Synchronous positioning and mapping method and system suitable for low-light condition

Abstract

The invention provides a synchronous positioning and mapping method and system suitable for a low light condition, electronic equipment and a storage medium, and aims to solve the problem that the existing SLAM technology is difficult to meet the actual operation requirements of real-time monitoring and quick response under a dark light condition. The method comprises the following steps: introducing an LSKAttention big kernel attention mechanism into an IAT-based image enhancement module, performing image enhancement processing to obtain an output result of image enhancement; and carrying out comprehensive extraction and clustering of image features on an output result through a preset closed-loop detection module. The problems of poor tracking image quality, poor mapping loopback effect and the like in the dark light environment can be solved, and accurate synchronous positioning and mapping of the environment in the dark light environment are realized.

Description

Technical Field

[0001] The present disclosure relates to the field of image processing technology, and in particular to a simultaneous positioning and mapping method suitable for low-light conditions, a simultaneous positioning and mapping system suitable for low-light conditions, an electronic device, and a computer-readable storage medium. Background Art

[0002] Currently, "digitalized" construction and governance have become a key direction for industry development. Three-dimensional visualization of underground environments is an integral part of this effort and crucial for achieving relevant construction goals. Accurate and efficient underground environmental exploration and governance can provide solid data support and intuitive spatial information display. However, traditional exploration and governance methods present numerous insurmountable challenges when faced with challenging environments, particularly in low-light conditions and complex, difficult-to-close environments, where their robustness falls far short of meeting practical requirements.

[0003] Traditional visual SLAM (Simultaneous Localization and Mapping) loop closure detection algorithms, such as the widely used Bag of Visual Words (BoVW) model and its derivatives (e.g., RTAB-MAP, methods based on the Gist descriptor, and methods using panoramic color image histograms combined with nearest neighbor learning), have demonstrated certain advantages in past research and practice. For example, RTAB-MAP successfully estimates loop closure probabilities using discrete Bayesian filters, meeting real-time requirements and providing a viable solution for online loop closure detection in large environments. The Gist descriptor extracts image information from two dimensions: direction and frequency, and uses Gabor filters to represent global image features. The method of panoramic color image histograms combined with nearest neighbor learning employed by Ulrich et al. also achieves image matching to a certain extent. However, these traditional methods share a key common issue: they all rely on manually designed features. This approach faces significant challenges in complex and changing environments with low light conditions. Varying lighting conditions, complex terrain, and diverse equipment make it difficult to manually design universal and effective feature descriptions. This not only limits the algorithm's ability to accurately extract image information, resulting in limited accuracy of closed-loop detection and inability to accurately identify and match image features, which in turn affects the positioning and mapping accuracy of the entire SLAM system, but also the process of manually designing features is cumbersome and time-consuming, which seriously restricts the real-time performance of the algorithm and makes it difficult to meet the actual operational needs of real-time monitoring and rapid response in dark light conditions. Summary of the Invention

[0004] In order to at least solve the problem that existing SLAM technologies in the prior art are unable to meet the actual operational requirements of real-time monitoring and rapid response in low-light conditions, the present disclosure provides a method for simultaneous positioning and mapping under low-light conditions, a simultaneous positioning and mapping system under low-light conditions, an electronic device, and a computer-readable storage medium, which can solve the problems of poor tracking image quality and poor mapping loopback effects in low-light environments, and achieve accurate simultaneous positioning and mapping of the environment under low-light conditions.

[0005] In a first aspect, the present disclosure provides a method for simultaneous positioning and mapping in low-light conditions, the method comprising:

[0006] Introducing LSKA (Large-Scale Kernel Attention) into the image enhancement module based on IAT (Image Adaptive Transformer)

[0007] Spatial attention)ttention large core attention mechanism, perform image enhancement processing, and obtain the output result of image enhancement;

[0008] The output results are subjected to comprehensive extraction and clustering of image features through a preset closed-loop detection module.

[0009] Furthermore, the LSKAttention large-core attention mechanism is introduced into the IAT-based image enhancement module to perform image enhancement processing, including:

[0010] Copy the input image to generate a copy;

[0011] For the original input image, DWConv (Depthwise SeparableConvolution) is used to perform convolution processing in the horizontal and vertical directions, and then two spatial dilation convolution DW-D-Conv operations are performed. Then, conv1 is used to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally the feature map attn (Attention Weight Map) is obtained.

[0012] A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

[0013] Furthermore, the loss function of the image enhancement module for image enhancement processing training adopts a comprehensive loss function L, and:

[0014] L=L1+α1L2+α2L SsIM

[0015] Among them, L1 is the mean absolute error, L2 is the mean square error, α1 is the L2 weight, L SsIM =1-SSIM, SSIM is the structural similarity index, α2 is L SSIM Weight.

[0016] Furthermore, the preset closed-loop detection module includes a feature extraction network and a NetVLAD (Network Vector of Locally Aggregated Descriptors) clustering network;

[0017] The feature extraction network uses the ConvNeXt-Tiny (Convolutional Next-Generation Tiny Architecture) network as the basic architecture to perform initial feature extraction on the image; and

[0018] By adding the ParC (Position-aware Circular Convolution) module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, the ParC-ConvNeXt Block module is obtained to further improve the feature extraction network's ability to extract image position information features; and

[0019] A CBAM (Convolutional Block Attention Module) attention mechanism module is also set after the ConvNeXt-Tiny network to make the feature information more concentrated;

[0020] The NetVLAD clustering network obtains representative global features by clustering local features of the image.

[0021] Furthermore, the output results are subjected to comprehensive extraction and clustering of image features through a preset closed-loop detection module, including:

[0022] The output result is used as the input image. The input image is passed through the feature extraction network to output a feature map, which is flattened into N D-dimensional local feature vectors. Each vector represents the semantic features of the local area of ​​the image.

