Outdoor frosting detection method based on high and low frequency signal fusion

Abstract

The application relates to the field of visual detection, and discloses an outdoor frosting detection method based on high-low frequency signal fusion. The method first performs high-frequency component enhancement preprocessing on an input image to highlight surface details; then a double-branch detection network is constructed, wherein a main branch extracts global semantic features, and a high-frequency branch enhances high-frequency information through wavelet transformation and a learnable scaling; subsequently, the features of the two branches are learnably fused across layers, and a channel-spatial joint attention mechanism is introduced for calibration; finally, the fusion features are used to output a frosting category, a confidence and a region position. Through explicit enhancement and utilization of image high-frequency details, the application effectively improves the detection accuracy and robustness of low-contrast and fine-texture frosting regions in a complex outdoor environment, and can realize real-time and automatic frosting monitoring and early warning.

Description

Technical Field

[0001] This invention relates to the field of visual inspection, and more specifically to an outdoor frost detection method based on high- and low-frequency signal fusion. Background Technology

[0002] With global climate change and the increasing frequency of extreme low temperatures and high humidity, frost is prone to form on outdoor surfaces such as roads, bridges, power transmission lines, airport runways, vehicle windshields, and building facades. Frost not only affects visual transparency but also leads to reduced surface friction coefficients, decreased electrical insulation performance, and increased structural loads, seriously threatening traffic safety, power transmission safety, aviation operations, and the normal use of public facilities.

[0003] Currently, common methods for detecting frost can be mainly categorized as follows: Detection methods based on physical sensors, such as temperature sensors, humidity sensors, and frost point sensors, indirectly determine the risk of frost formation by measuring environmental parameters. These methods suffer from slow response times, high installation and maintenance costs, limited spatial coverage, and the inability to directly identify the morphology and distribution of frost.

[0004] Manual inspection methods rely on on-site observation or periodic checks by personnel, which have drawbacks such as strong subjectivity, low efficiency, inability to monitor in real time, and difficulty in operation in harsh environments.

[0005] Visual detection methods based on traditional image processing utilize image processing techniques such as edge detection, texture analysis, and threshold segmentation to identify frost layers. However, because frost-covered areas typically exhibit low contrast, blurred edges, and fine textures, traditional methods suffer from poor robustness and high false positive and false negative rates in complex scenarios such as changing lighting, rain and snow interference, and nighttime imaging.

[0006] Deep learning-based detection methods: In recent years, convolutional neural networks (CNNs) have made significant progress in image recognition, and some studies have attempted to apply them to frost detection. However, conventional CNN models tend to extract global semantic features of images and are insufficient in perceiving high-frequency details (such as tiny ice crystal boundaries, abrupt changes in surface reflection, etc.), resulting in limited detection accuracy in tasks with subtle textures and obvious local changes, such as frost.

[0007] Therefore, existing technologies still lack an automatic frost detection method that can effectively integrate high-frequency and low-frequency information from images and achieve high precision, real-time performance, and robustness in complex outdoor environments. Summary of the Invention

[0008] The purpose of this invention is to provide an outdoor frost detection method based on high and low frequency signal fusion, thereby solving the above-mentioned technical problems.

[0009] The objective of this invention can be achieved through the following technical solutions: The outdoor frost detection method based on high- and low-frequency signal fusion includes the following steps: S1. Acquire the outdoor scene image to be detected; S2. Perform high-frequency component enhancement preprocessing on the image to obtain an enhanced image containing high-frequency detail information; S3. The enhanced image is input into a dual-branch frost detection network for processing. The dual-branch network includes: The main feature extraction branch is used to extract global semantic features of the image; The high-frequency attention enhancement branch is used to extract and enhance high-frequency detail features of the image; S4. Perform cross-layer fusion of the features extracted by the main feature extraction branch and the features extracted by the high-frequency attention enhancement branch to obtain fused features; S5. Based on the fusion features, output the frosting detection results, which include frosting type, confidence level, and frosting area location information.

[0010] As a further technical solution, the high-frequency component enhancement preprocessing in step S2 includes: S21. Perform Gaussian filtering on the input image to obtain the low-frequency component; S22. Subtract the low-frequency component from the original image to obtain the high-frequency component; S23. The original image and the high-frequency component are stitched together along the channel dimension to form a multi-channel enhanced image.

