Using a classification model For the Sub-region amplitude normalization characteristics Perform binary classification to obtain the probabilities of positive and negative categories, as follows: ； ； in, For the first The probability of each sub-region category Indicates the first The probability that each sub-region is a positive sample. Indicates the first The probability that each sub-region is a negative sample; for vector; This is an activation function used to... Mapped to a probability distribution; (22) For each sub-region, the negative sample probability is taken as the defect confidence level, as follows: ； in, For the first The confidence that a sub-region is defective is defined as follows: ； in, Weights for each sub-region; The confidence level of the current frame being defective, if If the current frame is classified as defective, its corresponding confidence level is 1. Otherwise, the current frame detection and classification result is no defect, and its corresponding confidence level is 1. .

5. The rigid body ultrasonic flaw detection system based on image feature extraction according to claim 4, characterized in that, The results display module performs the following steps: The detection results are sent to the host computer via the Modbus protocol. The Modbus protocol data packet format is as follows: ； in, The entire data packet to be sent includes: the request address. The master station requests the slave station to perform the operation type Func, the current frame detection classification result and its corresponding confidence level encoding Data, and the cyclic redundancy check code CRC.

6. The rigid body ultrasonic flaw detection system based on image feature extraction according to claim 1, characterized in that, The system also includes an acceleration module for improving detection speed through inter-frame similarity detection, specifically including: Calculate the current time Compared to the previous moment global feature map If the similarity is greater than a certain threshold, the previous time step is used directly. The detection and classification results are used as the current moment The result, similarity is defined as: ； in, For inter-frame similarity; , These are the global feature maps for the current frame and the previous frame, respectively. This represents the dot product operation. This represents taking the modulus after expanding the vector; smoothing the similarity of consecutive frames ensures that waveform changes are judged based on the cumulative trend within the sliding window, thus avoiding misjudgments caused by single-frame noise. The expression is as follows: ； in, The similarity index is after smoothing. This is a similarity metric to the previous frame. As a smoothing factor, As the similarity threshold, when At that time, output the previous frame. The detection results are as follows: otherwise, super-resolution normalization is performed on the segmented ultrasonic image feature sub-regions, and waveform classification detection and corresponding confidence calculation are performed on the rigid body flaw detection waveform images of uniform size.

7. A rigid body ultrasonic flaw detection method based on image feature extraction, based on the system described in claims 1-6, characterized in that, The steps are as follows: 1) Decode the real-time video stream into time-series frames and extract multi-scale features to segment the multi-angle regional rigid body flaw detection waveform images; 2) Perform super-resolution normalization on the rigid body flaw detection waveform image after region segmentation to obtain a rigid body flaw detection waveform image of uniform size; 3) Perform similarity detection on adjacent frames of multi-angle segmented rigid body flaw detection waveform images to determine whether the current frame uses the result of the previous frame; 4) Perform waveform detection and classification on rigid body flaw detection waveform images of uniform size to obtain the detection and classification results and their corresponding confidence levels; 5) Visualize the detection and classification results and their corresponding confidence levels.

Images