A visual slam system under low illumination conditions

Abstract

The application relates to a visual SLAM system under low-illumination conditions, which comprises an image preprocessing module and a SLAM module. The image preprocessing module comprises an image contrast enhancement submodule based on histogram equalization and an image denoising submodule based on domain transfer. The image contrast enhancement submodule based on histogram equalization is used for performing contrast enhancement on a low-illumination image to obtain a contrast-enhanced image, and the image denoising submodule based on domain transfer is used for performing denoising on the contrast-enhanced image. The visual SLAM system under low-illumination conditions has the advantages that the image contrast is improved and the noise is processed in the image preprocessing module, the low-illumination image detail expression capability is improved, the problem of poor SLAM input image quality under low-illumination conditions is solved from the input level, and the system has high precision and good stability.

Description

A visual SLAM system under low light conditions Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a visual SLAM system under low-light conditions. Background Technology

[0002] When mobile robots or other unmanned equipment work in dimly lit mines or enclosed indoor environments, the sensors are affected by low illumination conditions due to the lack of sufficient light sources, resulting in problems such as low contrast and low signal-to-noise ratio in the acquired images.

[0003] However, Visual SLAM (VSLAM, visual localization and mapping) systems heavily rely on the quality of the input image. Images captured in poorly lit environments may suffer from lost information in dark areas and unexpected noise. When the contrast of the input image is too low, the detectability and recognizability of feature points may decrease, making it difficult to extract feature points from the image. The inability to accurately extract feature points can further lead to matching errors, thus affecting the accuracy and precision of VSLAM pose estimation. Furthermore, when the signal-to-noise ratio is low, noise dominates, potentially causing image distortion, reduced or lost details. VSLAM algorithms typically use feature points for camera pose estimation and map construction. Under low signal-to-noise ratio conditions, feature points may not be extracted or incorrect feature points may be extracted. Furthermore, extracted feature points are prone to edge blurring, making it difficult to utilize effective and high-quality feature points in the front-end visual odometry. This can lead to errors or drift during matching, thus reducing the stability of the VSLAM algorithm. Summary of the Invention

[0004] Therefore, it is necessary to provide a visual SLAM system under low-light conditions to address the above-mentioned technical problems. Before performing simultaneous localization and map reconstruction, a preprocessing module that can enhance image contrast and reduce noise is added, which effectively improves the accuracy and stability of the entire system.

[0005] This invention provides a visual SLAM system under low-light conditions, comprising:

[0006] The image preprocessing module is used to enhance the contrast and remove noise from low-light images to obtain enhanced images;

[0007] The SLAM module is used for pose estimation based on the augmented image;

[0008] The image preprocessing module includes an image contrast enhancement submodule based on histogram equalization and an image denoising submodule based on domain transfer. The image contrast enhancement submodule based on histogram equalization is used to enhance the contrast of low-light images to obtain contrast-enhanced images, and the image denoising submodule based on domain transfer is used to denoise the contrast-enhanced images.

[0009] The image contrast enhancement submodule based on histogram equalization includes:

[0010] The feature extraction unit is used to extract features from low-light images;

[0011] The projection unit is connected to the output of the feature extraction unit and is used to project the reflection map and illumination map into the feature space corresponding to the features extracted by the feature extraction unit.

[0012] The histogram equalization unit is used to perform histogram equalization processing on low-light images. It extracts feature information rich in detail through a deep learning network, thereby obtaining the histogram prior information of the low-light image.

[0013] Constraint unit, used to constrain the reflectance map using histogram prior information of the low-light image;

[0014] The reconstruction unit is used to reconstruct the illumination map and the reflection map processed by the constraint module to obtain a contrast-enhanced image.

[0015] In one embodiment, the feature extraction unit includes an upper branch, a lower branch, and a connection subunit. The output terminals of the upper branch and the lower branch are both connected to the input terminal of the connection subunit. The low-light image is input to the feature extraction unit from the input terminals of the upper branch and the lower branch, respectively.

[0016] The upper branch, along the input to output direction of the low-light image, is sequentially connected to a 9×9 convolutional layer, three 3×3 convolutional + ReLU layers, a 3×3 deconvolutional + ReLU layer, and a 3×3 convolutional + ReLU layer. The input of the last 3×3 convolutional + ReLU layer is also connected to the output of the 9×9 convolutional layer.

[0017] The lower branch has a 3×3 convolution + ReLU layer.

[0018] In one embodiment, the domain-transfer-based image denoising submodule includes:

[0019] A noise encoder is used to extract noise information from contrast-enhanced images.

