A method for identifying hull defects

Abstract

The invention discloses a hull defect recognition method, which relates to the technical field of ships. The method comprises the following steps: heating the hull through external electromagnetic excitation to obtain an infrared image and a corresponding infrared heat distribution field image; performing wavelet decomposition on the infrared image, enhancing a low-frequency sub-image through a nonlinear transformation and a guided filtering algorithm with improved weights, enhancing a high-frequency sub-image through Gaussian filtering and a second-order differential operator, and performing wavelet reconstruction on the enhanced low-frequency sub-image and the high-frequency sub-image to obtain a static information image; performing frame-by-frame difference on the temperature value of the infrared heat field distribution image, establishing a temperature difference matrix through vectorization operation, obtaining principal value eigenvalues ​​through corresponding covariance matrix calculation, and obtaining a dynamic information image through grayscale quantization; performing defect detection using a trained recognition model, performing feature fusion and detection on the static information image and the dynamic information image through a feature extraction module, a feature fusion module, and a detection module to complete defect recognition.

Description

Technical Field

[0001] The present invention relates to the field of ship technology, and in particular to a method for identifying hull defects. Background Art

[0002] In recent years, the shipping industry has become increasingly important for China's economic development. China is not only the world's second largest economy, but also the world's largest cargo trading country. Among the large amount of imported and exported goods, more than 95% need to be transported by ships. China's shipping routes and trade networks have spread across major countries and regions in the world, and indicators such as port scale, number of crew members, and size of shipping fleets are among the best in the world. At the same time, in addition to shipping, China's inland shipping has also made great progress. Fixed asset investment in inland waterways continues to grow, and the throughput of various types of goods and containers in inland ports has shown a steady growth trend. The booming development of the shipping industry is inseparable from ships in good condition, but due to the complex manufacturing process and special working environment of ships, hull defects are a common problem. Therefore, efficient and accurate hull defect identification methods are crucial to ship safety. Infrared thermal imaging analysis is a commonly used hull defect identification method, which uses the principle that the infrared radiation uniformity of the defective part and the intact part of the hull is different for defect identification. Not only can it be non-destructive, but it also has many advantages such as low operating difficulty and large detection area.

[0003] At present, the Chinese invention patent with application number CN202211194557.4 discloses an infrared detection and recognition system and method. The recognition system includes: a sensing and computing integrated infrared detection module, which is used to perform convolution operations on the infrared image to be recognized according to the dynamically configured modulation weights to obtain the feature information of the infrared image. The modulation weights are obtained by training the pre-set deep neural network model using the pre-collected training sample set; an image processing module, which is used to classify and recognize the infrared image based on the trained deep neural network model according to the feature information to obtain the semantic information of the infrared image; a semantic interpretation processing conversion module, which is used to convert the semantic information into information in a preset form and transmit the information in the preset form to the user. The system improves the efficiency of convolution operations and can be applied to applications such as infrared image acquisition, target detection, classification and recognition that require long-term online operation. However, directly collecting infrared images through infrared imagers is not only susceptible to environmental thermal radiation, but also has the problems of low contrast and high noise in the infrared images themselves. There is a large error when extracting features from the original infrared image, and when using traditional infrared thermal imaging methods for detection, it is often necessary to use an exhaustive method to adjust the temperature of each frame of the image. The calculation time complexity is high, resulting in slow infrared image processing speed and insufficient defect recognition accuracy. Summary of the invention

[0004] The technical problem solved by the present invention is that when the traditional infrared thermal imaging method is used to identify hull defects, there are problems such as slow infrared image processing speed and low recognition degree of defective parts in the image.

[0005] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0006] A method for identifying hull defects comprises the following steps:

[0007] Step 1, heating the hull part to be inspected, and after the heating stops, collecting an infrared image of the heated area and a corresponding infrared thermal field distribution image through an infrared imager;

[0008] Step 2, selecting the first frame of infrared image obtained after stopping heating, performing wavelet decomposition on the first frame of infrared image to obtain a low-frequency sub-image and a high-frequency sub-image, enhancing the low-frequency sub-image and the high-frequency sub-image respectively, and obtaining a static information image through wavelet reconstruction;

[0009] Step 3, subtract the temperature values ​​of the pixel points of the infrared thermal field distribution image frame by frame, establish the infrared thermal field distribution temperature difference matrix through vectorization operation, calculate the eigenvalues ​​and eigenvectors of the corresponding covariance matrix, select the main value eigenvalue by calculating the influence coefficient, and perform grayscale quantization on the eigenvectors corresponding to the main value eigenvalues ​​to obtain a dynamic information image;

[0010] Step 4, using the trained recognition model to perform defect detection, and performing image fusion and detection on the static information image and the dynamic information image through the feature extraction module, the feature fusion module, and the detection module to complete defect recognition.

