Light thorn identification method adaptive to vehicle-mounted CMS and storage medium

Abstract

The invention relates to the technical field of image recognition, and provides a light thorn recognition method adaptive to a vehicle-mounted CMS and a storage medium, a lightweight detection model is constructed based on consideration of board end chip computing power and a parallel architecture, requirements for light thorn detection precision and a lightweight neural unit SFB, multi-dimensional feature information of light thorns can be deeply excavated, and the detection accuracy of the light thorns can be improved. The recognition accuracy is improved; the device is light enough, can adapt to the characteristic of low computing power of a vehicle end chip, and makes a response within a given time; intelligent post-processing operation is explored, and dependence on a three-party library is reduced as much as possible.

Description

Technical Field

[0001] The present invention relates to the field of image recognition technology, and in particular to a light thorn recognition method and storage medium adapted to a vehicle-mounted CMS. Background Art

[0002] The vehicle-mounted CMS (Camera-Monitoring System) is a vehicle-mounted camera monitoring system, which is a system that uses camera technology to monitor the vehicle's surrounding environment.

[0003] During the driving process, if the CMS external camera is dirty, oily, raindrops, etc., there will be obvious light spikes in the image fed back at night. At this time, even if the driver directly views the video, it is not easy to detect. Therefore, it is necessary to use the light spike recognition function to prompt the CMS that there is dirt and remind the driver to wipe the lens to reduce the occurrence of light spikes. The light spike recognition function is usually composed of traditional image algorithms, such as binarization and histogram equalization to distinguish pixel differences.

[0004] The detection accuracy of light thorn detection algorithms based on traditional image algorithms is generally lower than that of light thorn detection algorithms based on convolutional neural networks. In addition, a lot of post-processing is required to perform template matching operations. Taking the binary light thorn detection method as an example, the collected image needs to be divided into multiple areas and binarized. If the pixel value of the area where the bright light is located is too high, the area is detected as a light thorn. If the proportion of the number of light thorn areas to the number of areas divided in the entire image is greater than a certain set threshold, it means that there are light thorns in the image.

[0005] In the feature extraction stage of light thorn recognition using traditional image algorithms, due to the relatively simple and limited feature descriptions used by traditional algorithms, they can only capture relatively obvious and superficial information in the image. It is difficult to fully and accurately describe the complex and ever-changing shapes, textures, and subtle differences between light thorns and the surrounding environment. When there are slight grayscale changes or blurred edges in light thorns, traditional algorithms may not be able to accurately extract the complete light thorn features, resulting in inaccurate recognition. That is, the traditional method has limited performance in light thorn detection, as follows:

[0006] 1. The detection capability of tiny light thorns is limited. This method mainly judges whether light thorns exist based on the pixel value and the proportion of the area where the light thorns are located. However, it may not be able to effectively detect some tiny light thorns with small size and no obvious pixel value changes. Because the area occupied by these tiny light thorns is small, their pixel values ​​have little effect on the statistical results of the entire area and are easily ignored.

[0007] 2. Affected by ambient light and noise, in the actual image acquisition process, the unevenness of ambient light and the interference of various noises are difficult to avoid. Changes in ambient light may cause inconsistent brightness in different areas of the image, thus affecting the effect of binarization and the detection results of light thorns. At the same time, noise may cause some isolated bright spots or dark spots to appear in the image. These noise points may be misjudged as light thorns after binarization or affect the accurate judgment of the light thorn area, reducing the accuracy and reliability of detection.

[0008] Therefore, when faced with different lighting conditions (such as direct strong light, low light environment, or uneven lighting scenes), key parameters such as the binarization threshold of the traditional algorithm are difficult to adjust adaptively, resulting in large deviations in the light thorn recognition results. At the same time, if there is noise interference in the image, whether it is Gaussian noise, salt and pepper noise, or other types of noise, it is very easy for traditional algorithms to make misjudgments. For example, noise points may be mistakenly identified as light thorns, or the boundaries and shapes of light thorns may be severely distorted, thereby affecting subsequent analysis and judgment.