[0020] The first content encoder is used to extract content and detail information from the contrast-enhanced image;

[0021] The second content encoder is used to extract content and detail information from the noise-free image, which is the same size as the contrast-enhanced image.

[0022] A noise domain image processor is used to combine noise information from a contrast-enhanced image with content information from a noise-free image to obtain a noisy image.

[0023] A noise-free domain image processor is used to process the content information of contrast-enhanced images to obtain noise-free enhanced images.

[0024] In one embodiment, the input of the noise encoder is connected to the output of the reconstruction unit, the output of the noise encoder is connected to the input of the noise domain image processor, the input of the first content encoder is connected to the output of the reconstruction unit, the output of the first content encoder is connected to the input of the noise-free domain image processor, and the output of the second content encoder is connected to the input of the noise domain image processor.

[0025] In one embodiment, the detail information includes texture information of the contrast-enhanced image.

[0026] In one embodiment, the low-light image is input into the image preprocessing module in the form of a model;

[0027] The model for low-light images is

[0028]

[0029] In the formula, I represents the low-light image, R represents the reflectance map, L represents the illumination map, and N represents the noise component. This indicates element-wise multiplication. This represents the noise under the incident component. This represents a reflection map containing noise.

[0030] The beneficial effects of the present invention are as follows: The visual SLAM system under low illumination conditions of the present invention adds an image preprocessing module before the SLAM system, and completes the image contrast enhancement and noise processing in this module. Starting from improving the detail expression ability of low illumination images, it solves the problem of poor input image quality of SLAM under low illumination conditions from the input level, and effectively improves the accuracy and stability of the system. Attached Figure Description

[0031] Figure 1 is a schematic diagram of the framework of the visual SLAM system under low illumination conditions provided in an embodiment of the present invention;

[0032] Figure 2 is a schematic diagram of the image contrast enhancement submodule based on histogram equalization provided in an embodiment of the present invention;

[0033] Figure 3 is a schematic diagram of the network structure of the image preprocessing module provided in an embodiment of the present invention;

[0034] Figure 4 shows the result of the ORB-SLAM2 system provided in the embodiment of the present invention failing to initialize under the fr2-desk sequence. Detailed Implementation

[0035] 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.

[0036] In one embodiment, as shown in FIG1, FIG1 is a schematic diagram of the framework of a visual SLAM system under low illumination conditions provided in an embodiment of the present invention. The visual SLAM system under low illumination conditions of the present invention includes:

[0037] The image preprocessing module is used to enhance the contrast and remove noise from low-light images to obtain enhanced images.

[0038] The SLAM module is used for pose estimation based on the augmented image.

[0039] The image preprocessing module includes an image contrast enhancement submodule based on histogram equalization and an image denoising submodule based on domain transfer. The image contrast enhancement submodule based on histogram equalization is used to enhance the contrast of low-light images to obtain contrast-enhanced images, and the image denoising submodule based on domain transfer is used to denoise the contrast-enhanced images.

[0040] The image preprocessing module can improve the contrast and reduce noise in low-light images. Starting from improving the detail expression of low-light images, it solves the problem of poor input image quality in visual SLAM under low-light conditions from the input level.

[0041] As shown in Figure 2, the image contrast enhancement submodule based on histogram equalization includes:

[0042] The feature extraction unit is used to extract features from low-light images.

[0043] The projection unit is connected to the output of the feature extraction unit and is used to project the reflection map and illumination map into the feature space corresponding to the features extracted by the feature extraction unit.

[0044] The histogram equalization unit is used to perform histogram equalization on low-light images. It extracts feature information rich in detail through a deep learning network to obtain the histogram prior information of the low-light image.

[0045] The constraint unit is used to constrain the reflectance map using the histogram prior information of the low-light image.

[0046] The reconstruction unit is used to reconstruct the illumination map and the reflection map processed by the constraint module to obtain a contrast-enhanced image.

[0047] In one embodiment, as shown in Figure 3, this embodiment is based on Retinex theory and combined with histogram equalization method to suppress noise in the image while restoring details in the low-light image, thereby completing the construction of the low-light image preprocessing module network (Retinex, Histogram equalization and DeNosing, RHDN).

[0048] The feature extraction unit in the image preprocessing module network includes an upper branch, a lower branch, and a connection subunit. The output of the upper branch and the output of the lower branch are both connected to the input of the connection subunit. Low-light images are input to the feature extraction unit from the input of the upper branch and the input of the lower branch, respectively.