[0011] As a further technical solution, the high-frequency attention enhancement branch in step S3 processes the enhanced image according to the following process: S31: Perform discrete wavelet transform on the input image and decompose it into four sub-bands: LL, LH, HL, and HH. S32: The HH subband is scaled and enhanced using a trainable parameter γ, which is initialized to 1.0; S33: Perform inverse wavelet transform on the enhanced sub-band to reconstruct the high-frequency enhanced image; S34: Extract the features of the high-frequency enhanced image as high-frequency features using a lightweight convolutional neural network.

[0012] As a further technical solution, the cross-layer fusion in step S4 specifically refers to: The high-frequency features are weighted and fused with the second-stage output features and the third-stage output features of the main feature extraction branch, respectively, where the fusion weights are learnable parameters.

[0013] As a further technical solution, at the node of the weighted fusion, a channel space joint attention mechanism is introduced to calibrate the fused features; The channel-space joint attention mechanism includes a channel attention module and a spatial attention module executed sequentially. The channel attention module is used to calculate the weights of each channel, and the spatial attention module is used to generate a spatial saliency mask.

[0014] As a further technical solution, the dual-branch frost detection network is trained end-to-end using the following loss function: ;in, , These are the weighting coefficients. For classifying losses, To pinpoint the loss.

[0015] As a further technical solution, the data augmentation methods used during training include at least one of the following: random brightness adjustment, Burmester noise simulation fogging, Gaussian noise addition, and random cropping.

[0016] As a further technical solution, in step S5, the location information of the frosted area is output in the form of a bounding box or a heat map, and an alarm is triggered when the confidence level exceeds a set threshold.

[0017] The beneficial effects of this invention are: (1) By designing high-frequency component enhancement preprocessing and high-frequency attention enhancement branches, this invention enables the model to explicitly capture subtle texture changes and edge information in the frosted area, effectively overcoming the problem of insufficient detail perception ability of traditional visual methods in low contrast and complex lighting environments, and significantly improving the detection accuracy and stability in interference scenarios such as rain, snow, and night. (2) A dual-branch parallel architecture is adopted. The main branch is responsible for understanding the global scene semantics, while the high-frequency branch focuses on extracting local detail features. The two complement and enhance each other through a learnable cross-layer fusion mechanism and attention module, realizing the effective fusion of high and low frequency information. This allows the network to obtain richer feature representations without blindly increasing the depth or number of parameters, thus ensuring model efficiency. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a logical schematic diagram of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Please see Figure 1 As shown, this invention is an outdoor frost detection method based on high and low frequency signal fusion. Step S1: Construct outdoor frost image data; Outdoor scene images are collected under different time periods, climatic conditions, and geographical locations. The presence and location of frost are labeled to form a labeled dataset. The images include visible light images, infrared thermal images, or multimodal fusion images. The label categories include at least: no frost and frost. Long-term continuous acquisition is carried out using fixed cameras or vehicle-mounted vision systems, and a large-scale dataset containing seasonal changes is established.

[0022] Step S2: Design a high-frequency component enhancement preprocessing module; Given an input image Its size is ,in: Indicates the height of the image (in pixels). Indicates the width of the image (in pixels). This indicates the number of channels in the image; for RGB images, the value is 3. First, Gaussian kernel convolution is performed on the image to obtain the low-frequency components. : ; in, Represents the Gaussian kernel function. x and y are the coordinates of the pixels, respectively. The standard deviation of the Gaussian kernel controls the degree of blurring; in this embodiment, it is set to 1.5. * represents the convolution operator. Then the high-frequency components are obtained. : The original image is stitched together with the high-frequency components to create a nine-channel enhanced image: ; This operation highlights subtle differences in surface texture, enhances the sensitivity of subsequent networks to frosting edges, and enables the model to focus on the subtle texture perturbations caused by frosting in the initial stage.

[0023] Step S3: Construct a two-branch frost detection network; This invention designs a deep neural network structure containing two parallel branches: Branch 1: Main Feature Extraction Branch; Classic convolutional neural networks, such as ResNet-50 and Darknet, are used as the backbone to extract multi-level semantic features of images.

[0024] The input image is Let the first The feature map output by the layer is in: For the height of the feature map, The width of the feature map, This represents the number of channels in the feature map; these feature maps are sequentially passed to subsequent detection heads to generate classification and localization results.