[0009] 3. Lack of comprehensive consideration of spatial information. Only the proportion of light thorn areas is used to judge whether there are light thorns in the image, ignoring the spatial distribution information of light thorns in the image. In fact, spatial information such as the distribution position and concentration degree of light thorns is also very important for accurately judging the light thorn situation in the image. For example, even if the proportion of light thorn areas does not reach the set threshold, if these light thorns are concentrated in the key areas of the image, it may still have a significant impact on the image quality or subsequent processing, and this method cannot effectively capture this spatial distribution feature. Summary of the invention

[0010] The present invention provides a light thorn recognition method and storage medium adapted to a vehicle-mounted CMS, which solves the technical problems of large computational complexity, poor anti-interference ability, low light thorn detection accuracy, inability to efficiently detect light thorns, and low detection accuracy of existing light thorn detection algorithms.

[0011] In order to solve the above technical problems, the present invention provides a light thorn recognition method adapted to a vehicle-mounted CMS, comprising:

[0012] Capturing images of the on-site environment to obtain on-site images;

[0013] Marking the target area in the scene image, and performing image preprocessing on the marked scene image, wherein the image preprocessing includes scale scaling;

[0014] Creating a lightweight neural unit SFB, and building a light thorn detection model based on the neural unit SFB;

[0015] The scene image is input into the light thorn detection model to perform light thorn detection, and the detection result is output.

[0016] In a further embodiment, model training is also included before performing optical thorn recognition, and the model training includes:

[0017] Capturing images of different scene environments to obtain view images;

[0018] Annotating a target area in the view image, and performing a scale scaling process on the annotated view image;

[0019] generating a sample set based on the processed view image;

[0020] A training set and a validation set are created based on the view images in the sample set, and model training is performed on the optical thorn detection model according to the training set and the validation set.

[0021] This solution collects images from different scene environments to obtain view images for model training. Images from different angles can provide more samples, increase data diversity, help the model better learn and understand different features, and improve model accuracy.

[0022] In a further embodiment, the marking of the target area comprises:

[0023] Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn, and obtain a first region;

[0024] Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn to obtain a second region;

[0025] An overlapping area between the first area and the second area corresponds to an area where the light source is located.

[0026] This solution takes the light source that generates the light thorn as the corner point, and marks the regions of interest for the light thorn areas extending on both sides, accurately locates the light thorn area, reduces the calculation and processing of irrelevant areas, and reduces the amount of calculation, thereby significantly improving the processing speed and accuracy.

[0027] In a further embodiment, the image preprocessing comprises:

[0028] Acquire the annotated scene image, and scale the scene image based on a predefined standard; perform image extraction on the scaled scene image, acquire corresponding image Y component information from the scene image based on the YCbCr color model, and use the image Y component information as image data of the scaled scene image.

[0029] This solution scales and preprocesses the collected images. On the premise of fully retaining the image feature information, it directly reduces the data size of the on-site images to adapt to the real-time requirements of the low-computing-power board-side on-board chips.

[0030] In a further embodiment, the image preprocessing also includes image enhancement processing, and the image enhancement processing includes at least one of image blurring, mosaic enhancement, and random erasing, and the blur kernel size of the image blurring is 5×5.

[0031] Before the light thorn detection model performs light thorn detection, this solution performs image enhancement processing such as image blurring, mosaic enhancement, and random erasing on the input image to perform data expansion. By increasing the diversity and quantity of training data, data enhancement can help the model generalize better, reduce overfitting, and thus improve the robustness of the algorithm.

[0032] In a further implementation scheme, a lightweight neural unit SFB is created, including: forming a lightweight neural unit SFB according to a depth group convolution Depth Group Conv, a batch normalization layer BatchNorm Layer, a constant level activation function ReLU, and a channel shuffle unit shuffle channel.

