Night fast sky segmentation method and device based on deep learning

Abstract

The application discloses a kind of night fast sky segmentation method and device based on deep learning.The night fast sky segmentation method based on deep learning includes: obtaining the night sky picture after pre-processing;The night sky picture is input into the illumination enhanced image segmentation model for segmentation, to obtain the segmentation picture of night sky, wherein the illumination enhanced image segmentation model includes: illumination enhancement network and semantic segmentation network, the illumination enhancement network carries out illumination enhancement with the night sky picture as input, and the semantic segmentation network carries out night sky segmentation based on the output of the illumination enhancement network.The method can quickly and accurately distinguish sky and non-sky area in night environment, effectively solve the problem of sky segmentation in poor light conditions environment, greatly improve the efficiency of segmentation.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and image processing technology, and in particular to a method and apparatus for fast night sky segmentation based on deep learning. Background Technology

[0002] Image semantic segmentation is an important task in computer vision, enabling detailed understanding of specific objects by labeling each pixel in an image. This technology has wide applications in fields such as intelligent transportation devices (e.g., autonomous driving, 3D reconstruction, and video surveillance).

[0003] In recent years, deep convolutional neural networks have made significant progress and have become the mainstream approach in many fields. Compared with traditional methods, these networks achieve better results by providing reliable functionality to accomplish specific tasks. The earliest proposed fully convolutional neural network method was proposed by Long et al. for solving semantic segmentation tasks. Since then, many high-performance networks have emerged in this field, continuously driving the development of semantic segmentation technology.

[0004] Meanwhile, inspired by transfer learning, many systems transform general classification neural networks into backbone networks and integrate prior information to further enhance the model's feature encoding capabilities. There are many advanced sky segmentation models, such as academic papers 1 "Fakeye: Sky augmentation with real-time sky segmentation and texture blending" (CVPR Workshop on Computer Vision for Augmented and Virtual Reality, 2020) and 2 "Sky is not the limit: semantic-aware sky replacement" (ACM Transactions on Graphics, vol. 35, no. 4, pp. 149: 1–149: 11, 2016).

[0005] However, these networks are not optimized for nighttime scenes. Limited lighting conditions result in low image contrast, blurring subtle semantic features and boundaries, making accurate object segmentation difficult. Underexposure or overexposure issues lead to loss or distortion of image details, further complicating segmentation.

[0006] Image semantic segmentation is the process of dividing an image into different semantic regions, each of which is assigned a specific semantic label. For many deep neural network models, high-quality pixel-level night scene image annotation is a prerequisite for training, as they require accurate labeled data to learn semantic boundaries and regions. However, since most current semantic segmentation datasets are based on daytime scenes, neural networks trained in this environment perform poorly in low-light conditions, exhibiting a significant performance drop.

[0007] Chinese patent document CN115775344A discloses a method for constructing sky segmentation video and image data. The method includes acquiring video frame images and inertial sensor data through a mobile device camera; performing binary semantic segmentation on each frame image using deep learning to obtain a coarse real-world sky region mask; and employing a refinement operation to obtain a corresponding set of real-world sky segmentation images. A virtual scene is created based on the captured real-world sky video and / or a custom theme. The camera position and motion mode are set to acquire video frame images and inertial sensor data. The material information of objects and skyboxes in the virtual scene is modified to obtain a virtual sky region mask. Finally, the virtual sky region mask is added to the set of real-world sky segmentation images to obtain a precise and large dataset of sky segmentation images.

[0008] While the aforementioned scheme can generate a large dataset of sky-segmented images with inertial sensor data relatively quickly and accurately, it relies heavily on deep learning and inertial sensor data for sky segmentation. This may lead to reduced segmentation performance in situations with insufficient sensor data or in complex environments, such as at night or in low-light conditions. The scheme does not reveal any solutions to address the issue of unsatisfactory segmentation performance in low-light scenes. Summary of the Invention

[0009] In view of the above, the purpose of this invention is to provide a fast night sky segmentation method and apparatus based on deep learning, so as to solve the technical problem that the existing technology has unsatisfactory sky segmentation effect in scenes with low light intensity, thereby improving the accuracy and efficiency of night image segmentation.