[0023] The network has K learnable cluster centers with a shape of K×D, and the parameters of the cluster centers are automatically optimized through training;

[0024] For each local feature vector, calculate its residual with all cluster centers, and calculate the probability that each local feature belongs to each cluster through the Softmax function (normalized exponential function) based on the obtained residual;

[0025] For each cluster center k, aggregate the weighted residuals of all local features to obtain the descriptor of the cluster;

[0026] The descriptors of all clusters are concatenated into a K×D dimensional global feature vector and L2 normalized to enhance the scale invariance of the features.

[0027] In a second aspect, the present disclosure provides a simultaneous positioning and mapping system suitable for low-light conditions, the system comprising:

[0028] The image processing module is configured to introduce the LSKAttention large-core attention mechanism into the IAT-based image enhancement module to perform image enhancement processing and obtain the image enhanced output result;

[0029] The extraction and clustering module is configured to perform comprehensive extraction and clustering of image features on the output results through a preset closed-loop detection module.

[0030] Furthermore, the image processing module is specifically configured to:

[0031] Copy the input image to generate a copy;

[0032] For the original input image, DWConv is used to perform convolution processing in the horizontal and vertical directions, and then two spatial expansion convolution DW-D-Conv operations are performed. Then, conv1 is used to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally the feature map attn is obtained;

[0033] A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

[0034] Furthermore, the preset closed-loop detection module includes a feature extraction network and a NetVLAD clustering network;

[0035] The feature extraction network uses the ConvNeXt-Tiny network as the basic architecture to perform initial feature extraction on the image; and

[0036] By adding the ParC module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, a ParC-ConvNeXt Block module is obtained to further improve the feature extraction network's ability to extract image position information features; and,

[0037] A CBAM attention mechanism module is also set after the ConvNeXt-Tiny network to make the feature information more concentrated;

[0038] The NetVLAD clustering network obtains representative global features by clustering local features of the image.

[0039] In a third aspect, the present disclosure provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the simultaneous positioning and mapping method for low-light conditions as described in any one of the first aspects.

[0040] In a fourth aspect, the present disclosure provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for simultaneous positioning and mapping under low-light conditions described in any one of the first aspects above is implemented.

[0041] Beneficial effects:

[0042] This disclosure provides a simultaneous localization and mapping method, system, electronic device, and storage medium suitable for low-light conditions. The LSKA attention mechanism used in image enhancement significantly expands the network's receptive field by employing large-scale convolution kernels. This effectively utilizes large-scale feature information, improving the model's ability to understand images in complex low-light scenes. Loop closure detection enables comprehensive and effective extraction and clustering of image features, improving the accuracy and robustness of loop closure detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of a simultaneous positioning and mapping method for low-light conditions provided in the first embodiment of the present disclosure;

[0044] Figure 2 A schematic diagram of the structure of an LS_PEM provided in an embodiment of the present disclosure;

[0045] Figure 3 A schematic diagram of a ConvNeXt-Tiny network architecture provided in an embodiment of the present disclosure;

[0046] Figure 4 A schematic diagram of the architecture of a ParC_ConvNeXt Block provided in an embodiment of the present disclosure;

[0047] Figure 5 A schematic diagram of a NetVLAD clustering network architecture provided in an embodiment of the present disclosure;

[0048] Figure 6 A ParC_ConvNeXtT_NetVLAD provided in the embodiment of the present disclosure

[0049] Network architecture diagram;

[0050] Figure 7 A schematic diagram of the LSS-IAT Network model structure provided in an embodiment of the present disclosure;

[0051] Figure 8 A photographic image of a step portion including height differences provided in the second embodiment of the present disclosure;

[0052] Figure 9 A partial photo image of a coal mining tunnel with relatively good lighting conditions provided in the second embodiment of the present disclosure;

[0053] Figure 10 A photographic image of an unconstructed tunnel portion with poor lighting provided in the second embodiment of the present disclosure;

[0054] Figure 11 A comparison chart of image enhancement effects of different models provided in Example 2 of the present disclosure;

[0055] Figure 12 This is an architecture diagram of a simultaneous positioning and mapping system suitable for low-light conditions, provided in Example 3 of the present disclosure;

[0056] Figure 13 This is an architectural diagram of an electronic device provided in Example 4 of the present disclosure. DETAILED DESCRIPTION

[0057] To enable those skilled in the art to better understand the technical solutions of the present disclosure, the present disclosure is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments and drawings described herein are only used to explain the present disclosure, rather than to limit the present disclosure.

[0058] It should be noted that the terms "first", "second", etc. in the specification and claims of the present disclosure and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence; and, in the absence of conflict, the embodiments and features in the embodiments of the present disclosure can be arbitrarily combined with each other.

[0059] The terms used in the embodiments of the present disclosure are for the purpose of describing specific embodiments only and are not intended to limit the present disclosure. The singular forms "a," "an," "the," and "the" used in the embodiments of the present disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0060] In the subsequent description, suffixes such as "module," "component," or "unit" used to represent elements are used only to facilitate the description of the present disclosure and have no specific meaning. Therefore, "module," "component," or "unit" may be used interchangeably.

[0061] Existing technologies have various drawbacks when conducting underground environmental exploration and management in the dark environment of coal mines, such as:

[0062] Histogram equalization: This method enhances overall contrast by adjusting the distribution of the image's histogram. However, in practice, it can result in a reduction in image grayscale levels, obscuring some detailed information. In particular, when the image histogram contains high-density areas, the enhanced image often exhibits excessive contrast enhancement. This not only distorts the visual quality of the image but, more seriously, can obscure key texture features, equipment identification, or safety warnings in the environment, thus impacting subsequent analysis and decision-making.

[0063] Retinex: This method enhances images based on variable lighting conditions and the color reflectivity of objects. However, it can easily introduce additional image noise, which is particularly noticeable in complex environments. This increased noise can cause localized image distortion, blurring previously clear structures, equipment outlines, or signs of human activity, significantly reducing image reliability and usability.

[0064] Deep learning methods: While deep learning technology has made some progress in image enhancement, it still has limitations when processing images in complex lighting environments. When an image contains both bright and dark areas, applying global illumination strategies can lead to oversaturation of bright areas while underperforming the enhancement of dark areas. This unbalanced processing results in a decrease in overall image visual quality, making it difficult to effectively extract key information and failing to meet the requirements of high-precision image enhancement.