[0011] As a preferred embodiment of a method for identifying hull defects described in the present invention, an induction coil with high-frequency current is placed close to the hull part to be detected. The hull is usually made of alloys such as carbon steel. Therefore, when the induction coil is placed close, the hull will rapidly generate an induced current with a reciprocating direction due to the electromagnetic effect. The generated induced current will overcome the resistance and be converted into heat energy when it is conducted in the hull, so that the temperature of the hull to be tested will rise rapidly. The heating time is 5 to 10 seconds, which can take into account the hull heating effect and the detection rate. After stopping the heating, turn on the infrared imager that has been fixed in place in advance to record the video, and collect the infrared image of the heating area and the corresponding infrared thermal field distribution image. Directly obtaining infrared images at room temperature is susceptible to environmental interference and the boundary clarity of the defective part is low, making it difficult to identify. The method of electromagnetic excitation heating can make full use of the difference in thermal conductivity between the defective part and the intact part of the hull, which is conducive to highlighting the defective part and the boundary in infrared imaging.

[0012] As a preferred solution of the ship hull defect identification method described in the present invention, the first frame of infrared image is subjected to wavelet transform by a two-dimensional fast wavelet transform filter, and the first frame of infrared image is decomposed into high-frequency components and low-frequency components according to the frequency, so as to obtain high-frequency sub-images and low-frequency sub-images. When decomposing the infrared image by wavelet transform, no information is lost or redundant information is generated, and the image can still be perfectly reconstructed after the high-frequency sub-image and the low-frequency sub-image are enhanced respectively.

[0013] As a preferred solution of the ship hull defect identification method described in the present invention, the texture feature of the low-frequency sub-image is enhanced, and the low-frequency sub-image is nonlinearly gray-mapped by the Ln function to obtain a gray-mapped image, the calculation expression of which is:

[0014]

[0015] in, Represents the horizontal coordinate of the pixel point, Represents the horizontal coordinate of the pixel point, represents the gray value of the low-frequency sub-image, represents the gray value of the low-frequency sub-image after nonlinear transformation, Represents the mapping coefficient. The low-frequency sub-image retains most of the detail information of the original first frame of infrared image. The low-frequency sub-image is transformed nonlinearly using the logarithmic function, which can expand the low gray value while compressing the high gray value relatively, maintaining the image brightness balance while highlighting the image details.

[0016] As a preferred solution of a hull defect identification method described in the present invention, the low-frequency sub-image after grayscale mapping is further enhanced, the guided filtering algorithm is optimized by combining the improved weights of the Sobel operator, the gradient image is obtained by performing a derivative operation on the grayscale mapping image, the gradient image and the grayscale mapping image are linearly superimposed to obtain the guided image, a linear regression equation of the guided image and a corresponding cost function are established, the regularization parameter in the cost function is replaced by the improved weights combined with the Sobel operator, the coefficients of the linear regression equation of the guided image are calculated by the cost function, and the coefficients are brought back to the equation to obtain a low-frequency sub-image with enhanced texture features, and the calculation expression is:

[0017] =

[0018]

[0019]

[0020] in, represents the gradient corresponding to the low-frequency sub-image, represents partial derivative, I represents the guide image, represents the superposition coefficient, represents the regularization parameter, represents the improved weight, represents the sum of the pixels in the guided filter window, i represents the low-frequency sub-image pixel, represents the pixel point of the guidance image, represents the Sobel operator operation, represents the variance operation, The guided filtering algorithm has low computational time complexity, can quickly process a large amount of hull infrared image data, and fully retains the spatial structure information of the original image, which is conducive to subsequent image reconstruction.