[0010] In a first aspect, embodiments of the present invention provide a fast night sky segmentation method based on deep learning, comprising:

[0011] Obtain the preprocessed night sky image;

[0012] The night sky image is input into an illumination-enhanced image segmentation model for segmentation to obtain a segmented night sky image.

[0013] The illumination-enhanced image segmentation model includes an illumination enhancement network and a semantic segmentation network. The illumination enhancement network uses the night sky image as input for illumination enhancement, and the semantic segmentation network segments the night sky based on the output of the illumination enhancement network.

[0014] According to an embodiment of the present invention, the acquisition of the preprocessed night sky image includes:

[0015] A sky dataset is acquired, and a night sky dataset is obtained through a generative adversarial network, wherein the generative adversarial network converts daytime sky images in the sky dataset into nighttime sky data images;

[0016] Data augmentation operations are performed on a night sky dataset to obtain a preprocessed night sky image. The data augmentation operations include at least one of the following: random cropping, random rotation, random translation, and random brightness.

[0017] By employing generative adversarial networks, the nighttime dataset was effectively expanded, further enhancing the model's training performance and generalization capabilities.

[0018] According to an embodiment of the present invention, the illumination enhancement network is trained twice. The first training uses low-light images as samples to select the network with the best evaluation index, which is then used as part of the illumination enhancement image segmentation model for a second fine-tuning training.

[0019] According to an embodiment of the present invention, the preferred evaluation index for the optimal network is evaluated using peak signal-to-noise ratio and structural similarity image index.

[0020] According to an embodiment of the present invention, the semantic segmentation network adopts an encoder-decoder structure, wherein the encoder obtains a downsampled feature map by rapidly downsampling the night sky image;

[0021] The decoder includes a softmax layer; the decoder receives the downsampled feature map, and through the softmax layer predicts the feature map into two categories: sky and non-sky, and upsamples each category to the size of the night sky image.

[0022] According to an embodiment of the present invention, the encoder includes three depthwise separable convolutional modules and four depth feature extraction modules;

[0023] The depth-separable convolution module reduces the size of the encoder-based input by adjusting the depth and point directions.

[0024] The depth feature extraction module includes: a dilated convolution module and a depth-separable convolution module; the dilated convolution module, based on the encoder input, inserts holes between consecutive pixels to expand the kernel and increase the receptive field of the output feature map.

[0025] According to an embodiment of the present invention, the size of the sampled night sky image includes:

[0026] The predicted feature map is upsampled using bilinear interpolation;

[0027] Perform a stitching operation on feature maps of different resolutions.

[0028] According to an embodiment of the present invention, the illumination-enhanced image segmentation model is trained using a warm-up learning rate of 0.001 for 2000 steps.

[0029] Alternatively, the training process of the illumination-enhanced image segmentation model uses the sum of the cross-entropy loss function and the Dice loss function as the loss function during network training;

[0030] Alternatively, after the illumination-enhanced image segmentation model stabilizes, an initial learning rate of 0.01 is set, and the model is trained for 20,000 epochs, using a cosine annealing decay strategy.

[0031] Secondly, embodiments of the present invention also provide a fast night sky segmentation device based on deep learning, comprising:

[0032] The preprocessing module is used to obtain preprocessed night sky images;

[0033] The illumination-enhanced image segmentation module is used to input the night sky image into the illumination-enhanced image segmentation model for segmentation to obtain a segmented image of the night sky.

[0034] The illumination-enhanced image segmentation model includes an illumination enhancement network and a semantic segmentation network. The illumination enhancement network uses the night sky image as input for illumination enhancement, and the semantic segmentation network segments the night sky based on the output of the illumination enhancement network.

[0035] Thirdly, embodiments of the present invention also provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores a tagged ornament, and the processor executes the computer program to implement the steps of the above-described deep learning-based fast night sky segmentation method.

[0036] Fourthly, embodiments of the present invention also provide a computer-readable medium having a computer program stored thereon, characterized in that the computer program, when processed and executed, implements the steps of the above-described deep learning-based fast night sky segmentation method.