[0065] It can be seen that the various existing methods are difficult to meet the actual operational needs of real-time monitoring and rapid response in dark light conditions.

[0066] The following is a detailed description of the technical solutions of the present invention and how the technical solutions of the present invention solve the technical problems in the prior art with specific embodiments. It will be appreciated that, in the embodiments of the present application, the execution subject may perform some or all of the steps in the embodiments of the present application, and these steps or operations are merely examples. The embodiments of the present application may also perform other operations or variations of various operations. In addition, the various steps may be performed in different orders as presented in the embodiments of the present application, and it may not be necessary to perform all the operations in the embodiments of the present application. Furthermore, the following specific embodiments may be combined with each other, and the same or similar concepts or processes may not be described in detail in certain embodiments.

[0067] Figure 1 A flowchart of a method for simultaneous positioning and mapping under low-light conditions provided in the first embodiment of the present disclosure is shown as follows: Figure 1 As shown, the method includes:

[0068] Step S101: introducing the LSKAttention large-core attention mechanism into the IAT-based image enhancement module to perform image enhancement processing and obtain an image enhancement output result;

[0069] Step S102: performing comprehensive extraction and clustering of image features on the output results through a preset closed-loop detection module.

[0070] The present invention aims to provide a simultaneous localization and mapping (SLAM) method suitable for low-light conditions, so as to solve the problems of poor tracking image quality and poor mapping loopback effect existing in the existing technology in dark environments (such as coal mines, tunnels, underground, etc.), and realize accurate simultaneous localization and mapping of the environment in dark environments.

[0071] The simultaneous localization and mapping process includes image enhancement (pre-processing) and loop closure detection (post-processing). Image enhancement improves image quality in low-light conditions, resolving issues such as noise, low contrast, and blurred details. Loop closure detection identifies whether the current scene has been visited before, thereby correcting the accumulated errors of the SLAM system.

[0072] The SLAM process is low-light original image → enhanced image (clear details, reduced noise) → feature extraction network → high discriminative features.

[0073] Image enhancement is fundamental. Image quality determines feature quality. Providing high-quality input data to subsequent modules directly impacts overall system performance, improving the robustness of loop closure detection. Enhanced images reduce blur and noise, making the generated global descriptor more discriminative and significantly reducing the mismatch rate. Loop closure detection is the goal: high-precision positioning and mapping are achieved through enhanced data, which together form a closed technical loop.

[0074] In the disclosed embodiment, for low-light original images, image enhancement processing is performed in an IAT-based image enhancement module. IAT is a lightweight transformer model for image enhancement and exposure correction. The basic principle is to restore normal-light sRGB images under low-light or over / underexposure conditions by decomposing the image signal processor (ISP) pipeline into local and global image components. By introducing the LSKAttention large-kernel attention mechanism in the IAT-based image enhancement module, the network's ability to capture global contextual information of the image can be enhanced by expanding the receptive field of the convolution kernel, while dynamically focusing on key areas in combination with the attention mechanism. Taking the coal mine scene as an example, a large convolution kernel (such as 7×7) covers a wider area and can capture global structural information such as tunnel direction and equipment layout. Then, the effective area (such as safety signs and crack details) is dynamically enhanced through attention weights to suppress noise interference. In addition, the LSKA module takes less time and can meet real-time requirements.

[0075] The output results are then subjected to comprehensive image feature extraction and clustering through a preset closed-loop detection module. The preset closed-loop detection module includes two parts: feature extraction and feature clustering. Feature extraction can be performed using the ConvNeXt-Tiny network or the EfficientNet series network. EfficientNet optimizes the network structure and parameter configuration to improve feature extraction capabilities while ensuring computational efficiency. A position encoding module (such as the position encoding in the Transformer) is added to the appropriate position in the network to accurately extract image position information features. At the same time, the SK attention mechanism (Selective Kernel Attention) is used to adaptively select appropriate convolution kernels based on the different scales and importance of image features, thereby more effectively extracting the channel and spatial information of the feature map. Feature clustering can be performed using NetVLAD clustering or density-based clustering algorithms such as the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm. This algorithm clusters data points according to their density distribution. It does not require the pre-specification of the number of clusters and can automatically discover areas of different densities in the data set and divide density-connected data points into the same cluster. It may have better adaptability and accuracy for clustering image features in the complex environment of underground coal mines.

[0076] The function of feature extraction is to extract distinguishable environmental information from the image, use convolutional networks to detect local features such as corners, edges, and textures, and strengthen key areas (such as lane corners and equipment identification) through attention mechanisms (such as CBAM), thereby outputting high-dimensional feature maps (such as 384×14×14 tensors) that carry spatial and semantic information. The function of feature clustering is to aggregate scattered local features into compact global descriptors, perform soft allocation clustering on features through specific algorithms, generate a fixed-length vector (such as 64 dimensions) to represent the entire image, and output a global descriptor (such as a 64×384 vector), which can ensure local Figure 1 Consistency, in scene re-identification, for example, when the robot returns to a historical area, the closed loop is identified by descriptor similarity matching (such as cosine similarity > 0.85).

[0077] The disclosed embodiment introduces the LSKAttention mechanism in the image enhancement module, and significantly expands the receptive field of the network by adopting large-size convolution kernels. This enables the model to break through the limitations of traditional small convolution kernels and capture a wider range of contextual information in the image, thereby better understanding the relationship between different areas in the image. At the same time, by combining the attention mechanism with large convolution kernels, the model can more intelligently focus on key feature areas in the image and effectively learn long-distance dependencies between features. This capability is particularly important in complex scenes underground in coal mines. For example, under low-light conditions, it can better identify key information such as the outline of equipment in the distance, the direction of the tunnel, and potential safety hazards, thereby improving the entire system's ability to understand complex scenes and providing strong support for subsequent precise image enhancement and accurate positioning and mapping.

