Battery pack panoramic image splicing method and system for new energy automobile transportation cabin

Abstract

The invention discloses a new energy automobile transportation cabin-oriented battery pack panoramic image splicing method and system. The method comprises the following steps of collecting a visible light image and a thermal imaging image of a new energy automobile power battery pack; image fusion is carried out after image preprocessing; the high and low frequency adaptive enhanced super-resolution network performs super-resolution processing on the fused image; performing pairwise feature matching on the images after super-resolution processing, and screening matching results to obtain a feature matching point set; performing random sampling on the feature matching point set, estimating candidate homography matrixes and calculating corresponding re-projection errors, selecting the homography matrix with the largest number of inner points as an initial geometric transformation model, and performing optimization through a nonlinear least square method to obtain an optimal homography matrix; and splicing the matched images after geometric transformation estimation to obtain a panoramic image of the battery pack. According to the invention, high-quality battery pack imaging in weak light and weak texture environments is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and system for stitching panoramic images of battery packs in the cargo hold of new energy vehicle transport ships. Background Technology

[0002] Power batteries are transported in enclosed environments for extended periods during transnational maritime transport, making their operational safety and temperature status monitoring a major concern. Inside the hold of roll-on / roll-off (Ro-Ro) vessels, the dense parking of vehicles, insufficient lighting, and limited air circulation make it difficult to clearly obtain the external structure and temperature distribution of the power battery pack using ordinary imaging equipment. Furthermore, the periodic rolling and vibration of ships during navigation can cause viewing angle shifts, image blurring, or slight distortion in imaging equipment, thus affecting image stability.

[0003] To acquire images of the battery pack area, visible light cameras or infrared thermal imagers are typically used as the primary means of inspection. However, visible light imaging is greatly limited by the lighting conditions in the ship's cabin, resulting in high image noise, insufficient dynamic range, and difficulty in revealing the structural details of the battery pack surface. While infrared images can characterize temperature distribution, they contain extremely little texture information and have obvious blurred edges. When used alone, they are difficult to support high-precision spatial analysis and subsequent matching. Relying directly on a single modality often fails to obtain high-quality images suitable for structural identification and temperature analysis.

[0004] Traditional image enhancement and feature point matching algorithms also face significant bottlenecks in such environments. For example, conventional super-resolution models have a large number of parameters and high computational resource requirements, making them unsuitable for deployment on edge computing platforms relied upon by ship inspection equipment. Commonly used feature point algorithms, such as SIFT, ORB, and AKAZE, have low matching success rates in weak texture images, cross-modal images, and images with large changes in viewpoint. They are prone to problems such as insufficient matching points and high mismatch rates, which can lead to image misalignment or stitching failure.

[0005] On the other hand, when the inspection equipment moves inside the cabin, it continuously collects multiple frames of images. The field of view of a single frame image is limited, and it is usually necessary to rely on stitching multiple images to generate a larger view of the battery pack. If the feature correspondence between adjacent images is unstable, or the image resolution is insufficient to provide enough feature points, it is impossible to accurately estimate the geometric transformation, and ultimately it is difficult to obtain a complete panoramic image without obvious seams.

[0006] Therefore, existing technologies have significant shortcomings in improving image quality in low-light environments, enhancing feature consistency under cross-modal conditions, and stably stitching multi-frame images. There is an urgent need for an image enhancement and stitching technology that combines low computational power, strong robustness, and cross-modal adaptability to generate higher resolution, larger field of view, and clearer structural visual information of power battery packs, providing a reliable foundation for subsequent temperature anomaly detection and risk assessment. Summary of the Invention

[0007] To overcome the defects and shortcomings of existing technologies, this invention provides a panoramic image stitching method and system for battery packs in the cargo hold of new energy vehicle transport vessels. This invention achieves high-quality battery pack imaging in low-light and low-texture environments through lightweight dual-modal fusion enhancement, high- and low-frequency adaptive enhanced super-resolution networks, and dense feature matching and stitching framework.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] This invention provides a method for stitching panoramic images of battery packs in the cargo hold of new energy vehicle transport vessels, comprising the following steps:

[0010] Acquire visible light and thermal images of power battery packs for new energy vehicles;

[0011] Image preprocessing is performed, and the preprocessed visible light image is fused with the thermal imaging image.