[0049] The upper branch, along the input-to-output direction of the low-light image, sequentially connects a 9×9 convolutional layer, three 3×3 convolutional + ReLU layers, a 3×3 deconvolutional + ReLU layer, and a 3×3 convolutional + ReLU layer. The input of the last 3×3 convolutional + ReLU layer is also connected to the output of the 9×9 convolutional layer. The lower branch has a 3×3 convolutional + ReLU layer.

[0050] In this embodiment, in the upper branch, the low-light input image first extracts features from the input image through a 9×9 convolutional layer, then passes through three 3×3 convolutional + ReLU layers and one 3×3 deconvolutional + ReLU layer. The residual features from the convolutional layers and the features from the deconvolutional layers are then fused before being fed into the next 3×3 convolutional + ReLU layer. To ensure the detail similarity between the enhanced image and the original image, in the lower branch, the low-light image is directly used as input. Image features are extracted through a 3×3 convolutional + ReLU layer, and these features are then concatenated with the previously stitched residual features from the upper layer. Next, a 3×3 convolutional layer projects the reflectance map and illumination map from the feature space, and a Sigmoid function is added to constrain the feature information within the range [0, 1]. Finally, histogram prior information is used to constrain the reflectance map, and the processed reflectance map and illumination map are reconstructed, thereby completing the contrast enhancement of the low-light image.

[0051] In the feature extraction unit of the image preprocessing module network in this embodiment, the low-light image is first decomposed into an illumination map and a reflection map. Then, the feature information after histogram equalization is used as a constraint. At the same time, the illumination smoothing loss is used to suppress the texture and color information in the illumination map, so as to preserve the inherent attributes as much as possible to restore the reflection map.

[0052] Considering that reflectance images are susceptible to noise interference, this invention utilizes a denoiser to denoise the contrast-enhanced image to further improve image quality. It eliminates noise while preserving content information in the reflectance image through domain transfer, thereby generating the final enhanced image. In one embodiment, the domain transfer-based image denoising submodule includes:

[0053] A noise encoder is used to extract noise information from contrast-enhanced images.

[0054] The first content encoder is used to extract content information and detail information from the contrast-enhanced image. The detail information includes the texture information of the contrast-enhanced image.

[0055] The second content encoder is used to extract content and detail information from the noise-free image, which is the same size as the contrast-enhanced image.

[0056] A noise domain image processor is used to combine noise information from a contrast-enhanced image with content information from a noise-free image to obtain a noisy image.

[0057] A noise-free domain image processor is used to process the content information of contrast-enhanced images to obtain noise-free enhanced images.

[0058] In the image preprocessing module network, the input of the noise encoder is connected to the output of the reconstruction unit, the output of the noise encoder is connected to the input of the noise domain image processor, the input of the first content encoder is connected to the output of the reconstruction unit, the output of the first content encoder is connected to the input of the noise-free domain image processor, and the output of the second content encoder is connected to the input of the noise domain image processor.

[0059] In one embodiment, the low-light image is input into the image preprocessing module in the form of a model. Taking into account environmental noise, and based on Retinex principles, the degraded low-light image can be modeled as follows:

[0060]

[0061] Since the illumination map determines the dynamic range of a low-light image, it is unaffected by noise; however, the reflectance component represents an inherent property of the image and is frequently affected by noise during imaging. Therefore, the model for low-light images can be further modified as follows:

[0062]

[0063] In the formula, I represents the low-light image, R represents the reflectance map, L represents the illumination map, and N represents the noise component. This indicates element-wise multiplication. This represents the noise under the incident component. This represents a reflection map containing noise.

[0064] When a low-light image is input using the model described above, the image contrast enhancement submodule based on histogram equalization first decomposes the low-light image into an illumination map L and a reflectance map. Then, the image I is subjected to histogram equalization. h The feature information, i.e., prior information, is used to constrain the reflectance map to preserve the detail information of the image during the enhancement process, resulting in a constrained reflectance map. The constrained reflection image is then fused with the illumination image to obtain the contrast-enhanced image I. X .