[0037] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0038] This invention constructs an image segmentation model for nighttime skies, which can quickly and accurately distinguish between sky and non-sky regions in nighttime environments, effectively solving the problem of sky segmentation in poor lighting conditions and significantly improving segmentation efficiency.

[0039] The illumination enhancement network constructed in this invention utilizes a lightweight illumination enhancement network to preprocess nighttime scenes, effectively improving image quality and thus enhancing segmentation accuracy.

[0040] This invention has wide applicability and can be applied to various night scene segmentation scenarios, thus having significant practical value. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating a fast night sky segmentation method based on deep learning provided in an embodiment.

[0043] Figure 2 This is a schematic diagram of a lightweight semantic segmentation network model with an encoder-decoder structure provided in the embodiment;

[0044] Figure 3 This is a semantic segmentation prediction result diagram provided in the embodiment;

[0045] Figure 4 This is a structural block diagram of a fast night sky segmentation device based on deep learning provided in the embodiment;

[0046] Figure 5 This is a schematic diagram of the structure of the electronic device provided in the embodiment. Detailed Implementation

[0047] 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 of the invention and do not limit the scope of protection of this invention.

[0048] like Figure 1 As shown in the figure, an embodiment of the present invention provides a fast night sky segmentation method based on deep learning, comprising:

[0049] S101, Obtain the preprocessed night sky image.

[0050] Obtain a segmented dataset containing the sky, and use an adversarial generative network model to convert the daytime sky dataset into a nighttime sky dataset. After preprocessing, construct the original dataset containing nighttime sky images.

[0051] Specifically, through the following steps:

[0052] S1011, obtain the sky dataset, and obtain the night sky dataset through an adversarial generative network.

[0053] In this embodiment, two public datasets, ADE20K and SkyFinder, were used.

[0054] ADE20K is a large-scale scene parsing dataset containing over 20,000 images, covering a variety of complex scenes, and providing detailed annotations for each pixel, including annotations for objects and parts.

[0055] SkyFinder is a dataset specifically designed for sky detection and segmentation. It contains a large number of outdoor images with precise annotations of the sky regions within the images. While both ADE20K and SkyFinder are segmentation datasets that include sky data, covering various daytime and nighttime scenes, the dataset contains relatively few nighttime scenes, with the majority consisting of daytime images.

[0056] Therefore, in this embodiment, the daytime sky dataset is converted into the nighttime sky dataset using an adversarial generative network model and then merged into the original nighttime dataset.

[0057] The steps to generate a night sky dataset are as follows:

[0058] First, daytime and nighttime sky datasets are input separately into a generative adversarial network (GAN) for training, resulting in a generator that converts daytime images into nighttime images. Second, using the ADE20K and SkyFinder datasets as raw data, images containing sky elements are input into the generator to produce corresponding nighttime sky datasets at the same resolution. The initial labels are used as the labels for these nighttime datasets. Finally, the generated nighttime sky datasets are merged into the original nighttime datasets to obtain the final nighttime sky dataset.

[0059] S1012, perform data augmentation on the night sky dataset to obtain a preprocessed night sky dataset.

[0060] This embodiment employs a series of data augmentation operations: random cropping, random rotation, random translation, and random brightness. The purpose of this step is to improve the generalization ability of the model.

[0061] S1013 divides the preprocessed night sky dataset into training, validation, and test datasets in a ratio of 7:2:1.

[0062] S102, the night sky image is input into the illumination-enhanced image segmentation model for segmentation to obtain a segmented image of the night sky.

[0063] The illumination-enhanced image segmentation model includes:

[0064] The system comprises an illumination enhancement network and a semantic segmentation network. The illumination enhancement network takes the night sky image as input for illumination enhancement, and the semantic segmentation network segments the night sky based on the output of the illumination enhancement network.

[0065] Specifically, through the following steps:

[0066] S1021 enhances nighttime scenes using a lightweight lighting enhancement network, preserving the best-performing model.

[0067] First, a lightweight illumination enhancement network is designed. This end-to-end illumination enhancement network is used to learn high-level features of an image, such as illumination.