[0078] The image enhancement of the disclosed embodiments can be used to achieve global context perception and solve the problem of easy loss of local features; perform dynamic area enhancement and improve the utilization rate of key information in dark areas; and achieve efficient computing, balancing performance and real-time requirements. Then, through closed-loop detection, comprehensive and effective extraction and clustering of image features can be achieved, thereby improving the accuracy and robustness of closed-loop detection.

[0079] Furthermore, the LSKAttention large-core attention mechanism is introduced into the IAT-based image enhancement module to perform image enhancement processing, including:

[0080] Copy the input image to generate a copy;

[0081] For the original input image, DWConv is used to perform convolution processing in the horizontal and vertical directions, and then two spatial expansion convolution DW-D-Conv operations are performed. Then, conv1 is used to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally the feature map attn is obtained;

[0082] A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

[0083] The embodiment of the present disclosure introduces the LSKAttention large core attention mechanism into the image enhancement module based on IAT, realizes the effective utilization of large-scale feature information through a specific operation process, and improves the model's ability to understand complex scene images in coal mines. The structure of the core enhancement submodule LS_PEM in the IAT image enhancement network is as follows: Figure 2 As shown, it is designed for low-light environments. Its core functions include: edge sharpening: solving the problem of blurred edges in dark-light images (such as equipment outlines and roadway cracks); noise suppression: reducing high-frequency noise in dark areas (ISO noise variance reduced by 60%+); local contrast optimization: enhancing the visual significance of key areas (such as safety signs).

[0084] The operation process includes:

[0085] (1) After the input image enters the module, it is first copied to generate a copy. This copy will be used to generate the final output result in the subsequent calculation process.

[0086] (2) For the original input image, DWConv is used to perform convolution processing in the horizontal and vertical directions. This step can effectively extract the spatial features of the image and preliminarily mine the key information in the image.

[0087] (3) Then, two spatially expanded convolution (DW-D-Conv) operations are performed. In this way, the network can capture a wider range of contextual information, thereby better understanding the overall structure and feature distribution relationship of the image, laying the foundation for subsequent accurate processing.

[0088] (4) After the above processing, conv1 (1x1 convolutional layer) is used to perform feature integration operation on the obtained feature map, and finally the feature map attn is obtained.

[0089] (5) After completing all the above processing, a copy of the original input image is multiplied with the feature map attn obtained after processing by the attention mechanism to obtain the final output result. This process realizes the organic fusion of the original image information and the feature information filtered and enhanced by the attention mechanism, so that the output image highlights the key features while retaining the original information, thereby improving the quality and recognizability of the image.

[0090] The purpose of DWConv's horizontal and vertical separation convolution is to decompose the standard 2D convolution into two 1D convolutions (horizontal + vertical), reducing the amount of computation and preserving spatial features.

[0091] Horizontal convolution (H-DWConv): Use a [1×k] convolution kernel (such as [1×3]) to extract horizontal texture and edge features of the image (such as the horizontal structure of downhole equipment).

[0092] Vertical convolution (V-DWConv): Use [k×1] convolution kernel (such as [3×1]) to capture longitudinal features (such as the vertical profile of a mine tunnel).

[0093] The purpose of DW-D-Conv spatial dilation convolution is to expand the receptive field through the dilation rate and capture large-scale context.

[0094] Operation details:

[0095] First dilated convolution: Use [3×3] depthwise separable convolution with a smaller dilation rate (such as r=2) to initially cover mid-range features (such as the relationship between local devices and background).

[0096] Second dilated convolution: Increase the dilation rate (e.g., r = 4) to cover a larger area (e.g., the global illumination distribution of the entire alley).

[0097] Depth-separable design:

[0098] First, perform channel-by-channel convolution (DW) and then point-by-point convolution (1×1Conv) to balance computational efficiency and feature expression capabilities.

[0099] The local features extracted by DWConv and the contextual information captured by DW-D-Conv are jointly input into the large-core attention mechanism to generate the weight map attn. Ultimately, through multiplication, a balance is achieved between local detail enhancement and global structural optimization. Horizontal and vertical convolutions enhance lane orientation features, while dilated convolutions address long-range blur caused by dust and fog, significantly improving image usability in complex environments.

[0100] Furthermore, the loss function of the image enhancement module for image enhancement processing training adopts a comprehensive loss function L, and:

[0101] L=L1+α1L2+α2L SSIM

[0102] Among them, L1 is the mean absolute error, L2 is the mean square error, α1 is the L2 weight, L SSIM =1-SSIM, SSIM is the structural similarity index, α2 is L SSIM Weight.

[0103] The image enhancement module is trained using an improved loss function (FS–MISSloss) to optimize model parameters to improve image enhancement effects.

[0104] Calculation method:

[0105] The L1 loss function plays a crucial role in image enhancement. It primarily calculates the difference between corresponding pixel values. This calculation method allows the L1 loss function to effectively preserve edge information when processing images, which is crucial for maintaining the image's basic structure and contours. However, its drawback is that while focusing on edge information, it can ignore subtle textures and details as relatively unimportant changes, leading to the loss of some detailed information during the image enhancement process.

[0106] The calculation principle of the L2 loss function (mean square error MSE) is based on the average of the sum of squares of the differences between the predicted value and the true value. Its formula is:

[0107]

[0108] Where y represents the true value, represents the predicted value, and n is the number of samples. The characteristic of the L2 loss function is that it imposes a greater penalty on larger errors, thereby making the results generated by the model smoother and reducing the impact of outliers.

[0109] SSIM (Structural Similarity Index) is a metric that comprehensively evaluates image similarity from multiple dimensions. It comprehensively considers the brightness, contrast, and structure of the image. Its value ranges from -1 to 1, where 1 indicates that the two images are identical. For two images x and y, the SSIM formula is defined as follows:

[0110]

[0111] Where μ x and μ y are the average brightness of images x and y respectively: and are the variances of images x and y respectively; σ xy is the covariance of x and y; c1 and c2 are small constants introduced to stabilize the denominator to prevent division by zero.