[0025] For example, using a ResNet-50 pre-trained on ImageNet as the backbone, an augmented image is input to extract multi-scale semantic features. The output includes feature maps from the second, third, and fourth stages, denoted as follows: , , .

[0026] Branch Two: High-Frequency Attention Enhancement Branch; This branch is specifically responsible for capturing high-frequency details in the image. The specific workflow is as follows: First, the input image Perform Discrete Wavelet Transform (DWT) to decompose it into four sub-bands: ; in, (Low-Low) subbands indicate that low-pass filtering has been applied in both the horizontal and vertical directions. The (Low-High) sub-band indicates low-pass filtering in the horizontal direction and high-pass filtering in the vertical direction. (High-Low) sub-bands represent high-pass filtering in the horizontal direction and low-pass filtering in the vertical direction. (High-High) subband indicates that high-pass filtering has been applied in both the horizontal and vertical directions.

[0027] Secondly, the HH subband mentioned above is amplified using a learnable scaling factor γ: γ is a trainable parameter, initialized to 1.0, which is automatically adjusted during training to enhance or suppress high-frequency contributions.

[0028] Then, the enhanced subband is merged with other subbands using inverse wavelet transform (IDWT) to reconstruct a new image that emphasizes high-frequency information. ; ; Will Input a small-scale CNN network (such as a shallow layer of ResNet-18) and extract the corresponding high-frequency feature maps. ; High frequency features The feature maps were upsampled using bilinear interpolation to the outputs of the second and third stages of the backbone network. and Perform cross-layer fusion: ; ; , The fused feature map , These are learnable weights used to control the intensity of high-frequency features, with an initial value set to 0.5; UpSample() represents the upsampling operation. , This represents the upsampling factor.

[0029] Add a channel-space joint attention module at each fusion node: Channel attention: Calculate the importance weight of each channel and highlight high-frequency relevant channels; Spatial attention: Generates a spatial mask that focuses on areas in the image where frost may exist (such as areas with dense edges).

[0030] This module first generates channel weights through global average pooling and a fully connected layer to recalibrate the feature channels; then, a convolutional layer generates a spatial attention mask to highlight areas suspected of frost. The calibrated features are then fed into the subsequent detection head.

[0031] Finally, model training: We used a self-built outdoor frost image dataset, which contains more than 10,000 labeled images, covering various scenes such as roads, power lines, bridges, and vehicles, as well as different time periods and weather conditions.

[0032] The model is trained using supervised training, and the loss function is: ; in: For classifying losses, , For tags, Output for the model; Indicates the total number of samples in the batch; To determine the localization loss, a commonly used deep learning bounding box loss function is employed, such as: , For tags, For network output. , These are weighting parameters, which should be adjusted according to the actual situation. This represents the total number of prediction boxes in a batch. Indicates the summation index; The following strategies were used during training: Optimizer: Adam, initial learning rate lr = 1 x 10 -4 Batch size: 16; Data augmentation: random brightness adjustment, Burmester noise to simulate fogging, Gaussian noise addition, random cropping; Training cycle: 100 epochs.

[0033] Step S5: Model Deployment and Real-Time Inference; Export the trained model in a lightweight format (such as ONNX or TensorRT) and deploy it on one of the following platforms: edge computing devices (such as NVIDIA Jetson series); cloud servers; camera built-in chips; the process is as follows: Receive images from the camera; Execute the forward reasoning process in S2 to S4; Output detection results, including: Frost type (no frost / with frost); confidence score; bounding box or heatmap of frost area; If the confidence level exceeds the set threshold, an alarm signal will be triggered and uploaded to the management center.

[0034] The detection head consists of a series of convolutional layers, responsible for decoding the fused and calibrated features. The final output consists of two parts: Categorized output: For each preset anchor, output confidence scores for two categories: frost-free and frost-free.

[0035] Positioning output: Each anchor point outputs the offset of a bounding box (center point coordinates, width, height).

[0036] After post-processing with non-maximum suppression (NMS), the final detection results are filtered out, including the frost category label, confidence score, and bounding box coordinates of the frost area. When the confidence score of the frost category exceeds a threshold (e.g., 0.7), the system automatically triggers an audible and visual alarm or sends a warning message to the monitoring center.