[0033] This solution inserts some neural activation functions into the original Shuffle block structure, sets the Depth Group Conv, the BatchNorm Layer, the constant level activation function ReLU, and the shuffle channel to form a lightweight neural unit SFB, which can reduce the SIZE of the model itself to adapt to the low-computing-power board-end vehicle chip, and further effectively improve the model accuracy, thereby improving the accuracy of optical thorn recognition.

[0034] In a further embodiment, constructing a light thorn detection model based on the neural unit SFB includes:

[0035] Building a backbone network based on the neural unit SFB, the backbone network is used to extract feature information from the scene image;

[0036] A Neck layer is constructed using a PANet structure, and the Neck layer is used to perform multi-scale fusion on the input feature information to obtain output features;

[0037] Setting a model decoupling head to detect the light thorns according to each of the input output features to obtain detection data;

[0038] The detection data is input into a preset loss function to implement gradient backpropagation for parameter update to obtain a trained convolutional neural network.

[0039] In a further embodiment, a backbone network is constructed based on the neural unit SFB, comprising:

[0040] A convolutional layer Stardard Conv, a batch normalization layer BatchNorm Layer, and a Relu activation layer form a convolution unit CBR;

[0041] After the input layer, one layer of convolutional unit CBR is sequentially arranged, and after the convolutional unit CBR, six layers of neural unit SFB are arranged;

[0042] An SPP layer is set at the end of the neural unit SFB to form a feature extraction network, wherein the SPP layer uses an Add operation to add multi-scale layers.

[0043] The present invention also provides a storage medium on which a computer program is stored, and the computer program is used to implement the above-mentioned image processing method for improving the accuracy of optical thorn detection. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0044] Based on the consideration of the computing power and parallel architecture of the board-side chip and the requirement for the accuracy of light thorn detection, the present invention constructs a lightweight detection model based on the lightweight neural unit SFB, which can not only deeply mine the multi-dimensional feature information of the light thorn to improve the accuracy of recognition; but also is lightweight enough to adapt to the low computing power of the vehicle-side chip and respond within a given time; explores intelligent post-processing operations to minimize dependence on third-party libraries. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a workflow diagram of a light thorn recognition method adapted to a vehicle-mounted CMS provided by an embodiment of the present invention;

[0046] Figure 2 is a scene collection requirement diagram provided by an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of optical thorn marking provided by an embodiment of the present invention;

[0048] Figure 4 is a single-channel image of the Y component of an image provided by an embodiment of the present invention;

[0049] Figure 5 is a schematic diagram of the SFB structure provided by an embodiment of the present invention;

[0050] Figure 6Another SFB structure diagram provided by an embodiment of the present invention is

[0051] Figure 7 is a comparative performance table of SFB algorithms provided by an embodiment of the present invention;

[0052] Figure 8 It is a schematic diagram of the operation of lightweight pyramid pooling SPPA provided by an embodiment of the present invention;

[0053] Fig. 9 is a schematic diagram of the network structure of the optical thorn detection model provided by an embodiment of the present invention;

[0054] Fig.10 is a comparative performance table of algorithms of the optical thorn detection model provided by the embodiment of the present invention;

[0055] Fig.11 This is a random scene test effect diagram provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0057] Example 1

[0058] The embodiment of the present invention provides a light thorn recognition method adapted to a vehicle-mounted CMS, such as Figures 1 to 11 As shown, in this embodiment, steps S0 to S4 are included:

[0059] S0. Model training, wherein the model training includes:

[0060] Capturing images of different scene environments to obtain view images;

[0061] Annotating a target area in the view image, and performing a scale scaling process on the annotated view image;

[0062] generating a sample set based on the processed view image;

[0063] A training set and a validation set are created based on the view images in the sample set, and model training is performed on the optical thorn detection model according to the training set and the validation set.