[0112] Based on the definition of SSIM, the SSIM loss function is defined as L SSIM =1-SSIM In summary, the final loss function used is as follows:

[0113] SSIMloss=L1+α1L2+α2L SSIM

[0114] In this embodiment, α1 can be set to 0.2 and α2 can be set to 0.1. By properly adjusting these two weight coefficients, a balance is achieved between the L1, L2, and SSIM loss terms. This allows the loss function to retain image edges while placing greater emphasis on image visual quality and structural preservation, effectively reducing the loss of image details and generating high-quality enhanced images.

[0115] In the complex lighting environment of coal mines, the traditional single loss function is difficult to meet the various requirements for image enhancement effects. This embodiment introduces the L2 loss term and the SSIM loss term, and combines them with the L1 loss function to form a more comprehensive and effective loss function combination. This combination method gives full play to the advantages of each loss term and overcomes the limitations of a single loss function. Specifically, the edge-preserving property of the L1 loss function ensures that the basic structure and contour of the image are preserved, so that the enhanced image will not have structural deformation or blurred contours. The smoothing property of the L2 loss term makes the image smoother in details, reduces the influence of noise and outliers, and improves the overall quality of the image. The SSIM loss term utilizes the image structure information, making the enhanced image more similar to the original image in terms of brightness, contrast and structure, and more in line with the visual perception of the human eye. In summary, the improved loss function can generate enhanced images that perform well in terms of visual quality, structural integrity and detail richness, providing more reliable and accurate image data for environmental monitoring and mapping in coal mines.

[0116] Furthermore, the preset closed-loop detection module includes a feature extraction network and a NetVLAD clustering network;

[0117] The feature extraction network uses the ConvNeXt-Tiny network as the basic architecture to perform initial feature extraction on the image; and

[0118] By adding the ParC module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, a ParC-ConvNeXt Block module is obtained to further improve the feature extraction network's ability to extract image position information features; and,

[0119] A CBAM attention mechanism module is also set after the ConvNeXt-Tiny network to make the feature information more concentrated;

[0120] The NetVLAD clustering network obtains representative global features by clustering local features of the image.

[0121] The loop closure detection module based on deep learning includes a feature extraction network and a NetVLAD clustering network;

[0122] Feature extraction network structure composition: The embodiment of the present disclosure uses ConvNeXt-Tiny network (ConvNeXt) as the basic architecture to perform initial feature extraction on the image. The ConvNeXt-Tiny structure is as follows Figure 3 As shown in Figure 1, the ConvNeXt network has demonstrated excellent performance in the field of image classification. Its unique network structure and parameter configuration can effectively extract the basic features of the image. In order to further improve the network's ability to extract image position information features, the ParC-Net network is carefully introduced in stage 3 and stage 4 of the ConvNeXt network. Specifically, the last three ConvNeXtBlock modules in stage 3 (structured as Figure 3 As shown in part b of the figure, ParC is added to the last ConvNeXt Block module in stage4. After such replacement operation, the ParC-ConvNeXt Block module is obtained. The ParC-ConvNeXt Block module structure is as follows Figure 4 shown.

[0123] The ParC-Net network introduces ultra-large deep convolutions (with kernel size approximately twice the size of the input features), models long-term dependencies through global receptive fields, and enhances position sensitivity through implicit position encoding. It adds bifurcated gate units (BGUs), including spatial BGUs (to extract multi-scale spatial dependencies) and channel BGUs (to optimize channel mixing), which are more compact and efficient than traditional SE modules. It designs a unified local-global convolution block that combines large-kernel global convolutions with local deep convolutions. After fusing channel information through point-by-point convolution, the network processes the information in two branches, achieving more balanced feature interactions and improving the model's ability to capture multi-scale information. In the last three modules of stage 3 and the last module of stage 4 of the ConvNeXt network, the original deep convolutions are replaced with ParC modules (including position-aware convolutions) to form the ParC-ConvNeXt Block. After inputting the feature map, this module first normalizes it through LayerNorm. It then uses the ParC module to fuse pixel coordinate information (x, y) with feature values, modulating the convolution kernel weights with the position coordinates to achieve differentiated responses to features at different locations. Finally, the output is processed through GELU activation, point-by-point convolution, and residual connections. By explicitly introducing position coordinates and interacting with features, this structure can accurately capture the spatial relationship between features in the image, enhancing the model's ability to locate objects and model global structure. In particular, it improves the joint representation of position information and semantics at the mid- and high-level semantic feature stages.

[0124] To more comprehensively extract channel and spatial information from feature maps, the disclosed embodiments also introduce a CBAM attention mechanism. This mechanism can perform in-depth analysis of feature maps from both channel and spatial dimensions, dynamically adjusting weights based on feature importance. This highlights key features and suppresses unimportant information, providing richer and more representative feature information for subsequent image matching.

[0125] CBAM consists of a channel attention module and a spatial attention module, and can be seamlessly integrated into the existing CNN architecture. It captures the dependencies between channels through global average pooling and global maximum pooling, and then sends the two pooled features to a shared multi-layer perceptron to extract features, generating a channel attention map and a spatial attention map respectively, so as to focus on "which channels are important" and "where is the informative part". Since CBAM is simple, lightweight and universal in design, it can usually be added directly after the ParC-ConvNeXt Block module. The embodiment of the present disclosure adds a CBAM module after the entire ConvNeXt network, providing the model with a more comprehensive and effective feature extraction capability without major architectural modifications, and the additional computational overhead is also very small.

[0126] Implementation of feature extraction function: The advantage of the ConvNeXt network lies in its efficient feature extraction capability, which can quickly and accurately capture the basic feature information in the image. The addition of the ParC-Net network brings the system the ability to accurately extract the image position information features. In the underground coal mine environment, location information such as the location of equipment, the direction of the tunnel, and the layout of various facilities is crucial for accurate positioning and mapping. Through its unique structure and algorithm, the ParC-Net network can effectively identify and extract these position-related feature information, and organically integrate it with the basic features extracted by the ConvNeXt network. The CBAM attention mechanism further optimizes the quality of the feature map, making the feature information more concentrated and prominent. Through this series of structural designs and technical combinations, the feature extraction network can provide high-quality, comprehensive and accurate feature information for the subsequent closed-loop detection process, laying a solid foundation for the precise operation of the entire system.