[0021] As a preferred solution of the hull defect identification method described in the present invention, the high-frequency sub-image is enhanced, and the high-frequency sub-image obtained by wavelet decomposition contains a lot of noise. The high-frequency sub-image is smoothed by Gaussian filtering to reduce image noise, and then the edge of the smoothed image is sharpened by a second-order differential operator to highlight the contour and spots of the image. The calculation expression is:

[0022]

[0023] in, represents the sharpening operator, represents the standard deviation, Represents Gaussian filter parameters. Hull defects are often manifested as sudden contours and spots in infrared images. By enhancing the contours and spots in high-frequency sub-images, it is helpful to identify hull defects in infrared images.

[0024] As a preferred solution of the ship hull defect identification method described in the present invention, the enhanced high-frequency sub-image and the low-frequency sub-image are subjected to a two-dimensional fast wavelet inverse transform, and a static information image is obtained through wavelet reconstruction.

[0025] As a preferred solution of a hull defect identification method described in the present invention, N frames of infrared thermal field distribution images are selected after heating is stopped, and the temperature values ​​of the pixel points corresponding to the first frame of infrared thermal field distribution image and the next N-1 frame of infrared thermal field distribution image are subtracted one by one to obtain N-1 infrared thermal field distribution temperature difference vectors, and an infrared thermal field distribution temperature difference matrix is ​​constructed, and its calculation expression is:

[0026] , N

[0027]

[0028] in, represents the temperature difference vector between the first frame of infrared thermal field distribution image and the u-th frame of infrared thermal field distribution image, represents the temperature difference matrix of infrared thermal field distribution, It represents the temperature difference between the pixel with coordinates (x, y) in the first frame of infrared thermal field distribution image and the uth frame of infrared thermal field distribution image, p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image, and p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image. By constructing the infrared thermal field distribution temperature difference matrix, it avoids adjusting the temperature of each frame of the image by exhaustive method.

[0029] As a preferred solution of a hull defect identification method described in the present invention, a covariance matrix corresponding to the temperature difference matrix of the infrared thermal field distribution is established, and N-1 eigenvalues ​​of the covariance matrix and corresponding N-1 eigenvectors are calculated. The influence coefficient is calculated by the eigenvalue of the covariance matrix, and the influence coefficient is sorted from high to low and accumulated one by one. When the accumulated sum of the current k influence coefficients reaches the influence threshold, the accumulation is stopped, wherein the influence threshold is usually taken as 0.9, and the eigenvalues ​​corresponding to the first k influence coefficients are selected as the main values, and the calculation expression is:

[0030]

[0031]

[0032] in, represents the covariance matrix, represents the mathematical expectation of the temperature difference matrix of infrared thermal field distribution, T represents the matrix transpose, represents the kth influence coefficient, represents the i-th covariance matrix eigenvalue arranged from large to small, represents the i-th covariance matrix eigenvalue arranged from large to small, p represents the maximum horizontal coordinate of the infrared thermal field distribution image pixel, and p represents the maximum horizontal coordinate of the infrared thermal field distribution image pixel. By establishing the temperature difference matrix, the N-frame image information can be fully utilized, and the covariance matrix and principal value selection can effectively reflect the dynamic temperature change characteristics of the hull detection area, while effectively reducing error interference.

[0033] As a preferred scheme of the hull defect identification method described in the present invention, the infrared thermal field feature vector is obtained by taking the average value of the feature vectors corresponding to the k principal values, the infrared thermal field feature vector is gray-quantized, and the infrared thermal field feature vector is represented as a matrix by establishing the inverse operation of the infrared thermal field distribution temperature difference matrix. The obtained matrix element positions correspond to the pixel point positions of the original infrared thermal field distribution map, and the matrix element values ​​are linearly mapped to grayscale values ​​from 0 to 255 to obtain a dynamic information image.