[0012] A high- and low-frequency adaptive enhanced super-resolution network is constructed to perform super-resolution processing on the fused image;

[0013] The super-resolution processed image is subjected to pairwise feature matching to extract matching points between images. A matching confidence score is assigned to each pair of matching points. The matching results are filtered according to a preset confidence score threshold to obtain a set of feature matching points.

[0014] The feature matching point set is randomly sampled, the candidate homography matrix is ​​estimated and the corresponding reprojection error is calculated. The homography matrix with the most interior points is selected as the initial geometric transformation model, and the homography matrix is ​​optimized by nonlinear least squares method to obtain the optimal homography matrix.

[0015] Geometric transformation estimation is performed on the matched images based on the optimal homography matrix, and the matched images after geometric transformation estimation are stitched together to obtain a panoramic image of the battery pack.

[0016] As a preferred technical solution, image preprocessing includes the following steps:

[0017] Brightness adjustment, contrast adjustment, and adaptive histogram equalization are performed on visible light images.

[0018] Adjust the brightness and contrast of the thermal imaging image.

[0019] As a preferred technical solution, image fusion is performed on the preprocessed visible light image and thermal imaging image, specifically including: pixel-by-pixel weighted fusion, modal feature stitching, and structural texture mask fusion.

[0020] As a preferred technical solution, a high- and low-frequency adaptive enhanced super-resolution network is constructed to perform super-resolution processing on the fused image, specifically including:

[0021] Perform windmill convolution processing on the fused image;

[0022] The convolutional image is fed into the high- and low-frequency feature adaptive module after passing through the residual group. The high- and low-frequency feature adaptive module performs pooling operation on the input feature map, extracts high- and low-frequency features from the pooled output, fuses the high- and low-frequency features, and then performs convolution processing and channel attention operation on the fused features.

[0023] As a preferred technical solution, the high- and low-frequency feature adaptive module performs pooling operations on the input feature map, as follows:

[0024] ;

[0025] ;

[0026] in, This represents the low-frequency components of the image after average pooling. This indicates the average pooling operation. Represents the high-frequency components of an image. This means upsampling and restoring the low-frequency components of the image to the input feature map. Original feature dimensions.

[0027] As a preferred technical solution, high- and low-frequency features are extracted from the pooled output, specifically including:

[0028] ;

[0029] ;

[0030] ;

[0031] ;

[0032] in, This represents the separated high-frequency feature portion. This indicates a high-frequency feature separation operation. This indicates the enhanced low-frequency characteristic portion. This represents the Sigmoid activation function. This represents the average pooling output after two fully connected layers. This represents the maximum pooling output after two fully connected layers. This indicates the first fully connected layer. This indicates the second fully connected layer. This represents the modified linear unit activation function.

[0033] As a preferred technical solution, pairwise feature matching is performed on the super-resolution processed image. Specifically, the LoFTR algorithm is used for feature matching, as follows:

[0034] ;

[0035] in, These are the coordinates of the i-th key point in the first image. It is the corresponding key point in the second image. It is the confidence score of the match. This represents the total number of corresponding points detected.

[0036] As a preferred technical solution, the feature matching point set is randomly sampled, the candidate homography matrix is ​​estimated, and the corresponding reprojection error is calculated, as follows:

[0037] ;

[0038] in, Represents the homography matrix. This represents the coordinates of the i-th matching point in the source image. express The third component;

[0039] The homography matrix with the largest number of interior points is selected as the initial geometric transformation model, and is expressed as:

[0040] ;

[0041] in, This represents the homography matrix that maximizes the number of interior points. Denotes the set of interior points. This indicates the preset reprojection error threshold. This indicates a homography matrix filtering operation that maximizes the number of interior points.