[0127] The NetVLAD clustering network obtains representative global features by clustering local features of the image.

[0128] Clustering principle: NetVLAD is an efficient clustering algorithm that is improved on the basis of VLAD (Vector of Locally Aggregated Descriptors). The idea of ​​VLAD is to aggregate all local descriptors (such as SIFT, SURF, ORB, etc.) in an image and form a vector to represent the global descriptor of the entire image. The NetVLAD clustering network is as follows: Figure 5 As shown. In VLAD, the cluster center c k It is obtained by clustering a large number of features extracted from the data set in advance, and then extracting N D-dimensional local features x from the image i , and then combine these features with the K*D dimensional cluster centers obtained in advance:

[0129]

[0130] Where a k is a sign function, when x i When it belongs to the kth cluster, a k (x i ) is 1, otherwise it is 0; x i (j) represents the value of the i-th dimension of the i-th feature; similarly, c k (j) represents the value of the jth dimension of the kth cluster center. Through the above formula, we can get a global feature with a feature vector number of K*D, and use this global feature to represent the image.

[0131] In the entire improved ParC_ConvNeXtT_NetVLAD network, Figure 6 As shown in the figure, the input dimension is 3*480*640. After the image passes through the ParC_ConvNeXt module, the output feature map has a feature dimension of 738*15*20. The feature map then passes through the CBAM module to extract the channel and spatial information while keeping the dimension of the feature map unchanged. The feature map is then sent to the NetVLAD module for dimensionality reduction, so that the feature map becomes the required K*D dimensional feature, that is, 64*738.

[0132] Furthermore, through Figure 7 The LSS-IAT network model shown optimizes feature representation to improve model performance;

[0133] The LSS-IAT network model is located between the feature extraction and reconstruction / prediction steps (belonging to the training phase (used to optimize network parameters during model training)). Its function is to optimize feature representation through specific mechanisms (such as improved attention or loss constraints) to improve model performance. It receives the output of the feature extraction network (such as the feature map of ParC-ConvNeXt) and optimizes the feature representation through a dual-branch structure:

[0134] Local branch: Enhance detail textures (such as rust and cracks on equipment)

[0135] Global branch: Modeling scenario structure (such as lane topology, equipment layout)

[0136] In the LocalBranch Network, LS_PEM acts as a kernel space pixel enhancement module, which strengthens edges through large horizontal / vertical convolution (7×1, 1×7) and retains details through pixel-level normalization; path design: main path: three-level LS_PEM series connection → deep enhancement of local features; parallel path: direct convolution → retaining original feature information; synergistic effect: preventing artifacts caused by over-enhancement.

[0137] Global Branch Network (GPM), Global Pyramid Pooling module: Global Pyramid Pooling, multi-scale pooling (1×1, 3×3, 5×5) captures the scene hierarchy and outputs the global topological representation of the lane; Convolution stacking: 4-5 layers of convolution gradually abstract high-level semantics.

[0138] Local-global feature fusion,weighted output for illumination-robust feature representation.

[0139] Furthermore, the output results are subjected to comprehensive extraction and clustering of image features through a preset closed-loop detection module, including:

[0140] The output result is used as the input image. The input image is passed through the feature extraction network to output a feature map, which is flattened into N D-dimensional local feature vectors. Each vector represents the semantic features of the local area of ​​the image.

[0141] The network has K learnable cluster centers with a shape of K×D, and the parameters of the cluster centers are automatically optimized through training;

[0142] For each local feature vector, calculate its residual with all cluster centers, and calculate the probability of each local feature belonging to each cluster through the Softmax function based on the obtained residual;

[0143] For each cluster center k, aggregate the weighted residuals of all local features to obtain the descriptor of the cluster;

[0144] The descriptors of all clusters are concatenated into a K×D dimensional global feature vector and L2 normalized to enhance the scale invariance of the features.

[0145] 1. Local feature preprocessing

[0146] The input image is passed through a feature extraction network (such as ParC-ConvNeXt) to output a feature map (such as 738×15×20), which is flattened into N D-dimensional local feature vectors (N=15×20=300, D=738), each of which represents the semantic features of a local area of ​​the image (such as edges, textures, and device contours).

[0147] 2. Learnable cluster centers

[0148] The network has K built-in learnable cluster centers (such as K=64) with a shape of K×D. The parameters of the cluster centers are automatically optimized through training, rather than relying on pre-trained clustering (such as K-means) like traditional VLAD.

[0149] The role of the cluster center: as the "prototype" of image features, representing different types of local features (such as "tunnel wall", "equipment metal surface", etc.).

[0150] 3. Residual Calculation and Soft Allocation

[0151] For each local eigenvector (f i ), calculate its relationship with all cluster centers (c k )’s residual:

[0152]

[0153] Soft assignment mechanism: The Softmax function is used to calculate the probability (weight) of each local feature belonging to each cluster, rather than hard assignment (0 / 1 attribution). The formula is: ‖f i -c k ‖ 2 is the square of the Euclidean distance between the feature and the center, ω ik ∈(0,1), (Probability normalization), the weight reflects the similarity between the local feature and the cluster center, allowing one feature to contribute to multiple cluster centers.

[0154] 4. Global Feature Aggregation

[0155] For each cluster center k, aggregate the weighted residuals of all local features to obtain the descriptor of the cluster: The descriptors of all clusters are concatenated into a K×D dimensional global feature vector (e.g., 64×738) and L2 normalized to enhance the scale invariance of the features.

[0156] In the complex environment of underground coal mines, image features are distributed in a complex and diverse manner. Traditional clustering methods struggle to accurately classify and characterize image features, which impacts the accuracy of loop closure detection. The NetVLAD clustering algorithm, through its unique design, effectively addresses this problem. It rationally clusters local image features to generate representative global features. During loop closure detection, these global features more accurately reflect the similarities and differences between images. By comparing the global features of different images, the system can more accurately determine whether images belong to the same location or scene, achieving higher-accuracy image matching. This plays a crucial role in improving the loop closure detection performance and positioning and mapping accuracy of the entire SLAM system in the complex environment of underground coal mines. It effectively avoids positioning deviations and mapping errors caused by image matching errors, ensuring stable and accurate system operation.