[0034] As a preferred solution of the hull defect recognition method described in the present invention, the static information image and the dynamic information image are fused and detected using a trained recognition model, and the recognition model is composed of a feature extraction module, a feature fusion module and a detection module, specifically including:

[0035] The feature extraction module is composed of two identical branches connected in parallel, each branch includes four convolutional layers, and the PReLU function is selected as the activation function of each convolutional layer. Multi-scale feature extraction is performed on the static information image and the dynamic information image through dense connection and residual connection to obtain a static information feature map and a dynamic information feature map;

[0036] The feature fusion module performs feature fusion through a dual attention map, multiplies the static information feature map and the dynamic information feature map element by element to obtain a joint feature map, performs multi-scale convolution and channel splicing on the joint feature map with the static information feature map and the dynamic information feature map respectively, performs normalization processing through a Logistic function to obtain a static information attention map and a dynamic information attention map, and adds the static information attention map and the dynamic information attention map element by element to obtain a fused feature map;

[0037] The detection module is composed of a region proposal network, a region of interest pooling layer, and a classification regression network. The region proposal network generates a candidate frame at the defect position on the fused feature map, and the region of interest pooling layer maps the candidate frame on the fused feature map to obtain the region of interest. The region of interest is segmented and pooled to obtain a standard feature map of a fixed size. The standard feature map is classified and predicted and the bounding box regression parameters are predicted by the fully connected layer in the classification regression network to complete the identification and positioning of the defective part. The recognition model based on the convolutional neural network can extract image features at multiple scales, and the dual attention map performs feature fusion to fully utilize the image feature information and improve the feature fusion level. The region proposal network, the region of interest pooling layer, and the classification regression network improve the accuracy and real-time performance of defect detection.

[0038] As a preferred solution of a hull defect recognition method described in the present invention, wherein: the recognition model has been trained through an image data set after being constructed, including: heating the defective hull part, collecting 120 infrared images and infrared thermal field distribution images of the defective part, enhancing the infrared image by wavelet transform to obtain the static information image, and enhancing the infrared thermal field distribution image by establishing a temperature difference matrix to obtain the dynamic information image. The defective part in the image is manually box-selected and marked to obtain a data set consisting of 240 files containing rectangular box annotations. The files of this data set are encoded in utf-8 format with a resolution of 640×480, covering 6 common types of hull defects, including: cracks, corrosion, wrinkles, dents, holes and desoldering. The data set is horizontally flipped, vertically flipped, and horizontally and vertically flipped respectively, and the data set is expanded to 4 times the original size. The expanded data set contains 960 pictures, which are divided into a training set and a validation set according to a ratio of 8:2, wherein the training set contains 768 pictures and the validation set contains 192 pictures. The expanded data set is used for training and verification to obtain a trained recognition model.

[0039] The beneficial effects of the present invention are as follows: before collecting images with an infrared imager, the hull is first heated by external electromagnetic excitation, which is beneficial to amplifying the difference in thermal conductivity between the defective part and the intact part of the hull; the infrared image is subjected to wavelet decomposition to obtain a low-frequency sub-image and a high-frequency sub-image, which are enhanced respectively and then subjected to wavelet reconstruction to obtain a static information image, which is beneficial to highlighting the contrast of the infrared image and enhancing the contour and texture features of the image; the infrared thermal field distribution image is subjected to difference and vectorization operations to establish a temperature difference matrix, the principal value eigenvalue is selected by covariance matrix calculation, and a dynamic information image is obtained by grayscale quantization, which is beneficial to identifying defects from the perspective of temperature changes, reducing the computational complexity by matrix operations, and effectively highlighting the abnormal temperature change points; a recognition model based on a convolutional neural network is established, which can extract image features of static information images and dynamic information images at multiple scales, and then fully utilize the extracted feature information through a dual attention map, and finally efficiently complete defect recognition through detection module selection. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of a basic framework provided for an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a flow chart provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0042] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0043] Example 1, reference Figure 1 and Figure 2 , as an embodiment of the present invention, provides a hull defect identification method, comprising:

[0044] Step 1: Heat the hull part to be inspected, and after the heating stops, collect the infrared image of the heated area and the corresponding infrared thermal field distribution image through the infrared imager;

[0045] Step 2: Select the first frame of infrared image obtained after stopping heating, perform wavelet decomposition on the first frame of infrared image to obtain low-frequency sub-image and high-frequency sub-image, enhance the low-frequency sub-image and high-frequency sub-image respectively, and obtain a static information image through wavelet reconstruction.