[0068] Secondly, the low-light and normal image datasets from the public datasets LOL and MIT-Adobe FiveK were used as training sets and input into the designed illumination enhancement network for training, resulting in a model that enhances nighttime images into well-lit images.

[0069] Finally, based on the image evaluation metrics PSNR (Peak Signal-to-Noise Ratio) and SSIM (Structural Similarity), the model with the best enhancement effect under the experimental conditions was obtained.

[0070] The formula for calculating the peak signal-to-noise ratio is as follows:

[0071]

[0072] In the formula, MAX is the maximum possible pixel value of the image, which is 255 here. MSE represents the mean squared error. Given two images A and B, with pixel values ​​A(i,j) and B(i,j) respectively, and the image size is M×N, then the MSE is calculated as follows:

[0073]

[0074] The formula for calculating the structural similarity is as follows:

[0075]

[0076] In the formula, A and B represent two input images, μ A and μ B It is the average pixel value of images A and B. and σ is the variance of images A and B. AB SSIM is the pixel covariance between images A and B. c1 and c2 are two constants used to increase numerical stability. The value of SSIM ranges from 0 to 1. The closer the value is to 1, the higher the structural similarity of the images.

[0077] S1022, design the deep feature extraction module and backbone network.

[0078] The deep feature extraction module of this embodiment is designed with reference to existing literature (Wang X, Lv Q, Chen G, Zhang J, Wei Z, Dong J, Fu H, Zhu Z, Liu J, Jin X. MobileSky: Real-Time Sky Replacement for Mobile AR. IEEE Trans VisComput Graph. 2023 Mar 20; PP. doi:10.1109 / TVCG.2023.3257840. Epub ahead of print. PMID:37030763.).

[0079] To capture greater spatial correlation, the number of parameters is reduced by utilizing dilated convolution modules and depthwise separable convolution modules.

[0080] Specifically, in this embodiment, the depth feature extraction module includes a dilated convolution module and a depthwise separable convolution module. The dilated convolution expands the kernel by inserting holes between its consecutive elements, thereby increasing the receptive field of the feature map. The depthwise separable convolution module reduces the size of the encoder-based input by adjusting the depth and point directions.

[0081] The backbone network based on the above modular design model combines ordinary convolution, dilated convolution, and depthwise separable convolution, enabling full channel fusion without increasing computational cost.

[0082] S1023 is a lightweight semantic segmentation network model with an encoder-decoder structure built based on a deep feature extraction module.

[0083] like Figure 2 As shown, a lightweight semantic segmentation network model with an encoder-decoder structure is constructed based on a deep feature extraction module. The overall network structure includes an encoder and a decoder.

[0084] The encoder performs rapid downsampling on the feature map.

[0085] First, three depthwise separable convolutional modules are used to reduce the size of the feature map to 1 / 8 of the original input image, where the kernel size of the depthwise separable convolutions is 7×7. Second, four depth feature extraction modules are used to extract features from the night sky and output downsampled feature maps. The decoder receives feature maps with different resolutions and fuses them together before upsampling them to the original size.

[0086] Finally, the decoder uses a softmax layer for final prediction, with each pixel having two categories: sky and non-sky.

[0087] To keep the network lightweight, only bilinear interpolation is used to upsample the 1 / 8 prediction results. The network also includes a feature map concatenation operation at different resolutions, which can fuse features of different sizes without losing information.

[0088] S1024, Obtain the complete illumination-enhanced image segmentation model.

[0089] An illumination-enhanced image segmentation model is constructed based on the best illumination enhancement network in S1021 and the lightweight semantic segmentation network model in S1023. The illumination-enhanced image segmentation model is trained using the preprocessed training set in S1013.

[0090] Specifically, in this embodiment, firstly, in the initial stage, the preprocessed training set in S1013 is used to train the illumination-enhanced image segmentation model. The training is performed using a warm-up learning rate of 0.001 for 2000 steps.

[0091] Secondly, the training process uses the sum of the cross-entropy loss function L and the Dice loss function as the loss function during network training.

[0092] The cross-entropy loss function is calculated in detail as follows:

[0093]

[0094] In the formula, N represents the number of samples, here the number of night sky images used for training, p i p represents the predicted probability of the i-th category. j p represents the total predicted probability for all categories before applying the softmax layer. j Not 1.