[0157] The disclosed embodiment introduces the LSK Attention large-core attention mechanism into the IAT-based image enhancement module, and through a specific operation process, it realizes the effective utilization of large-scale feature information, thereby improving the model's ability to understand complex scene images in coal mines. By improving the loss function, the L1, L2 and SSIM loss terms are reasonably combined to improve the image visual quality and structure preservation ability while maintaining the image edges. In the closed-loop detection module, the ConvNeXt-Tiny network, ParC-Net network, CBAM attention mechanism and NetVLAD are innovatively combined to achieve comprehensive and effective extraction and clustering of image features, improving the accuracy and robustness of closed-loop detection.

[0158] The second embodiment of the present disclosure also provides a synchronous positioning and mapping method suitable for low-light conditions, and constructs a special experimental platform to verify the effectiveness and practicality of the present disclosure. The iPhone 14 Pro with a handheld gimbal is used as a data acquisition device, and its built-in camera is fully utilized to collect high-quality color images, while the millimeter-wave radar is used to obtain accurate depth information. The experimental scene is selected in a representative real coal mine industrial environment, which covers a variety of complex scenes commonly seen in coal mines, including steps with obvious height differences (such as Figure 8 As shown in the figure), the coal mining tunnel section with relatively good lighting conditions (such as Figure 9 as shown) and the unconstructed tunnel section with poor lighting (as shown Figure 10 shown).

[0159] This diverse scene setting can fully test the performance of the disclosure under different lighting and terrain conditions.

[0160] Image enhancement experiment

[0161] In the image enhancement experiments, a large number of real-world images from coal mines under low-light conditions were collected as experimental data. These images were processed and compared using the IAT-based image enhancement module (LSS-IAT model) disclosed in this paper and other common image enhancement algorithms. The results are shown in Table 1 below.

[0162] Table 1: Performance comparison of different image enhancement algorithms

[0163]

[0164] Experimental results demonstrate that the LSS-IAT model demonstrates excellent performance on a real-world underground mine dataset. Its peak signal-to-noise ratio (PSNR) reaches 0.878, and its structural similarity index (SSIM) reaches 22.86. Compared to other commonly used image enhancement algorithms, the LSS-IAT model achieves at least 4.5% higher PSNR and 12.4% higher SSIM. Furthermore, the optimized LSS-IAT model has only 0.023 parameters (M), making it well-suited for real-time deployment. In practical applications, it can quickly process image data and provide timely and effective enhanced images.

[0165] Comparison of image enhancement effects of different models Figure 11 shown.

[0166] An analysis of images enhanced by different algorithms from the perspective of subjective visual effects shows that the brightness of images processed by the EFINet and Zero-DCE models is significantly dim, resulting in unsatisfactory overall image enhancement and difficulty in clearly presenting key information in the coal mine. While the GLADNet model performs well in terms of brightness enhancement, it has shortcomings in color restoration, resulting in image color distortion and affecting the image's authenticity and recognizability. While the images enhanced by the RUAS algorithm show improvements in both brightness and color, they lose some important image details, resulting in blurring and making some subtle geological features or equipment markings difficult to discern. Images enhanced by the IAT model can better restore brightness, but there is still room for improvement in restoring local image details. In contrast, images enhanced by the LS-IAT model perform well in terms of contrast and color restoration, more effectively preserving image details. Both the outlines of distant objects and nearby texture features are clearly presented, resulting in a significant subjective observation effect.

[0167] The third embodiment of the present disclosure also provides a synchronous positioning and mapping system suitable for low light conditions, such as Figure 12 As shown, the system includes:

[0168] An image processing module 11 is configured to introduce the LSKAttention large-core attention mechanism into the IAT-based image enhancement module to perform image enhancement processing and obtain an image enhanced output result;

[0169] The extraction and clustering module 12 is configured to perform comprehensive extraction and clustering of image features on the output results through a preset closed-loop detection module.

[0170] Furthermore, the image processing module 11 is specifically configured to:

[0171] Copy the input image to generate a copy;

[0172] For the original input image, DWConv is used to perform convolution processing in the horizontal and vertical directions, and then two spatial expansion convolution DW-D-Conv operations are performed. Then, conv1 is used to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally the feature map attn is obtained;

[0173] A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

[0174] Furthermore, the loss function of the image enhancement module for image enhancement processing training adopts a comprehensive loss function L, and:

[0175] L=L1+α1L2+α2L SSIM

[0176] Among them, L1 is the mean absolute error, L2 is the mean square error, α1 is the L2 weight, L SSIM =1-SSIM, SSIM is the structural similarity index, α2 is L SSIM Weight.

[0177] Furthermore, the preset closed-loop detection module includes a feature extraction network and a NetVLAD clustering network;

[0178] The feature extraction network uses the ConvNeXt-Tiny network as the basic architecture to perform initial feature extraction on the image; and

[0179] By adding the ParC module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, a ParC-ConvNeXt Block module is obtained to further improve the feature extraction network's ability to extract image position information features; and,

[0180] A CBAM attention mechanism module is also set after the ConvNeXt-Tiny network to make the feature information more concentrated;

[0181] The NetVLAD clustering network obtains representative global features by clustering local features of the image.

[0182] Furthermore, the extraction and clustering module 12 is specifically configured as follows:

[0183] The output result is used as the input image. The input image is passed through the feature extraction network to output a feature map, which is flattened into N D-dimensional local feature vectors. Each vector represents the semantic features of the local area of ​​the image.

[0184] The network has K learnable cluster centers with a shape of K×D, and the parameters of the cluster centers are automatically optimized through training;

[0185] For each local feature vector, calculate its residual with all cluster centers, and calculate the probability of each local feature belonging to each cluster through the Softmax function based on the obtained residual;

[0186] For each cluster center k, aggregate the weighted residuals of all local features to obtain the descriptor of the cluster;

[0187] The descriptors of all clusters are concatenated into a K×D dimensional global feature vector and L2 normalized to enhance the scale invariance of the features.