[0064] The view images are based on a self-built library, including several front views, each with a resolution of 1280×800, and capture typical indoor and outdoor scenes. Figure 2,Surround view images at different time periods and weather conditions.

[0065] This solution collects images from different scene environments to obtain view images for model training. Images from different angles can provide more samples, increase data diversity, help the model better learn and understand different features, and improve model accuracy.

[0066] S1, collecting images of the on-site environment to obtain on-site images;

[0067] S2, marking the target area in the scene image, and performing image preprocessing on the marked scene image, wherein the image preprocessing includes scale scaling;

[0068] In this embodiment, see Figure 3 , the target area annotation includes:

[0069] Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn, and obtain a first region;

[0070] Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn to obtain a second region;

[0071] An overlapping area between the first area and the second area corresponds to an area where the light source is located.

[0072] Among them, the annotation tool is used to annotate the area of ​​interest; Figure 3 The detection frame where the upper marked light is located is recorded as frame A, that is, the first area; the frame where the lower marked light is located is recorded as frame B, that is, the first area; the intersection of frames AB is the light source area (such as car lights).

[0073] This solution takes the light source that generates the light thorn as the corner point, and marks the regions of interest for the light thorn areas extending on both sides, accurately locates the light thorn area, reduces the calculation and processing of irrelevant areas, and reduces the amount of calculation, thereby significantly improving the processing speed and accuracy.

[0074] In this embodiment, the image preprocessing includes:

[0075] Obtain the annotated scene image, and perform a size scaling process on the scene image based on a predefined standard; see Figure 4 , performing image extraction on the scene image after the size scaling process, obtaining corresponding image Y component information from the scene image based on the YCbCr color model, and using the image Y component information as image data of the scene image after the size scaling process.

[0076] For example, the predefined standard is set to a width and height of 640×352. For the light thorn scene, the most obvious feature is the brightness information, so only the image Y component information needs to be extracted for detection. Through the above two operations, it can ensure that certain feature information can be retained after 32 times downsampling, and the real-time requirements of the on-board chip on the low-computing board can be met.

[0077] This embodiment performs scale scaling and preprocessing on the collected images, and directly reduces the data size of the on-site images while fully retaining the image feature information, so as to adapt to the real-time requirements of the low-computing-power board-side on-board chip.

[0078] In this embodiment, the image preprocessing also includes image enhancement processing, and the image enhancement processing includes at least one of image blurring, mosaic enhancement, and random erasing, and the blur kernel size of the image blurring is 5 × 5. Preferably, this embodiment adopts the three-in-one data enhancement of "image blurring + mosaic enhancement + random erasing".

[0079] In this embodiment, before the light thorn detection model performs light thorn detection, image enhancement processing such as image blurring, mosaic enhancement, and random erasing is performed on the input image to perform data expansion. By increasing the diversity and quantity of training data, data enhancement can help the model generalize better, reduce overfitting, and thus improve the robustness of the algorithm.

[0080] S3, creating a lightweight neural unit SFB, and building a light thorn detection model based on the neural unit SFB;

[0081] In this embodiment, a lightweight neural unit SFB is created, including: a lightweight neural unit SFB is formed according to a depth group convolution Depth GroupConv, a batch normalization layer BatchNorm Layer, a constant level activation function ReLU, and a channel shuffle unit shufflechannel.

[0082] Through experiments, it is found that inserting some neural activation functions (such as the constant level activation function ReLU in this embodiment) into the original Shuffle block structure can effectively improve the accuracy of the model. However, the network is mainly deployed on the CPU side of the vehicle chip, and the inserted activation function needs to consider the computational complexity. The neural unit SFB is the smallest structural unit of the model of this solution.