[0095] The Dice loss function is calculated as follows:

[0096]

[0097] In the formula, pred is the set of predicted values, true is the set of true values, pred∩true represents the intersection of predicted values ​​and true values, and pred∪true represents the union of predicted values ​​and true values.

[0098] Finally, after the model stabilizes, the initial learning rate is set to 0.01, and the model is trained for 20,000 epochs. A decay strategy using cosine annealing is employed to "escape" local minima and find a path to the global minimum.

[0099] S1025, Evaluate the illumination-enhanced image segmentation model based on the mIoU value.

[0100] mIoU is an important metric used to evaluate model accuracy in the field of image segmentation in computer vision and deep learning. It represents the mean intersection-over-union ratio, which is calculated by taking the IoU value of each class (the number of true samples divided by the number of true samples plus the number of false negative samples plus the number of false positive samples), and then averaging the IoU values ​​of all classes.

[0101] mIoU plays a crucial role in evaluating the quality and accuracy of semantic segmentation results for night skies. By calculating the intersection-union ratio (IU) for each class and averaging the IU, mIoU measures the segmentation accuracy of the model on each class and provides a comprehensive evaluation metric. A higher value indicates a more accurate segmentation result, and its value ranges from 0 to 1.

[0102] Specifically, the indicator is defined as follows:

[0103]

[0104] In the formula, p ij This represents the predicted number of pixels belonging to class i within class j. In other words, for a specific class i, p ii p ij and p ji These represent true positives, false positives, and false negatives, respectively. k is the number of categories. In this embodiment of the invention, the sky segmentation network divides pixels into two categories, with k = 2.

[0105] Based on the above operations, train the network model on the PC, input the test image for forward inference, and output the prediction result.

[0106] like Figure 3 As shown, the input night scene image and semantic segmentation prediction results of the model demonstrate that the classification results predicted by the method of this embodiment have good performance, high prediction accuracy, and an mIoU improvement of about 12% compared to a simple semantic segmentation network.

[0107] This invention employs a generative adversarial network model to transform daytime sky datasets into nighttime sky datasets to enhance existing datasets. Furthermore, an illumination enhancement network improves image quality, enhancing brightness, contrast, and color to improve visibility and clarity. Preprocessing the images before they enter the semantic image segmentation model provides better input for subsequent semantic image segmentation tasks.

[0108] Based on the same inventive concept, embodiments also provide a deep learning-based fast night sky segmentation device 400.

[0109] like Figure 4 As shown, there is a preprocessing module 410 and an illumination-enhanced image segmentation module 420.

[0110] The preprocessing module 410 is used to acquire a preprocessed night sky image;

[0111] The illumination-enhanced image segmentation module 420 is used to input the night sky image into the illumination-enhanced image segmentation model for segmentation to obtain a segmented image of the night sky.

[0112] It should be noted that the above-described embodiment of a deep learning-based fast night sky segmentation device should be illustrated using the above-described functional module division as an example when performing fast night sky segmentation. The functions can be assigned to different functional modules as needed, i.e., the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the above-described embodiment of a deep learning-based fast night sky segmentation device and a deep learning-based fast night sky segmentation method embodiment belong to the same concept; their specific implementation process is detailed in a deep learning-based fast night sky segmentation method embodiment, and will not be repeated here.

[0113] Based on the same inventive concept, the embodiment also provides a computing device, such as... Figure 5 As shown, at the hardware level, this deep learning-based fast night sky segmentation device includes a processor, internal bus, network interface, memory, and storage, and may also include other hardware required for other operations. The processor reads the corresponding computer program from the storage into the memory and then runs it to implement the aforementioned deep learning-based fast night sky segmentation method, including:

[0114] S101, Obtain the preprocessed night sky image;

[0115] S102, the night sky image is input into the illumination-enhanced image segmentation model for segmentation to obtain a segmented image of the night sky.