[0188] The simultaneous positioning and mapping system for low-light conditions in the embodiment of the present disclosure is used to implement the simultaneous positioning and mapping method for low-light conditions in method embodiment 1 and embodiment 2, so the description is relatively simple. For details, please refer to the relevant description in the previous method embodiment, which will not be repeated here.

[0189] In addition, if Figure 13 As shown, the fourth embodiment of the present disclosure further provides an electronic device, including a memory 100 and a processor 200, wherein the memory 100 stores a computer program. When the processor 200 runs the computer program stored in the memory 100, the processor 200 executes the above-mentioned various possible methods.

[0190] The memory 100 is connected to the processor 200 . The memory 100 may be a flash memory, a read-only memory, or other memory. The processor 200 may be a central processing unit or a single-chip microcomputer.

[0191] In addition, an embodiment of the present disclosure further provides a computer-readable storage medium, on which a computer program is stored, and the computer program is used by a processor to execute the above-mentioned various possible methods.

[0192] The computer-readable storage medium includes volatile or nonvolatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable read only memory), flash memory or other memory technology, CD-ROM (Compact Disc Read-Only Memory), Digital Versatile Disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer.

[0193] It is understood that the above embodiments are merely exemplary embodiments for illustrating the principles of the present disclosure, and the present disclosure is not limited thereto. Those skilled in the art may make various modifications and improvements without departing from the spirit and substance of the present disclosure, and such modifications and improvements are also considered to be within the scope of protection of the present disclosure.

Claims

1. A simultaneous positioning and mapping method suitable for low-light conditions, characterized in that: The method comprises: The LSKAttention large-core attention mechanism is introduced into the image enhancement module based on image adaptive transformation (IAT) to perform image enhancement processing and obtain the image enhanced output result; The output results are subjected to comprehensive extraction and clustering of image features through a preset closed-loop detection module.

2. The method according to claim 1, characterized in that The LSKAttention large-core attention mechanism is introduced into the IAT-based image enhancement module to perform image enhancement processing, including: Copy the input image to generate a copy; For the original input image, use the depth-wise separable convolution DWConv to perform convolution processing in the horizontal and vertical directions, then perform two spatial expansion convolution DW-D-Conv operations, and then use conv1 to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally obtain the feature map attention weight map attn; A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

3. The method according to claim 2, characterized in that The loss function of the image enhancement module for image enhancement processing training adopts a comprehensive loss function L, and: L=L1+α1L2+α2L SSIM Among them, L1 is the mean absolute error, L2 is the mean square error, α1 is the L2 weight, L SSIM =1-SSIM, SSIM is the structural similarity index, α2 is L SSIM Weight.

4. The method according to claim 1, wherein The preset closed-loop detection module includes a feature extraction network and a network local aggregation descriptor NetVLAD clustering network; The feature extraction network uses the ConvNeXt-Tiny network as the basic architecture to perform initial feature extraction on the image; and By adding the position-aware recurrent convolution ParC module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, the ParC-ConvNeXtBlock module is obtained to further improve the feature extraction network's ability to extract image position information features; and A convolutional block attention module (CBAM) is also set after the ConvNeXt-Tiny network to make the feature information more concentrated. The NetVLAD clustering network obtains representative global features by clustering local features of the image.

5. The method according to claim 4, characterized in that The output results are subjected to comprehensive extraction and clustering of image features through a preset closed-loop detection module, including: The output result is used as the input image. The input image is passed through the feature extraction network to output a feature map, which is flattened into N D-dimensional local feature vectors. Each vector represents the semantic features of the local area of ​​the image. The network has K learnable cluster centers with a shape of K×D, and the parameters of the cluster centers are automatically optimized through training; For each local feature vector, calculate its residual with all cluster centers, and calculate the probability that each local feature belongs to each cluster through the normalized exponential function Softmax function based on the obtained residual; For each cluster center k, aggregate the weighted residuals of all local features to obtain the descriptor of the cluster; The descriptors of all clusters are concatenated into a K×D dimensional global feature vector and L2 normalized to enhance the scale invariance of the features.

6. A simultaneous positioning and mapping system suitable for low-light conditions, characterized in that: The system comprises: The image processing module is configured to introduce the LSKAttention large-core attention mechanism into the IAT-based image enhancement module to perform image enhancement processing and obtain the image enhanced output result; The extraction and clustering module is configured to perform comprehensive extraction and clustering of image features on the output results through a preset closed-loop detection module.

7. The system according to claim 6, characterized in that The image processing module is specifically configured as follows: Copy the input image to generate a copy; For the original input image, DWConv is used to perform convolution processing in the horizontal and vertical directions, and then two spatial expansion convolution DW-D-Conv operations are performed. Then, conv1 is used to perform feature integration operation on the feature map obtained by the DW-D-Conv operation, and finally the feature map attn is obtained; A copy of the original input image is multiplied with the obtained feature map attn to obtain the image enhanced output.

8. The system according to claim 6, wherein: The preset closed-loop detection module includes a feature extraction network and a NetVLAD clustering network; The feature extraction network uses the ConvNeXt-Tiny network as the basic architecture to perform initial feature extraction on the image; and By adding the ParC module to the last three ConvNeXt Block modules in stage 3 and the last ConvNeXt Block module in stage 4 of the ConvNeXt-Tiny network, a ParC-ConvNeXt Block module is obtained to further improve the feature extraction network's ability to extract image position information features; and, A CBAM attention mechanism module is also set after the ConvNeXt-Tiny network to make the feature information more concentrated; The NetVLAD clustering network obtains representative global features by clustering local features of the image.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor runs the computer program stored in the memory, the processor executes the simultaneous positioning and mapping method suitable for low-light conditions according to any one of claims 1 to 5.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by the processor, the simultaneous positioning and mapping method suitable for low-light conditions according to any one of claims 1 to 5 is implemented.

Images