[0042] As a preferred technical solution, geometric transformation estimation is performed on the matched images based on the optimal homography matrix, and the matched images after geometric transformation estimation are stitched together, specifically including:

[0043] To perform transform boundary calculations, the four corner points of the source image are defined as follows:

[0044] ;

[0045] Mapping the corner points to the target coordinate system is represented as:

[0046] ;

[0047] in, Represents the mapped corner points. Represents the optimal homography matrix;

[0048] Calculate the output canvas size;

[0049] Construct the translation matrix, represented as:

[0050] ;

[0051] ;

[0052] in, Represents the translation matrix. This indicates the minimum value on the x-axis of the output canvas. This indicates the minimum value on the y-axis of the output canvas;

[0053] The first image is deformed based on the translation matrix and mapped to the output canvas coordinate system;

[0054] Map the second image onto the output canvas according to the translation offset;

[0055] The first image after deformation and the second image after translation are superimposed on the same output canvas and stitched together to form the final panoramic image of the battery pack.

[0056] This invention also provides a panoramic image stitching system for battery packs in the cargo hold of new energy vehicle transport vessels, used to implement the above-mentioned panoramic image stitching method for battery packs in the cargo hold of new energy vehicle transport vessels, including: a dual-modal image acquisition module, an image preprocessing module, an image fusion module, a network construction module, a feature matching module, an optimal homography matrix construction module, a geometric transformation estimation module, and a panoramic stitching module;

[0057] The dual-modal image acquisition module is used to acquire visible light images and thermal imaging images of the power battery pack of new energy vehicles;

[0058] The image preprocessing is used to preprocess the acquired visible light images and thermal imaging images;

[0059] The image fusion module is used to fuse the preprocessed visible light image with the thermal imaging image;

[0060] The network construction module is used to construct a high- and low-frequency adaptive enhanced super-resolution network to perform super-resolution processing on the fused image.

[0061] The feature matching module is used to perform pairwise feature matching on the super-resolution processed image, extract matching points between images, assign matching confidence to each pair of matching points, and filter the matching results according to a preset confidence threshold to obtain a set of feature matching points.

[0062] The optimal homography matrix construction module is used to construct the optimal homography matrix. It randomly samples the feature matching point set, estimates the candidate homography matrix and calculates the corresponding reprojection error, selects the homography matrix with the most interior points as the initial geometric transformation model, and optimizes the homography matrix by nonlinear least squares method to obtain the optimal homography matrix.

[0063] The geometric transformation estimation module is used to estimate the geometric transformation of the matched image based on the optimal homography matrix.

[0064] The panoramic stitching module is used to stitch together the matched images that have undergone geometric transformation estimation to obtain a panoramic image of the battery pack.

[0065] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0066] (1) In the existing ship cabin inspection scenario, the visible light image of the power battery pack is severely affected by weak light, the infrared image has sparse texture and a limited number of cross-modal feature points, resulting in a low success rate of image alignment methods. At the same time, the inspection equipment has limited computing power and is not suitable for deploying large-scale image enhancement models, making it difficult to construct a stable and reliable panoramic image of the battery pack. This invention achieves high-quality battery pack imaging in weak light and weak texture environments through lightweight dual-modal fusion enhancement, high- and low-frequency adaptive enhancement super-resolution network, and dense feature matching and stitching.

[0067] (2) This invention effectively improves the image clarity and feature stability in low light and low texture environments by fusing and enhancing visible light and infrared images, super-resolution reconstruction and dense feature matching. Based on the high and low frequency adaptive enhanced super-resolution network, it significantly enhances detail information, so that cross-modal images can still maintain reliable matching when the viewing angle changes greatly. Combined with robust geometric transformation estimation and weighted fusion strategy, it can generate continuous and low distortion panoramic images of power battery packs. The overall lightweight design is suitable for edge device deployment and can provide high-quality visual evidence for the safety monitoring of power battery packs in the cabin of transport ships.