[0065] The domain-transfer-based image denoising submodule takes a noisy image and a noiseless image as input, given a noisy image I. X and an image without noise I Y First, using the noise encoder E N Extract Image I X The noise information in the data is processed using the first content encoder. And the first content encoder second From image I respectively X and I Y Extract the content and details of the image; then use the noise domain image processor G... X will come from I X Noise information and I Y The content information is combined to output a noisy image. Using a noise-free domain image processor G Y Output I X Noise-free image

[0066] The visual SLAM system under low-light conditions of this invention first utilizes the RHDN low-light image enhancement network to process images acquired by the camera, using histogram equalization as a priori to improve the quality of the color input image while enhancing contrast and reducing noise. Next, the RHDN-enhanced low-light image is used as input for localization and mapping. First, the tracking module estimates the camera pose, then the local mapping and loop closure detection modules complete the construction of a local map and the back-end optimization of the estimated trajectory. Finally, together, the system estimates the global pose and trajectory and reconstructs the surrounding environment map.

[0067] In one embodiment, the loss function set during training for the sub-network corresponding to the histogram equalization-based image contrast enhancement sub-module consists of three parts: illumination smoothing loss, histogram equalization constraint loss, and decomposition and reconstruction loss. The loss function of the sub-network corresponding to the histogram equalization-based image contrast enhancement sub-module will be explained in detail below:

[0068] (1) Illumination smoothing loss

[0069] During contrast enhancement, it's crucial to preserve as much texture and detail as possible in the reflection map. Therefore, the texture information in the illumination map should be smooth and have known structural boundaries. For the decomposed illumination map, it is constrained only with reference to the input low-light image, encouraging smoothness while preserving structural boundaries. Simultaneously, constraining the relative structure of this illumination map to be consistent with the input image can reduce over-smoothing at structural boundaries. Therefore, the illumination smoothing loss... The calculation is as follows:

[0070]

[0071] In the formula, |·| represents the absolute value operator, and ∈ is set to 0.01 to avoid 0 being used as a denominator in fraction operations. Let ||·||1 represent the gradient, and let L1 norm be represented. This loss can encourage the image contrast enhancement submodule based on histogram equalization to learn more sparse and simple features, thereby achieving image smoothing by finding a way to minimize the differences between pixels in the image.

[0072] (2) Histogram Equilibrium Constraint Loss

[0073] To preserve more texture and other details in the image, histogram equalization is used to constrain the output reflectance map. The mean squared error between the output reflectance map feature map and the input image feature map is defined as the histogram equalization constraint loss. The calculation is as follows:

[0074]

[0075] In the formula, F(·) represents the image features of the Conv 1-2 layer extracted by the VGG-19 model in the image after histogram equalization, and I h This represents the image after histogram equalization of the original low-light image I. This represents a reflection map containing noise, and ||·||2 represents the L2 norm.

[0076] (3) Decomposition and reconstruction loss

[0077] Since the image contrast enhancement submodule based on histogram equalization decomposes an image into an illumination map and a reflection map based on Retinex theory, the two decomposed maps should reconstruct the input image. To ensure the quality of the generated image, a decomposition and reconstruction loss is introduced.

[0078]

[0079] In summary, considering the constraints on the illumination map, reflection map, and image reconstruction, the overall loss function of the image contrast enhancement submodule based on histogram equalization is... for:

[0080]

[0081] In the formula, λ d , λ RH These are weighting coefficients, which were set to 0.5 and 0.1 respectively in the experiment based on experience.

[0082] The loss function set during training for the sub-network corresponding to the domain-transfer-based image denoising sub-module consists of three parts: KL divergence loss, content consistency loss, and smoothing loss. The loss function of the sub-network corresponding to the domain-transfer-based image denoising sub-module will be explained in detail below:

[0083] (1) KL divergence loss

[0084] Since high-quality images paired with low-light images are lacking for guidance during training, KL distance is introduced to constrain the extracted noise to conform to a standard normal distribution. Furthermore, the KL divergence loss is minimized to suppress the content information contained in the noise features. The KL divergence loss is calculated as follows:

[0085]

[0086] In the formula, l represents the feature dimension of the noise information extracted by the noise encoder, and μ i and σi Let represent the mean and standard deviation of the i-th noise feature, respectively.

[0087] (2) Loss of content consistency

[0088] Because training lacks paired image supervision, the generated denoised images may lose some content information. To ensure that the denoised image and the contrast-enhanced image have consistent content information as much as possible, content consistency loss is used. The calculation is as follows:

[0089]

[0090] In the formula, I X This represents the image output by the contrast enhancement module, which is also the noisy image input by the denoising module. This represents the image after denoising.