[0046] Step 3: Subtract the temperature values ​​of the pixels of the infrared thermal field distribution image frame by frame, establish the infrared thermal field distribution temperature difference matrix through vectorization operation, calculate the eigenvalues ​​and eigenvectors of the corresponding covariance matrix, select the main value eigenvalue by calculating the influence coefficient, and perform grayscale quantization on the eigenvectors corresponding to the main value eigenvalues ​​to obtain a dynamic information image;

[0047] Step 4: Use the trained recognition model to perform defect detection, and perform image fusion and detection on the static information image and the dynamic information image through the feature extraction module, feature fusion module, and detection module to complete defect recognition.

[0048] In this embodiment, the defective hull portion is heated by external electromagnetic excitation, and after the heating is stopped, the heated area is recorded by an infrared imager at a fixed position to obtain an infrared image and an infrared thermal field distribution image.

[0049] In this embodiment, the first frame of infrared image obtained after stopping heating is selected, and wavelet transform is performed on the first frame of infrared image through a two-dimensional fast wavelet transform filter to decompose the first frame of infrared image into high-frequency components and low-frequency components to obtain high-frequency sub-images and low-frequency sub-images.

[0050] In this embodiment, a nonlinear transformation is performed on the low-frequency sub-image, and a nonlinear grayscale mapping is performed on the low-frequency sub-image through an ln function to obtain a grayscale extended image, and its calculation expression is:

[0051]

[0052] in, Represents the horizontal coordinate of the pixel point, Represents the horizontal coordinate of the pixel point, represents the gray value of the low-frequency sub-image, represents the gray value of the low-frequency sub-image after nonlinear transformation, Represents the mapping coefficient.

[0053] In this embodiment, a gradient image is obtained by performing a derivative operation on the grayscale extension image, and a guide image is obtained by linearly superimposing the gradient image and the grayscale mapping image. The guided filtering algorithm is optimized by combining the improved weights of the Sobel operator to obtain a low-frequency sub-image with enhanced texture features. The calculation expression is:

[0054] =

[0055]

[0056]

[0057] in, represents the gradient corresponding to the low-frequency sub-image, represents partial derivative, I represents the guide image, represents the superposition coefficient, represents the regularization parameter, represents the improved weight, represents the sum of the pixels in the filter window, i represents the low-frequency sub-image pixel, represents the pixel point of the guidance image, represents the Sobel operator operation, represents the variance operation, Indicates compensation parameters.

[0058] In this embodiment, the high-frequency sub-image is smoothed by Gaussian filtering, and the edge of the smoothed image is sharpened by a second-order differential operator to obtain a high-frequency sub-image with prominent edge features, and its calculation expression is:

[0059]

[0060] in, represents the sharpening operator, represents the standard deviation, Represents Gaussian filter parameters.

[0061] In this embodiment, N frames of infrared thermal field distribution images are selected after heating is stopped, and the temperature values ​​of the pixels corresponding to the first frame of infrared thermal field distribution image and the next N-1 frame of infrared thermal field distribution image are subtracted one by one to obtain N-1 infrared thermal field distribution temperature difference vectors, and an infrared thermal field distribution temperature difference matrix is ​​constructed, and its calculation expression is:

[0062] , N

[0063]

[0064] in, represents the temperature difference vector between the first frame of infrared thermal field distribution image and the u-th frame of infrared thermal field distribution image, represents the temperature difference matrix of infrared thermal field distribution, It represents the temperature difference between the pixel with coordinates (x, y) in the first frame of infrared thermal field distribution image and the u-th frame of infrared thermal field distribution image, p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image, and p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image.

[0065] In this embodiment, a covariance matrix is ​​constructed and eigenvalues ​​and eigenvectors are calculated. The influence coefficient is calculated by the eigenvalue of the covariance matrix. The influence coefficients are sorted from high to low and accumulated one by one. The accumulation is stopped when the sum of the current k influence coefficients reaches the influence threshold. The eigenvalues ​​corresponding to the first k influence coefficients are selected as the main values. The calculation expression is:

[0066]

[0067]

[0068] in, represents the covariance matrix, represents the mathematical expectation of the temperature difference matrix of infrared thermal field distribution, T represents the matrix transpose, represents the kth influence coefficient, represents the i-th covariance matrix eigenvalue arranged from large to small, represents the i-th covariance matrix eigenvalue arranged from large to small, p represents the maximum horizontal coordinate of the pixel point of the infrared thermal field distribution image, and p represents the maximum horizontal coordinate of the pixel point of the infrared thermal field distribution image.