[0068] (3) The present invention adopts a high-low frequency decoupling and adaptive enhancement strategy to perform targeted processing on high-frequency texture details and low-frequency structural information respectively, thereby improving image resolution while maintaining overall structural consistency.

[0069] (4) The present invention introduces windmill convolution to enhance structural features in different directions, effectively improving the ability to express complex geometric structures and edge information, and is suitable for complex structure scenarios on the surface of battery packs.

[0070] (5) This invention combines LoFTR feature matching and RANSAC homography matrix estimation method to effectively suppress the interference of mismatch on geometric transformation results and improve the stability and accuracy of image stitching.

[0071] (6) The panoramic image generated by the present invention through canvas construction and overlapping area fusion strategy has good performance in terms of geometric alignment and visual continuity, and is suitable for subsequent temperature analysis and fault detection tasks. Attached Figure Description

[0072] Figure 1 This is a flowchart illustrating the panoramic image stitching method for battery packs in the cargo hold of a new energy vehicle transport vessel according to the present invention.

[0073] Figure 2 This is a schematic diagram of the overall network architecture of the high- and low-frequency adaptive enhanced super-resolution network of the present invention. Detailed Implementation

[0074] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0075] Example 1

[0076] like Figure 1 As shown, this embodiment provides a method for stitching panoramic images of battery packs in the cargo hold of a new energy vehicle transport ship, including the following steps:

[0077] S1: Perform dual-modal image acquisition: Acquire visible light images and thermal images of the power battery pack of the new energy vehicle from the visible light camera and the thermal imaging camera respectively according to the synchronous frame rate. The visible light camera is used to acquire the texture and structure information of the battery pack surface, and the thermal imaging camera is used to acquire the temperature distribution information of the battery pack surface, so as to obtain dual-modal images with complementary characteristics.

[0078] S2: Dual-modal image preprocessing: Preprocessing is performed separately on the acquired visible light image and thermal imaging image, specifically including:

[0079] Brightness adjustment, contrast enhancement, and contrast-limited adaptive histogram equalization (CLAHE) are applied to visible light images to improve the overall contrast and detail of the images. Brightness and contrast are adjusted to thermal imaging images to enhance the discernibility of temperature distribution.

[0080] S3: Dual-modal image fusion: The pre-processed visible light image and thermal imaging image are fused. The fusion process includes pixel-by-pixel weighted fusion, modal feature stitching and structural texture mask fusion, thereby generating a fused image that contains both structural and temperature information.

[0081] S4: As Figure 2 As shown, a high-low frequency adaptive enhanced super-resolution network is constructed. The fused image is input into the high-low frequency adaptive enhanced super-resolution network to perform super-resolution processing on the fused image, thereby improving the spatial resolution and detail representation ability of the image. Specifically, this includes:

[0082] S41: Input image P-convolution is used to enhance directional structural features in an image by performing windmill-shaped convolution. The convolution process is represented as follows:

[0083] ;

[0084] in, This represents the image after convolution of the windmill. This represents the windmill convolution operation;

[0085] S42: Construct residual groups by combining multiple residual blocks to maintain the stability of feature propagation while deepening the number of structural layers in the network, thereby effectively improving the feature expression ability and avoiding the gradient vanishing problem. One residual block consists of a 3×3 windmill convolution and activation function layer and another windmill convolution layer.