[0091] (3) Smoothing loss

[0092] The background of the denoised image tends to be overly smooth. Gaussian blur is introduced here to obtain image features under multi-scale conditions. Smoothing loss. The calculation is as follows:

[0093]

[0094] In the formula, λ σ The hyperparameter B controls the error at different Gaussian blur levels. σ (·) denotes a Gaussian blur operator with a Gaussian blur kernel σ. For σ = 5, 9, and 15, λ is set accordingly. σ =0.25, 0.5, 1.0.

[0095] In summary, considering the lack of paired images during training, potential loss of content information, color deviation, and background smoothing issues in the denoising process, the overall loss function of the sub-network corresponding to the domain transfer-based image denoising sub-module is:

[0096]

[0097] In the formula, λ KL , λ c , λ sm The weighting coefficient is usually set to: λ KL =0.01, λ c =10, λ sm =5.

[0098] In a specific embodiment, HN-SLAM is used as an abbreviation for the visual SLAM system under low illumination conditions. Since the HN-SLAM system of the present invention is an improvement based on the ORB-SLAM2 method, and in order to verify the superiority of the HN-SLAM method proposed in the present invention, the supervised signal-to-noise ratio-based low illumination image enhancement method SNR is replaced by the RHDN network and incorporated into the system framework of the present invention. The ORB-SLAM2 system and the SNR-SLAM system are used as comparison systems to illustrate the performance of the visual SLAM system under low illumination conditions of the present invention.

[0099] In this embodiment, after the degradation of the fr1-desk, fr1-floor, and fr2-desk sequences is completed, the fr1-desk, fr1-floor, and fr2-desk sequences under low illumination conditions are input into the HN-SLAM, ORB-SLAM2, and SNR-SLAM systems, respectively.

[0100] The ORB-SLAM2 system failed to complete initialization under low illumination conditions in the fr1-desk, fr1-floor, and fr2-desk sequences, as shown in Figure 4. Figure 4 shows the result of the ORB-SLAM2 system failing to initialize under the fr2-desk sequence. To visually demonstrate the experimental scenario, some visualization processing was performed.

[0101] When the ORB-SLAM2 method is used directly in the above three sequences, the poor image quality makes it impossible to extract feature points, causing SLAM to remain in the initialization stage and thus preventing subsequent pose calculation and mapping tasks.

[0102] To better evaluate the performance of the HN-SLAM method and SNR-SLAM, this section uses ATE (Absolute Trajectory Error) to evaluate the estimated pose.

[0103] Tables 1, 2, and 3 show the comparison results of ATE errors under fr1-desk, fr1-floor, and fr2-desk, respectively, where the bolded black part represents the optimal result under that error.

[0104] Table 1 Comparison of ATE errors under fr1-desk sequences

[0105]

[0106] Table 2 Comparison of ATE errors under fr1-floor sequences

[0107]

[0108] Table 3 Comparison of ATE errors under fr2-desk sequences

[0109]

[0110] As shown in Tables 1 and 2, under the fr1-desk and fr1-floor sequences, HN-SLAM performs significantly better than SNR-SLAM. In particular, under the fr1-floor sequence, the RMSE of the absolute trajectory error of HN-SLAM is only 0.0129, while the RMSE of the absolute trajectory error of SNR-SLAM reaches 0.1518. All errors of the absolute trajectory error of HN-SLAM are smaller than those of SNR-SLAM, and its performance is significantly better than that of HN-SLAM.

[0111] As shown in Table 3, under the fr2-desk sequence, the RMSE, Median, Mean, and Min of the absolute trajectory error of HN-SLAM are larger than those of SNR-SLAM, while its Std and Max are smaller. SNR-SLAM performs slightly better than HN-SLAM, but the difference is not significant.

[0112] In scenarios where the fr1-floor sequence contains relatively little texture information, the HN-SLAM system proposed in this embodiment outperforms the SNR-SLAM system; in scenarios with richer information, their performance is similar. Overall, HN-SLAM is better adaptable to various scenarios, improving estimation accuracy significantly under both high and low texture conditions, and outperforming SNR-SLAM in these scenarios.

[0113] To better evaluate the drift between the HN-SLAM and SNR-SLAM methods, this section uses RPE (Relative Pose Error) to evaluate the estimated camera pose. Tables 4, 5, and 6 show the comparison results of RPE errors under fr1-desk, fr1-floor, and fr2-desk, respectively, where the bolded black part represents the optimal result under that error.