[0083] If the stride of the deep group convolution is 1, see Figure 5 , then the convolution structure of SFB is as follows;

[0084] The output of the Channel Split operation module is divided into two branches, one of which is directly connected to the Concat operation module; the other branch is sequentially connected to a standard convolution layer Conv, a depth group convolution Depth GroupConv, a batch normalization layer BatchNorm Layer, a constant level activation function ReLU, a standard convolution layer Conv, and finally connected to the Concat operation module; the output of the Concat operation module is connected to the input of the Channel Shuffle operation module.

[0085] When the stride of the depth group convolution is 2, see Figure 6 , then the convolution structure of SFB is as follows;

[0086] The input end is directly divided into two branches. One of the branches is connected to the Concat operation module after passing through 1 deep group convolution Depth GroupConv, 1 constant level activation function ReLU, and 1 standard convolution layer. This branch contains a deep convolution layer with a step size of 2; the other branch is connected to 1 standard convolution layer Conv, 1 deep group convolution DepthGroup Conv, 1 batch normalization layer BatchNorm Layer, 1 constant level activation function ReLU, 1 standard convolution layer Conv in sequence, and finally connected to the Concat operation module; the output of the Concat operation module is connected to the input of the Channel Shuffle operation module.

[0087] Among them, the deep group convolution Depth Group Conv and the standard convolution layer Conv constitute the DW Conv deep separable convolution.

[0088] This embodiment inserts some neural activation functions into the original Shuffle block structure, and sets the Depth Group Conv, the BatchNorm Layer, the constant level activation function ReLU, and the shuffle channel to form a lightweight neural unit SFB, which can reduce the SIZE of the model itself to adapt to the low-computing-power board-end vehicle chip, and further effectively improve the model accuracy, thereby improving the optical thorn recognition accuracy.

[0089] At the same time, the computing power of the on-board chip CPU is far less than that of NPU, DSP, BPU and other acceleration units. When designing the network backbone, it is taken into account that the on-board CPU can use SIMD multi-data parallel acceleration when performing matrix operations. By calling the float16x8_t register with a bandwidth of 8, 8-bit elements are calculated at one time. Therefore, the network backbone starts from the first layer and uses 8, 16, 32, 64 bits and other bidth with eight-bit aligned channels. In addition, since the original intention of the ShuffleNetv2 network design was to be used for accuracy rankings of large data sets such as ImageNet, but in actual engineering, it is not necessary to use large data scenarios such as ImageNet with 1000 categories. Based on this purpose, multiple experiments were conducted, and the final backbone network was determined as follows Fig. 9 shown.

[0090] Among them, the algorithm performance corresponding to the backbone network is as follows Figure 7 As shown in the list, Output Size is the dimension or size of the model output. The length of the input sequence is extended by Repeat operation, such as Figure 7 and Fig. 9 As shown, the Repeat of the six neural unit SFBs is 1:3:1:3:1:1. Output Channel The number of channels of the feature map output after the convolution operation.

[0091] In this embodiment, see Fig. 9 , constructing a light thorn detection model based on the neural unit SFB includes:

[0092] A1, constructing a backbone network based on the neural unit SFB, wherein the backbone network is used to extract feature information from the scene image;

[0093] A2. A Neck layer is constructed using a PANet structure. The Neck layer is used to perform multi-scale fusion on the input feature information to obtain output features.

[0094] Among them, the Neck part adopts the PANet structure. This structure establishes a new bottom-up path based on FPN to improve the positioning information of the low-resolution feature map. This structure has a certain improvement effect on the position information and positioning of the light thorn.

[0095] A3. Setting a model decoupling head, and detecting the light thorns according to each of the input output features to obtain detection data;

[0096] A4. Input the detection data into a preset loss function to implement gradient backpropagation for parameter update to obtain a trained convolutional neural network.