[0086] S43: The image after the residual group is input to the high- and low-frequency feature adaptive module for processing. First, the high- and low-frequency feature adaptive module performs pooling operation on the input feature map to separate features. The specific expression is as follows:

[0087] ;

[0088] ;

[0089] in, This represents the low-frequency components of the image after average pooling. This indicates the average pooling operation. Represents the high-frequency components of an image. Indicates to Partial upsampling, restoration to and input Existing feature dimensions;

[0090] S44: Subsequently, the features extracted in step S43 are separated into high-frequency features and enhanced low-frequency features;

[0091] The high-frequency feature separation formula is:

[0092] ;

[0093] In the formula, This represents the separated high-frequency feature portion; This indicates a high-frequency feature separation operation;

[0094] The formula for enhancing low-frequency characteristics is:

[0095] ;

[0096] In the formula, This indicates the enhanced low-frequency characteristic portion; This represents the Sigmoid activation function; This represents the average pooling output after two fully connected layers. This represents the maximum pooling output after two fully connected layers. The formulas for calculating both are:

[0097] ;

[0098] ;

[0099] In the formula, This indicates the second fully connected layer. This indicates the first fully connected layer; This represents the modified linear unit activation function;

[0100] Step S45: Perform feature fusion processing on the high- and low-frequency features extracted in step S44, using the following formula:

[0101] ;

[0102] In the formula, This indicates the fusion features after the fusion process; This means concatenating multiple feature tensors along a specified dimension to form a larger feature tensor.

[0103] Step S46: Perform convolution on the features fused in step S45 for subsequent attention mechanism calculations. The formula is:

[0104] ;

[0105] In the formula, This represents the fused features after convolution calculation; This indicates the processing of the first convolutional layer;

[0106] Step S47: Adaptively generate channel attention maps using the SEAttention module to dynamically weight features from different channels. This mechanism strengthens key features while suppressing irrelevant features, thus significantly improving the model's expressive power. The calculation formula is as follows:

[0107] ;

[0108] In the formula, This represents the output after processing by the attention mechanism; Indicates channel attention operation;

[0109] In this embodiment, the high- and low-frequency adaptive enhanced super-resolution network uses spatial downsampling and feature difference techniques to decouple high-frequency and low-frequency components, thereby realizing a frequency adaptive processing strategy. High-frequency components are enhanced through multi-scale dilated convolution to optimize fine textures and structural details, while low-frequency components are refined through channel attention mechanisms to maintain global structural consistency. After processing by the high- and low-frequency adaptive enhanced super-resolution network, the image pixel information is richer, which is beneficial for subsequent feature matching and image stitching.

[0110] Step S5: Perform pairwise feature matching on the super-resolution processed images from Step S4. Use the LoFTR algorithm based on deep learning to extract corresponding feature points between images and assign a matching confidence level to each pair of matching points. Then, filter the matching results according to the preset confidence threshold and remove low-confidence matching points to obtain a highly reliable set of feature matching points.

[0111] In this embodiment, the feature matching used in step S5 is specifically the LoFTR algorithm, which is expressed as follows:

[0112] ;

[0113] In the formula, These are the coordinates of the i-th key point in the first image; These are the corresponding key points in the second image; It is the confidence score of the match; This represents the total number of corresponding points detected.

[0114] To improve the reliability of geometric transformation estimation, matching results with low confidence scores are filtered out. The specific formula is as follows:

[0115] ;

[0116] In the formula, (Preferred to be 0.8) represents the confidence threshold, and M represents the number of remaining high-confidence matches.

[0117] Step S6: Under the RANSAC framework, random sampling is performed using the high-confidence feature matching points selected in Step S5 to estimate the candidate homography matrix and calculate the corresponding reprojection error; the homography matrix with the most interior points is selected as the initial geometric transformation model, and the homography matrix is ​​further optimized by nonlinear least squares method to obtain the final optimal homography matrix;

[0118] In this embodiment, geometric transformation estimation is performed on the image matched in step S5 to facilitate subsequent image stitching. Under the RANSAC framework, the candidate homography matrix is ​​estimated and the reprojection error is calculated through random sampling of the minimum point set, specifically as follows:

[0119] ;

[0120] In the formula, This represents the homography matrix, which is a matrix that describes the planar perspective transformation relationship between two images. This represents the coordinates (homogeneous coordinates) of the i-th matching point in the source image. express The third component (the scale component in homogeneous coordinates) is used for normalization;