[0037] For example: The trained model was converted to TensorRT engine format and deployed in an edge computing box (with a built-in NVIDIA Jetson Xavier NX) next to the highway. This box is directly connected to a 2-megapixel starlight-level camera for 24-hour road monitoring. The system analyzes the video stream in real time at 5 frames per second. Once frost is detected on the road surface and the confidence level is greater than 0.65, an alarm message (including screenshots and location) is immediately uploaded to the road maintenance center via a 4G network, and a red warning box for the frost-covered area is automatically overlaid on the monitoring screen. This method effectively reduces the risk of traffic accidents caused by frost, achieving unmanned and intelligent road safety early warning.

[0038] The method employs an end-to-end fully differentiable design, eliminating the need for complex manual feature engineering and ensuring stable training. Furthermore, the data augmentation strategies used simulate various severe weather conditions, significantly enhancing the model's adaptability to outdoor scenes across different seasons, regions, and imaging conditions, demonstrating excellent cross-scene generalization.

[0039] Meanwhile, this method can be flexibly deployed on edge devices, the cloud, or camera chips to meet real-time processing needs. It can be widely applied in fields such as intelligent transportation, power operation and maintenance, aviation safety, and smart cities to achieve automated and intelligent monitoring and early warning of infrastructure frost conditions, which helps prevent safety accidents, reduce operation and maintenance costs, and improve the level of public safety management.

[0040] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. An outdoor frost detection method based on high- and low-frequency signal fusion, characterized in that, Includes the following steps: S1. Acquire the outdoor scene image to be detected; S2. Perform high-frequency component enhancement preprocessing on the image to obtain an enhanced image containing high-frequency detail information; S3. The enhanced image is input into a dual-branch frost detection network for processing. The dual-branch network includes: The main feature extraction branch is used to extract global semantic features of the image; The high-frequency attention enhancement branch is used to extract and enhance high-frequency detail features of the image; S4. Perform cross-layer fusion of the features extracted by the main feature extraction branch and the features extracted by the high-frequency attention enhancement branch to obtain fused features; S5. Based on the fusion features, output the frosting detection results, which include frosting type, confidence level, and frosting area location information.

2. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 1, characterized in that, The high-frequency component enhancement pretreatment in step S2 includes: S21. Perform Gaussian filtering on the input image to obtain the low-frequency component; S22. Subtract the low-frequency component from the original image to obtain the high-frequency component; S23. The original image and the high-frequency component are stitched together along the channel dimension to form a multi-channel enhanced image.

3. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 2, characterized in that, The high-frequency attention enhancement branch in step S3 processes the enhanced image according to the following procedure: S31: Perform discrete wavelet transform on the input image and decompose it into four sub-bands: LL, LH, HL, and HH. S32: The HH subband is scaled and enhanced using a trainable parameter γ, which is initialized to 1.0; S33: Perform inverse wavelet transform on the enhanced sub-band to reconstruct the high-frequency enhanced image; S34: Extract the features of the high-frequency enhanced image as high-frequency features using a lightweight convolutional neural network.

4. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 3, characterized in that, The cross-layer fusion mentioned in step S4 specifically refers to: The high-frequency features are weighted and fused with the second-stage output features and the third-stage output features of the main feature extraction branch, respectively, where the fusion weights are learnable parameters.

5. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 4, characterized in that, At the node of the weighted fusion, a channel space joint attention mechanism is introduced to calibrate the fused features; The channel-space joint attention mechanism includes a channel attention module and a spatial attention module executed sequentially. The channel attention module is used to calculate the weights of each channel, and the spatial attention module is used to generate a spatial saliency mask.

6. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 1, characterized in that, The dual-branch frost detection network is trained end-to-end using the following loss function: ;in, , These are the weighting coefficients. For classifying losses, To pinpoint the loss.

7. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 6, characterized in that, The data augmentation methods used during training include at least one of the following: random brightness adjustment, Burmester noise to simulate fogging, Gaussian noise addition, and random cropping.

8. The outdoor frost detection method based on high- and low-frequency signal fusion according to claim 1, characterized in that, In step S5, the location information of the frosted area is output in the form of a bounding box or a heat map, and an alarm is triggered when the confidence level exceeds a set threshold.

Images