[0097] Among them, this embodiment uses EIoU as the regression loss function. Most modern object detectors will use an optimized IoU variant to replace the distance-based loss for better detection of bounding boxes. EIoU splits the aspect ratio loss term into the difference between the predicted width and height and the minimum external frame width and height, which accelerates the convergence of the predicted frame and improves the regression accuracy of the predicted frame. The EIoU regression loss function formula is as follows:

[0098]

[0099] In this embodiment, a backbone network is constructed based on the neural unit SFB, including:

[0100] A convolutional layer Stardard Conv, a batch normalization layer BatchNorm Layer, and a Relu activation layer form a convolution unit CBR;

[0101] After the input layer, one layer of convolutional unit CBR is sequentially arranged, and after the convolutional unit CBR, six layers of neural unit SFB are arranged;

[0102] See also Figure 8 An SPP layer is set at the end of the neural unit SFB to form a feature extraction network, wherein the SPP layer uses an Add operation to add multi-scale layers.

[0103] In the past, the SPP layer usually used a large convolution pooling layer to combine the features with the Concat operation to achieve the fusion of multi-category and multi-scale features. The formula is as follows:

[0104]

[0105] The Add operation is to superimpose the semantic information extracted multiple times, which will highlight the proportion of correct classification, which is beneficial to the final classification and achieves high activation of correct classification. The formula is as follows:

[0106] Among them, the superposition formula of the Add operation is as follows:

[0107]

[0108] This superposition formula is common knowledge in the art and will not be described in detail in this embodiment.

[0109] At present, some papers have pointed out that the Concat operation is more suitable for samples with large differences in category features. In the general scenario of detecting light spikes, since most light spikes are ray-shaped and the sample feature differences are small, superimposing feature channels through the Concat operation will increase the amount of calculation. In the case of small differences in category feature information, there is no difference between the Add operation and the Concat operation, and the amount of calculation is not increased. The multi-scale pyramid pooling operation SPPA that uses the Add operation to add layers is as follows: Figure 4 As shown, the MaxPool convolution kernel size is 3×3.

[0110] S4. Input the scene image into the light thorn detection model to perform light thorn detection, and output the detection result.

[0111] All experiments in this embodiment use recall and precision as evaluation indicators.

[0112] The recall rate Recall refers to the original labeled samples. It indicates how many positive examples in the samples are predicted to be correct. The original positive class is predicted as the positive class (TP). Another way is to predict the original positive class as the negative class (FN). The formula is as follows.

[0113] Recall=TP / (TP+FN)

[0114] Precision represents the overall prediction accuracy, including positive and negative samples. The denominator is the predicted positive class. The precision is proposed to make the existing prediction results of the model as error-free as possible, where FP is the wrong positive sample, that is, predicting the negative sample as a positive sample. The formula is as follows.

[0115] Precision=TP / (TP+FP)

[0116] Lightweight detection algorithms such as YOLOv5-Lite, NanoDet, and FastestDet are used for training and testing based on a self-built optical thorn database, and the recall rate and precision rate of the algorithm in this embodiment are compared.

[0117] See also Fig.10 Although NanoDet and YOLOv5-Lite have high training accuracy in the light thorn dataset, the inference time is also relatively high. Although the precision and recall of other lightweight algorithms are lower than the first two algorithms, the inference time is greatly reduced. At the same time, the model size of this embodiment is only 167KB, but the lightweight detection algorithm proposed in this embodiment has a much smaller number of parameters than the other three detection algorithms, and the accuracy has not dropped significantly; the inference time (Latency) is also the smallest, only 97ms, and the calculation rate is greatly improved.

[0118] See also Fig.11 When the random scene test is performed in this embodiment, the light thorn area can be accurately identified in different scenes.

[0119] Example 2

[0120] The embodiment of the present invention further provides a storage medium on which a computer program is stored, and the computer program is used to implement an image processing method for improving the accuracy of optical thorn detection as described in the above embodiment 1. The storage medium may be a magnetic disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0121] The embodiment of the present invention is based on the consideration of the computing power and parallel architecture of the board-side chip, as well as the requirement for the accuracy of light thorn detection. Based on the lightweight neural unit SFB, a lightweight detection model is constructed, which can not only deeply mine the multi-dimensional feature information of the light thorn to improve the accuracy of recognition; but also is lightweight enough to adapt to the low computing power of the vehicle-side chip and respond within a given time; explores intelligent post-processing operations to minimize dependence on third-party libraries.