[0121] The formula for selecting the homography matrix that maximizes the number of interior points is:

[0122] ;

[0123] In the formula, This represents the homography matrix that maximizes the number of interior points. Denotes the set of interior points. This represents the preset reprojection error threshold, preferably 3 to 8 pixels, used to determine whether a matching point is an interior point. This indicates a homography matrix filtering operation that maximizes the number of interior points;

[0124] Finally, the homography matrix is ​​further optimized using the nonlinear least squares method. ;

[0125] Step S7: Based on the optimal homography matrix obtained in Step S6, the two images for which geometric transformation estimation has been completed are stitched together, including transformation boundary calculation, output canvas size determination, perspective transformation, canvas construction, and overlapping region fusion, ultimately generating a continuous panoramic image, specifically including:

[0126] Step S71: Perform transformation boundary calculation, assuming the four corner points of the source image are:

[0127] ;

[0128] The formula for mapping corner points to the target coordinate system is:

[0129] ;

[0130] In the formula Represents the mapped corner points;

[0131] Step S72: Calculate the output canvas size using the following formula:

[0132] ;

[0133] ;

[0134] ;

[0135] ;

[0136] In the formula, This indicates the minimum value on the x-axis of the output canvas; This indicates the maximum value of the x-axis of the output canvas; This indicates the minimum value on the y-axis of the output canvas; This indicates the maximum value of the output canvas along the y-axis; Indicates the size of the reference image;

[0137] Step S73: To ensure that all pixels fall within the positive coordinate space, apply a translation matrix:

[0138] ;

[0139] ;

[0140] In the formula, This represents the translation matrix obtained after final calculation;

[0141] Step S74: Use the adjusted translation matrix The first image is deformed and mapped to the output canvas coordinate system using the following formula:

[0142] ;

[0143] In the formula, This represents the first image after deformation. This represents the first source image;

[0144] The second image is then mapped onto the output canvas according to the translation offset:

[0145] ;

[0146] In the formula, This represents the total canvas after stitching, i.e., the stitched panoramic image, with the same size as the original image. ; This represents the second image; the formula represents the total image. The coordinates correspond to the placement of the second image. ;

[0147] Finally, the first image after deformation The translated second image is then overlaid onto the same output canvas to create the final panoramic stitched image. Because... and Since the dimensions are the same, the final The image consists of the following two parts:

[0148] ;

[0149] ;

[0150] The output canvas serves as a unified coordinate space for the panoramic stitched images, simultaneously accommodating the first image after geometric transformation and the second image after translation. Its final content constitutes the panoramic stitching result.

[0151] By repeatedly performing the above steps on multiple images, a complete visible light-thermal dual-light panoramic stitched image of the automotive power battery pack was finally output.

[0152] Example 2

[0153] This embodiment provides a panoramic image stitching system for battery packs in the cargo hold of a new energy vehicle transport ship, which is used in the panoramic image stitching method for battery packs in the cargo hold of a new energy vehicle transport ship in Embodiment 1. The system includes: a dual-modal image acquisition module, an image preprocessing module, an image fusion module, a network construction module, a feature matching module, an optimal homography matrix construction module, a geometric transformation estimation module, and a panoramic stitching module.

[0154] In this embodiment, the dual-modal image acquisition module is used to acquire visible light images and thermal imaging images of the power battery pack of new energy vehicles. Specifically, the visible light camera acquires structural and texture information, and the infrared thermal imager acquires temperature distribution information. The two types of images are synchronized at the same frame rate and timestamps are recorded to ensure accurate subsequent fusion.

[0155] In this embodiment, image preprocessing is used to preprocess the acquired visible light image and thermal imaging image. The visible light image is subjected to brightness adjustment, contrast enhancement, and contrast-limited adaptive histogram equalization (CLAHE) to improve the overall contrast and detail performance of the image. The thermal imaging image is subjected to brightness and contrast adjustment to enhance the discernibility of temperature distribution.