[0114] Table 4 Comparison of RPE errors under fr1-desk sequences

[0115]

[0116] Table 5 Comparison of RPE errors under fr1-floor sequences

[0117]

[0118] Table 6 Comparison of RPE errors under fr2-desk sequences

[0119]

[0120] As shown in Tables 4 and 6, under the fr1-desk and fr2-desk sequences, all errors of the relative pose error of HN-SLAM are smaller than those of SNR-SLAM except for the minimum error, indicating that the drift of HN-SLAM is smaller under these two sequences.

[0121] As shown in Table 5, under the fr1-floor sequence, the RMSE of the relative pose error of HN-SLAM is 0.0089, which is an order of magnitude lower than the RMSE of 0.0606 for SNR-SLAM. The Std of the relative pose error of HN-SLAM is 0.2162, which is much smaller than the Std of 1.9557 for SNR-SLAM. Comparing these errors, it can be seen that HN-SLAM can control the relative pose error within a small range. Compared with SNR-SLAM, HN-SLAM has less drift in various scenarios and its system performance is more stable.

[0122] In summary, the HN-SLAM system of the present invention has high accuracy and good stability.

[0123] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0124] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of the present invention should be determined by the appended claims.

Claims

1. A visual SLAM system under low-light conditions, characterized in that, include: The image preprocessing module is used to enhance the contrast and remove noise from low-light images to obtain enhanced images; The SLAM module is used for pose estimation based on the enhanced image; the image preprocessing module includes an image contrast enhancement submodule based on histogram equalization and an image denoising submodule based on domain transfer. The image contrast enhancement submodule based on histogram equalization is used to enhance the contrast of the low-light image to obtain a contrast-enhanced image, and the image denoising submodule based on domain transfer is used to denoise the contrast-enhanced image. The image contrast enhancement submodule based on histogram equalization includes: a feature extraction unit for extracting features from the low-light image; a projection unit connected to the output of the feature extraction unit for projecting the reflection map and illumination map into the feature space corresponding to the features extracted by the feature extraction unit; a histogram equalization unit for performing histogram equalization processing on the low-light image, extracting feature information rich in detail through a deep learning network to obtain histogram prior information of the low-light image; a constraint unit for constraining the reflection map using the histogram prior information of the low-light image; and a reconstruction unit for reconstructing the illumination map and the reflection map processed by the constraint module to obtain a contrast-enhanced image.

2. The visual SLAM system under low illumination conditions according to claim 1, characterized in that, The feature extraction unit includes an upper branch, a lower branch, and a connection subunit. The outputs of both the upper and lower branches are connected to the input of the connection subunit. The low-light image is input to the feature extraction unit from the inputs of the upper and lower branches, respectively. The upper branch, along the input-to-output direction of the low-light image, sequentially connects a 9×9 convolutional layer, three 3×3 convolutional + ReLU layers, a 3×3 deconvolutional + ReLU layer, and a 3×3 convolutional + ReLU layer. The input of the last 3×3 convolutional + ReLU layer is also connected to the output of the 9×9 convolutional layer. The lower branch has a 3×3 convolutional + ReLU layer.

3. The visual SLAM system under low illumination conditions according to claim 2, characterized in that, The domain-transfer-based image denoising submodule includes: a noise encoder for extracting noise information from the contrast-enhanced image; a first content encoder for extracting content and detail information from the contrast-enhanced image; a second content encoder for extracting content and detail information from the noise-free image, wherein the noise-free image is the same size as the contrast-enhanced image; a noise domain image processor for combining the noise information from the contrast-enhanced image and the content information from the noise-free image to obtain a noisy image; and a noise-free domain image processor for processing the content information from the contrast-enhanced image to obtain a noise-free enhanced image.

4. The visual SLAM system under low illumination conditions according to claim 3, characterized in that, The input terminal of the noise encoder is connected to the output terminal of the reconstruction unit, the output terminal of the noise encoder is connected to the input terminal of the noise domain image processor, the input terminal of the first content encoder is connected to the output terminal of the reconstruction unit, the output terminal of the first content encoder is connected to the input terminal of the noise-free domain image processor, and the output terminal of the second content encoder is connected to the input terminal of the noise domain image processor.

5. The visual SLAM system under low illumination conditions according to claim 3, characterized in that, The detailed information includes the texture information of the contrast-enhanced image.

6. The visual SLAM system under low illumination conditions according to claim 5, characterized in that, Low-light images are input into the image preprocessing module in the form of a model; the model for low-light images is... In the formula, I represents the low-light image, R represents the reflectance map, L represents the illumination map, and N represents the noise component. This indicates element-wise multiplication. This represents the noise under the incident component. This represents a reflection map containing noise.

Images