[0069] In this embodiment, the feature vectors corresponding to the k principal values ​​are averaged to obtain the infrared thermal field feature vectors, and the infrared thermal field feature vectors are gray-quantized to obtain the dynamic information image.

[0070] In this embodiment, the trained recognition model is used to perform image fusion and detection on the static information image and the dynamic information image. The recognition model is composed of a feature extraction module, a feature fusion module and a detection module, and specifically includes:

[0071] The feature extraction module is composed of two identical branches connected in parallel, each branch includes four convolutional layers, and the PReLU function is selected as the activation function. Multi-scale feature extraction is performed on the static information image and the dynamic information image through dense connection and residual connection to obtain a static information feature map and a dynamic information feature map;

[0072] The feature fusion module performs feature fusion through a dual attention map, multiplies the static information feature map and the dynamic information feature map element by element to obtain a joint feature map, performs multi-scale convolution and channel splicing on the joint feature map with the static information feature map and the dynamic information feature map respectively, performs normalization processing through a Logistic function to obtain a static information attention map and a dynamic information attention map, and adds the static information attention map and the dynamic information attention map element by element to obtain a fused feature map;

[0073] The detection module consists of a region proposal network, a region of interest pooling layer and a classification regression network. The region proposal network generates a candidate box at the defect position on the fused feature map, and the region of interest pooling layer maps the candidate box to the fused feature map to obtain the region of interest. The region of interest is segmented and pooled to obtain a standard feature map of a fixed size. The fully connected layer in the classification regression network performs classification prediction and bounding box regression parameter prediction on the standard feature map to complete the identification and positioning of the defective part.

[0074] It should be understood by those skilled in the art that the embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program codes. Among them, the storage medium may be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic memory, flash memory, magnetic disk or optical disk. These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0075] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for identifying hull defects, characterized in that: The following steps are involved: Step 1, heating the hull part to be inspected, and after the heating stops, collecting an infrared image of the heated area and a corresponding infrared thermal field distribution image through an infrared imager; Step 2, selecting the first frame of infrared image obtained after stopping heating, performing wavelet decomposition on the first frame of infrared image to obtain a low-frequency sub-image and a high-frequency sub-image, enhancing the low-frequency sub-image and the high-frequency sub-image respectively, and obtaining a static information image through wavelet reconstruction; Step 3, subtract the temperature values ​​of the pixel points of the infrared thermal field distribution image frame by frame, establish the infrared thermal field distribution temperature difference matrix through vectorization operation, calculate the eigenvalues ​​and eigenvectors of the corresponding covariance matrix, select the main value eigenvalue by sorting the influence coefficients, and perform grayscale quantization on the eigenvectors corresponding to the main value eigenvalues ​​to obtain a dynamic information image; Step 4, using the trained recognition model to perform defect detection, and performing image fusion and detection on the static information image and the dynamic information image through the feature extraction module, the feature fusion module and the detection module to complete defect recognition; The low-frequency sub-image is subjected to nonlinear transformation, and nonlinear grayscale mapping is performed on the low-frequency sub-image through an ln function to obtain a grayscale expansion image, the calculation expression of which is: ； in, Represents the horizontal coordinate of the pixel point, Represents the horizontal coordinate of the pixel point, represents the gray value of the low-frequency sub-image, represents the gray value of the low-frequency sub-image after nonlinear transformation, represents the mapping coefficient; The high-frequency sub-image is smoothed by Gaussian filtering, and the edge of the smoothed image is sharpened by a second-order differential operator to obtain a high-frequency sub-image with prominent edge features, and its calculation expression is: ； in, represents the sharpening operator, represents the standard deviation, represents Gaussian filter parameters, e represents a natural constant; Select N frames of infrared thermal field distribution images after stopping heating, and make a difference one by one between the temperature values ​​of the pixels corresponding to the first frame of infrared thermal field distribution image and the next N-1 frame of infrared thermal field distribution image to obtain N-1 infrared thermal field distribution temperature difference vectors, and construct an infrared thermal field distribution temperature difference matrix, whose calculation expression is: ， ； ； in, represents the temperature difference vector between the first frame of infrared thermal field distribution image and the u-th frame of infrared thermal field distribution image, represents the temperature difference matrix of infrared thermal field distribution, It represents the temperature difference between the pixel with coordinates (x, y) in the first frame of infrared thermal field distribution image and the u-th frame of infrared thermal field distribution image, p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image, p represents the maximum horizontal coordinate of the pixel in the infrared thermal field distribution image, and N represents the total number of frames of infrared thermal field distribution image.