[0116] The memory can be volatile memory at the near end, such as RAM, or non-volatile memory, such as ROM, FLASH, floppy disk, hard disk, etc., or even a remote storage cloud. The processor can be a central processing unit (CPU), microprocessor (MPU), digital signal processor (DSP), or field-programmable gate array (FPGA), meaning that the steps of the deep learning-based fast night sky segmentation method can be implemented using these processors.

[0117] Based on the same inventive concept, the embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when processed and executed, implements the above-described deep learning-based fast night sky segmentation method, including:

[0118] S101, Obtain the preprocessed night sky image;

[0119] S102, the night sky image is input into the illumination-enhanced image segmentation model for segmentation to obtain a segmented image of the night sky.

[0120] The computer-readable storage medium can be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0121] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based fast night sky segmentation method, characterized in that, The method comprises the following steps: obtaining a pre-processed night sky picture; inputting the night sky picture into an illumination-enhanced image segmentation model for segmentation to obtain a segmented picture of the night sky, wherein the illumination-enhanced image segmentation model comprises an illumination enhancement network and a semantic segmentation network, the illumination enhancement network takes the night sky picture as input for illumination enhancement, the illumination enhancement network is trained twice, the first training uses low-illumination pictures as samples, and an optimal network with optimal evaluation indicators is used as part of the illumination-enhanced image segmentation model for secondary fine-tuning training; the semantic segmentation network performs night sky segmentation based on the output of the illumination enhancement network, the semantic segmentation network adopts an encoder-decoder structure, the encoder obtains a down-sampled feature map by rapidly down-sampling the night sky picture; the decoder comprises a softmax layer; the decoder receives the down-sampled feature map, predicts the feature map into two categories of sky and non-sky through the softmax layer, and respectively up-samples to the size of the night sky picture; the encoder comprises three depth separable convolution modules and four depth feature extraction modules; the depth separable convolution module reduces the size of the input based on the encoder through the depth direction and the point direction; the depth feature extraction module comprises a hole convolution module and a depth separable convolution module; the hole convolution module inserts holes between consecutive pixels based on the input of the encoder to expand the kernel and increase the receptive field of the output feature map; the illumination-enhanced image segmentation model is trained in a warm-up preheating learning rate mode, the learning rate is set to 0.001, and 2000 steps are trained; or, the sum of the cross-entropy loss function and the dice loss function is used as the loss function during network training in the training process of the illumination-enhanced image segmentation model; or, after the model is stable, the initial learning rate is set to 0.01, 20000 epochs are trained, and a cosine annealing strategy is used for the decay strategy.

2. The deep learning based fast night sky segmentation method of claim 1, wherein, The pre-processed night sky picture is obtained by: obtaining a sky data set and obtaining a night sky data set through a generative adversarial network, wherein the generative adversarial network converts daytime sky images in the sky data set into night sky data images; performing data enhancement operations on the night sky data set to obtain pre-processed night sky pictures, wherein the data enhancement operations include at least one of the following: random cropping, random rotation, random translation, and random brightness. 3.The deep learning based fast night sky segmentation method of claim 1, wherein, The optimal network with optimal evaluation indicators is evaluated using the peak signal-to-noise ratio and the structural similarity image index.

4. The deep learning based fast night sky segmentation method of claim 1, wherein, The up-sampling to the size of the night sky picture comprises: using bilinear interpolation to up-sample the predicted feature map; performing a stitching operation on feature maps of different resolution sizes. 5.A deep learning based device for fast night sky segmentation, characterized in that, The method for realizing the deep learning-based night fast sky segmentation method of any one of claims 1-4 comprises: a preprocessing module for obtaining a pre-processed night sky picture; The image segmentation module under light enhancement is used for inputting the night sky picture into the image segmentation model under light enhancement for segmentation to obtain a segmented picture of the night sky, The image segmentation model under light enhancement comprises a light enhancement network and a semantic segmentation network, the light enhancement network performs light enhancement on the night sky picture as input, and the semantic segmentation network performs night sky segmentation based on the output of the light enhancement network.

6. A computing device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein, The memory has a jewelry with a label, and the processor executes the computer program to realize the steps of the deep learning-based night fast sky segmentation method in any one of claims 1-4.

7. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to realize the steps of the deep learning-based night fast sky segmentation method in any one of claims 1-4.

Images