[0156] In this embodiment, the image fusion module is used to perform image fusion on the preprocessed visible light image and thermal imaging image. By fusing the structural texture of the visible light image and the temperature distribution of the thermal imaging, the enhancement network can utilize both types of information simultaneously. The fusion methods include: pixel-wise weighted fusion, modal feature stitching, structural texture mask fusion, and the fusion result retains visible light details while including key infrared temperature distribution features.

[0157] In this embodiment, the network construction module is used to construct a high- and low-frequency adaptive enhanced super-resolution network to perform super-resolution processing on the fused image;

[0158] In this embodiment, the feature matching module is used to perform pairwise feature matching on the super-resolution processed image, extract matching points between images, assign matching confidence to each pair of matching points, and filter the matching results according to a preset confidence threshold to obtain a set of feature matching points. This module can effectively overcome the matching failure problem caused by the inconsistency between visible light and infrared features.

[0159] In this embodiment, the optimal homography matrix construction module is used to construct the optimal homography matrix. It randomly samples the feature matching point set, estimates the candidate homography matrix and calculates the corresponding reprojection error, selects the homography matrix with the most interior points as the initial geometric transformation model, and optimizes the homography matrix by nonlinear least squares method to obtain the optimal homography matrix.

[0160] In this embodiment, the geometric transformation estimation module is used to estimate the geometric transformation of the matched images based on the optimal homography matrix. This module can maintain image alignment accuracy even when there are a few mismatches.

[0161] In this embodiment, the panoramic stitching module is used to stitch the matched images that have completed geometric transformation estimation to obtain a panoramic image of the battery pack.

[0162] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for stitching panoramic images of battery packs in the cargo hold of a new energy vehicle transport vessel, characterized in that, Includes the following steps: Acquire visible light and thermal images of power battery packs for new energy vehicles; Image preprocessing is performed, and the preprocessed visible light image is fused with the thermal imaging image. A high- and low-frequency adaptive enhanced super-resolution network is constructed to perform super-resolution processing on the fused image; The super-resolution processed image is subjected to pairwise feature matching to extract matching points between images. A matching confidence score is assigned to each pair of matching points. The matching results are filtered according to a preset confidence score threshold to obtain a set of feature matching points. The feature matching point set is randomly sampled, the candidate homography matrix is ​​estimated and the corresponding reprojection error is calculated. The homography matrix with the most interior points is selected as the initial geometric transformation model, and the homography matrix is ​​optimized by nonlinear least squares method to obtain the optimal homography matrix. Geometric transformation estimation is performed on the matched images based on the optimal homography matrix, and the matched images after geometric transformation estimation are stitched together to obtain a panoramic image of the battery pack.

2. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Image preprocessing includes the following steps: Brightness adjustment, contrast adjustment, and adaptive histogram equalization are performed on visible light images. Adjust the brightness and contrast of the thermal imaging image.

3. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Image fusion is performed on the preprocessed visible light image and thermal imaging image, specifically including: pixel-wise weighted fusion, modal feature stitching, and structural texture mask fusion.

4. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Constructing a high- and low-frequency adaptive enhanced super-resolution network to perform super-resolution processing on the fused image, specifically including: Perform windmill convolution processing on the fused image; The convolutional image is fed into the high- and low-frequency feature adaptive module after passing through the residual group. The high- and low-frequency feature adaptive module performs pooling operation on the input feature map, extracts high- and low-frequency features from the pooled output, fuses the high- and low-frequency features, and then performs convolution processing and channel attention operation on the fused features.

5. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 4, characterized in that, The high- and low-frequency feature adaptive module performs a pooling operation on the input feature map, as shown below: ； ； in, This represents the low-frequency components of the image after average pooling. This indicates the average pooling operation. Represents the high-frequency components of an image. This means upsampling and restoring the low-frequency components of the image to the input feature map. Original feature dimensions.

6. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 4, characterized in that, Extracting high- and low-frequency features from the pooled output, specifically including: ； ； ； ； in, This represents the separated high-frequency feature portion. This indicates a high-frequency feature separation operation. This indicates the enhanced low-frequency characteristic portion. This represents the Sigmoid activation function. This represents the average pooling output after two fully connected layers. This represents the maximum pooling output after two fully connected layers. This indicates the first fully connected layer. This indicates the second fully connected layer. This represents the modified linear unit activation function.

7. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Pairwise feature matching is performed on the super-resolution processed image, specifically using the LoFTR algorithm for feature matching, as shown below: ； in, These are the coordinates of the i-th key point in the first image. It is the corresponding key point in the second image. It is the confidence score of the match. This represents the total number of corresponding points detected.

8. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Random sampling is performed on the feature matching point set to estimate the candidate homography matrix and calculate the corresponding reprojection error, as specifically expressed as: ； in, Represents the homography matrix. This represents the coordinates of the i-th matching point in the source image. express The third component; The homography matrix with the largest number of interior points is selected as the initial geometric transformation model, and is expressed as: ； in, This represents the homography matrix that maximizes the number of interior points. Denotes the set of interior points. This indicates the preset reprojection error threshold. This indicates a homography matrix filtering operation that maximizes the number of interior points.

9. The method for panoramic image stitching of battery packs for new energy vehicle transport ship compartments according to claim 1, characterized in that, Geometric transformation estimation is performed on the matched images based on the optimal homography matrix, and the matched images after geometric transformation estimation are stitched together, specifically including: To perform transform boundary calculations, the four corner points of the source image are defined as follows: ； Mapping the corner points to the target coordinate system is represented as: ； in, Represents the mapped corner points. Represents the optimal homography matrix; Calculate the output canvas size; Construct the translation matrix, represented as: ； ； in, Represents the translation matrix. This indicates the minimum value on the x-axis of the output canvas. This indicates the minimum value on the y-axis of the output canvas; The first image is deformed based on the translation matrix and mapped to the output canvas coordinate system; Map the second image onto the output canvas according to the translation offset; The first image after deformation and the second image after translation are superimposed on the same output canvas and stitched together to form the final panoramic image of the battery pack.

10. A panoramic image stitching system for battery packs in the cargo hold of a new energy vehicle transport vessel, characterized in that, The method for panoramic image stitching of battery packs for transporting new energy vehicles as described in any one of claims 1-9 includes: a dual-modal image acquisition module, an image preprocessing module, an image fusion module, a network construction module, a feature matching module, an optimal homography matrix construction module, a geometric transformation estimation module, and a panoramic stitching module. The dual-modal image acquisition module is used to acquire visible light images and thermal imaging images of the power battery pack of new energy vehicles; The image preprocessing is used to preprocess the acquired visible light images and thermal imaging images; The image fusion module is used to fuse the preprocessed visible light image with the thermal imaging image; The network construction module is used to construct a high- and low-frequency adaptive enhanced super-resolution network to perform super-resolution processing on the fused image. The feature matching module is used to perform pairwise feature matching on the super-resolution processed image, extract matching points between images, assign matching confidence to each pair of matching points, and filter the matching results according to a preset confidence threshold to obtain a set of feature matching points. The optimal homography matrix construction module is used to construct the optimal homography matrix. It randomly samples the feature matching point set, estimates the candidate homography matrix and calculates the corresponding reprojection error, selects the homography matrix with the most interior points as the initial geometric transformation model, and optimizes the homography matrix by nonlinear least squares method to obtain the optimal homography matrix. The geometric transformation estimation module is used to estimate the geometric transformation of the matched image based on the optimal homography matrix. The panoramic stitching module is used to stitch together the matched images that have undergone geometric transformation estimation to obtain a panoramic image of the battery pack.

Images