2. A method for identifying hull defects according to claim 1, characterized in that: The defective hull part is heated by external electromagnetic excitation. After the heating is stopped, the heated area is recorded by an infrared imager at a fixed position to obtain an infrared image and an infrared thermal field distribution image.

3. A method for identifying hull defects according to claim 1, characterized in that: The first infrared image frame obtained after stopping heating is selected, and the first infrared image frame is subjected to wavelet transform through a two-dimensional fast wavelet transform filter, so as to decompose the first infrared image frame into a high-frequency component and a low-frequency component to obtain a high-frequency sub-image and a low-frequency sub-image.

4. A method for identifying hull defects according to claim 1, characterized in that: The gradient image is obtained by performing derivative calculation on the grayscale extension image, and the guide image is obtained by linearly superimposing the gradient image and the grayscale mapping image. The low-frequency sub-image with enhanced texture features is calculated by combining the improved weighted guide filtering algorithm with the Sobel operator. The calculation expression is: = ； ； ； in, represents the gradient corresponding to the low-frequency sub-image, represents partial derivative, I represents the guide image, represents the superposition coefficient, represents the regularization parameter, represents the improved weight, represents the sum of the pixels in the filter window, i represents the low-frequency sub-image pixel, represents the pixel point of the guidance image, represents the Sobel operator operation, represents the variance operation, Indicates compensation parameters.

5. A method for identifying hull defects according to claim 1, characterized in that: The covariance matrix corresponding to the temperature difference matrix of the infrared thermal field distribution is constructed, and the influence coefficient is calculated by the eigenvalue of the covariance matrix. The influence coefficients are sorted from high to low and accumulated one by one. The accumulation is stopped when the cumulative sum of the current k influence coefficients reaches the influence threshold. The eigenvalues ​​corresponding to the first k influence coefficients are selected as the main values. The calculation expression is: ； ； in, represents the covariance matrix, represents the mathematical expectation of the temperature difference matrix of infrared thermal field distribution, T represents the matrix transpose, represents the kth influence coefficient, represents the i-th covariance matrix eigenvalue arranged from large to small, represents the i-th covariance matrix eigenvalue arranged from large to small, p represents the maximum horizontal coordinate of the pixel point of the infrared thermal field distribution image, and p represents the maximum horizontal coordinate of the pixel point of the infrared thermal field distribution image.

6. A method for identifying hull defects according to claim 5, characterized in that: The characteristic vectors corresponding to the k principal values ​​are averaged to obtain an infrared thermal field characteristic vector, and the infrared thermal field characteristic vector is gray-quantized to obtain a dynamic information image.

7. A method for identifying hull defects according to claim 1, characterized in that: Perform image fusion and detection on the static information image and the dynamic information image using a trained recognition model, wherein the recognition model includes a feature extraction module, a feature fusion module and a detection module; The feature extraction module is composed of two identical branches connected in parallel, each branch includes four convolutional layers, and the PReLU function is selected as the activation function. Multi-scale feature extraction is performed on the static information image and the dynamic information image through dense connection and residual connection to obtain a static information feature map and a dynamic information feature map; The feature fusion module performs feature fusion through a dual attention map, multiplies the static information feature map and the dynamic information feature map element by element to obtain a joint feature map, performs multi-scale convolution and channel splicing on the joint feature map with the static information feature map and the dynamic information feature map respectively, performs normalization processing through a Logistic function to obtain a static information attention map and a dynamic information attention map, and adds the static information attention map and the dynamic information attention map element by element to obtain a fused feature map; The detection module includes a region proposal network, a region of interest pooling layer and a classification regression network. The region proposal network generates a candidate box at the defect position on the fused feature map, and the region of interest pooling layer maps the candidate box on the fused feature map to obtain the region of interest. The region of interest is segmented and pooled to obtain a standard feature map of a fixed size. The fully connected layer in the classification regression network performs classification prediction and bounding box regression parameter prediction on the standard feature map to complete the identification and positioning of the defective part.

Images