[0122] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A light thorn recognition method adapted to a vehicle-mounted CMS, characterized in that: include: Capturing images of the on-site environment to obtain on-site images; Marking the target area in the scene image, and performing image preprocessing on the marked scene image, wherein the image preprocessing includes scale scaling; Creating a lightweight neural unit SFB, and building a light thorn detection model based on the neural unit SFB; The scene image is input into the light thorn detection model to perform light thorn detection, and the detection result is output.

2. The optical thorn recognition method adapted to the vehicle-mounted CMS according to claim 1, characterized in that: Before performing light thorn recognition, model training is also included, and the model training includes: Capturing images of different scene environments to obtain view images; Annotating a target area in the view image, and performing a scale scaling process on the annotated view image; generating a sample set based on the processed view image; A training set and a validation set are created based on the view images in the sample set, and model training is performed on the optical thorn detection model according to the training set and the validation set.

3. The optical thorn recognition method adapted to the vehicle-mounted CMS according to claim 1, characterized in that: The marking of the target area includes: Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn, and obtain a first region; Identify the light thorn area in the representative image, take the corresponding light source as the corner point and mark the region of interest based on the extension direction of one side of the corresponding light thorn to obtain a second region; An overlapping area between the first area and the second area corresponds to an area where the light source is located.

4. The optical thorn recognition method adapted to the vehicle-mounted CMS as claimed in claim 1, characterized in that: The image preprocessing comprises: Acquire the annotated scene image, and scale the scene image based on a predefined standard; perform image extraction on the scaled scene image, acquire corresponding image Y component information from the scene image based on the YCbCr color model, and use the image Y component information as image data of the scaled scene image.

5. The optical thorn recognition method adapted to the vehicle-mounted CMS according to claim 1, characterized in that: The image preprocessing also includes image enhancement processing, and the image enhancement processing includes at least one of image blurring, mosaic enhancement, and random erasing, and the blur kernel size of the image blurring is 5×5.

6. The optical thorn recognition method adapted to the vehicle-mounted CMS as claimed in claim 1, characterized in that: Create a lightweight neural unit SFB, including: based on the depth group convolution Depth Group Conv, the batch normalization layer BatchNormLayer, the constant level activation function ReLU, and the channel shuffle unit shuffle channel to form a lightweight neural unit SFB.

7. The optical thorn recognition method adapted to the vehicle-mounted CMS as claimed in claim 1, characterized in that: Constructing a light thorn detection model based on the neural unit SFB includes: Building a backbone network based on the neural unit SFB, the backbone network is used to extract feature information from the scene image; A Neck layer is constructed using a PANet structure, and the Neck layer is used to perform multi-scale fusion on the input feature information to obtain output features; Setting a model decoupling head to detect the light thorns according to each of the input output features to obtain detection data; The detection data is input into a preset loss function to implement gradient backpropagation for parameter update to obtain a trained convolutional neural network.

8. The optical thorn recognition method adapted to the vehicle-mounted CMS as claimed in claim 7, characterized in that: The backbone network is constructed based on the neural unit SFB, including: A convolutional layer Stardard Conv, a batch normalization layer BatchNorm Layer, and a Relu activation layer form a convolution unit CBR; After the input layer, one layer of convolutional unit CBR is sequentially arranged, and after the convolutional unit CBR, six layers of neural unit SFB are arranged; An SPP layer is set at the end of the neural unit SFB to form a feature extraction network, wherein the SPP layer uses an Add operation to add multi-scale layers.

9. A storage medium having a computer program stored thereon, characterized in that: The computer program is used to implement a light thorn recognition method adapted for a vehicle-mounted CMS as described in any one of claims 1-8.

Images