An Infrared-Visible Dual-Mode Adaptive Fusion Method for Detecting Leaks in Liquid Media

Abstract

This invention discloses an infrared and visible light dual-modal adaptive fusion method for detecting liquid media leaks, belonging to the field of industrial vision inspection technology. The specific method includes: acquiring infrared thermal imaging images and visible light images; preprocessing the images using geometric registration and scale normalization; constructing a dual-branch backbone network to independently extract infrared features and dual-modal features from the visible light and infrared images respectively; inputting the infrared features into a multi-scale wavelet convolution module for frequency domain noise reduction and structural enhancement; performing adaptive feature fusion to obtain fused features; sequentially inputting the fused features into a multi-scale convolution kernel module and a high-efficiency upsampling channel offset convolution block for multi-scale context enhancement and spatial semantic alignment; inputting the optimized features into a decoupled detection head to complete the localization and classification of the liquid media leak area, and outputting the leak target box, confidence level, and category. This method has significant advantages in detecting liquid media leaks in low-contrast, weak-texture, and strong-interference environments.

Description

Technical Field

[0001] This invention relates to the fields of industrial visual inspection, deep learning target detection, and infrared and visible light multimodal fusion technology, specifically to an infrared and visible light dual-modal adaptive fusion liquid medium leakage detection method. Background Technology

[0002] In industrial production sites such as petrochemical, chemical, and oil and gas transportation, equipment such as pipelines, valves, flanges, and pump sets for transporting liquid media are subjected to high temperatures, high pressures, corrosion, and vibration for extended periods. This can easily lead to problems such as seal aging, weld defects, material fatigue, and erosion damage, resulting in leaks of toxic, flammable, and corrosive liquid media. Such leaks are characterized by their high degree of concealment, weak initial signs, and rapid spread. If not detected in time, they can easily cause fires, explosions, poisoning, environmental pollution, and major casualties, posing a serious threat to production safety, public safety, and the ecological environment.

[0003] Current methods for detecting liquid media leaks in industrial settings mainly include manual inspection, portable gas detectors, fixed-point sensor monitoring, ultrasonic detection, and fiber optic sensing. These methods generally have the following limitations: manual inspection relies on personnel experience, has limited coverage, slow response times, and cannot achieve 24 / 7 continuous monitoring; portable detection equipment can only sample at fixed points, making continuous monitoring of the entire area difficult; fixed sensors can only monitor localized locations and are insufficient for detecting leaks in large areas, over long distances, at high altitudes, or in concealed locations; ultrasonic and fiber optic sensing are costly to deploy, complex to construct, and difficult to maintain, hindering large-scale deployment.

[0004] With the development of machine vision and infrared thermal imaging technologies, vision-based leak detection methods have gradually become a research hotspot. Infrared thermal imaging can sense the temperature difference between the medium and the environment, and can work in environments such as low light, nighttime, dust, and haze, compensating for the shortcomings of visible light imaging; visible light images have the advantages of clear texture, rich detail, and accurate positioning. However, in actual industrial scenarios, single-modal imaging has obvious defects: infrared images lack texture, have blurred edges, and are susceptible to interference from thermal noise, background radiation, and sunlight reflection; visible light images are easily affected by lighting, shadows, occlusion, and fog, and may completely fail under extreme conditions. Therefore, single-modal imaging cannot meet the requirements of high-precision, high-robustness, and all-weather leak detection in complex industrial environments.

[0005] In recent years, multimodal fusion detection technology has made some progress, but existing methods still have significant shortcomings:

[0006] 1) The fusion strategy is simple, mostly using image-level front-end stitching or detection result-level back-end weighting, without achieving feature-level deep fusion. Fusion and detection are independent of each other, and end-to-end joint optimization is not possible. 2) Cross-modal features have problems such as spatial misalignment, scale inconsistency, and semantic shift. Simple superposition can easily introduce modal interference, causing weak targets to be submerged. 3) There is a lack of dedicated enhancement modules for the thermal and texture features of liquid media leakage, resulting in insufficient ability to identify micro-leakage, low-contrast leakage, and diffuse leakage. 4) The network structure is complex, with a large number of parameters and slow inference speed, making it difficult to deploy in real time on explosion-proof pan-tilt units, edge computing units, and embedded devices. 5) It cannot adaptively switch between single and dual modal modes. When a camera fails or imaging fails, the detection performance drops sharply.

[0007] In summary, existing technologies struggle to simultaneously meet the demands of high-precision, high-robustness, all-weather operation, low false alarm rate, easy deployment, and adaptive mode switching for intelligent detection of liquid media leaks in industrial settings. Therefore, there is an urgent need to propose a dual-modal intelligent leak detection method that can deeply integrate infrared and visible light modes, adaptively suppress noise, accurately locate the leak area, adapt to industrial explosion-proof environments, and be deployable in real time. Summary of the Invention

[0008] This invention addresses the technical bottlenecks in existing liquid medium leakage detection technologies for the chemical and petrochemical industries, including insufficient single-modal imaging information, easy missed detection of weak feature leaks, strong interference from complex backgrounds, insufficient cross-modal feature fusion, severe spatial semantic shift, and difficulty in industrial deployment of models. It discloses an infrared-visible dual-modal adaptive fusion method for liquid medium leakage detection, which specifically includes the following steps: Infrared thermal imaging images and visible light images were acquired, and the images were preprocessed using geometric registration and scale normalization. A dual-branch backbone network is constructed to extract infrared features and dual-modal features from visible light images and infrared images independently. The dual-branch backbone network includes a multi-scale wavelet convolution module, an adaptive weight feature fusion module, a multi-scale convolution kernel module, an efficient upsampling channel offset convolution block, and a decoupled detection head. The process involves inputting infrared features into a multi-scale wavelet convolution module for frequency domain noise reduction and structural enhancement; inputting dual-modal features into an adaptive weighted feature fusion module to dynamically allocate the contribution ratios of infrared and visible light modes through learnable channel weights and perform adaptive feature fusion to obtain fused features; inputting the fused features sequentially into a multi-scale convolution kernel module and an efficient upsampling channel offset convolution block for multi-scale context enhancement and spatial semantic alignment; inputting the optimized features into a decoupled detection head to complete the localization and classification of the liquid medium leakage area and output the leakage target box, confidence level, and category.

[0009] Furthermore, the multi-scale wavelet convolution module employs cascaded Haar wavelet transform to decompose infrared features into low-frequency and high-frequency subbands, suppressing high-frequency noise while preserving target structural information.

[0010] Furthermore, the adaptive weighted feature fusion module includes an infrared feature enhancement branch and a visible light feature enhancement branch. The infrared feature enhancement branch is used for edge and saliency enhancement through the FLEEModule, while the visible light feature enhancement branch is used to generate a spatial attention map and perform weighted processing on the bimodal features, and can learn channel weight parameters. and ,satisfy Used to dynamically adjust modal contribution.

[0011] Furthermore, the efficient upsampling channel offset convolutional block aligns the feature space through channel grouping, spatial cyclic displacement, and channel shuffling, thereby mitigating the semantic offset phenomenon caused by upsampling.

[0012] Furthermore, the multi-scale convolution kernel module employs parallel multi-scale convolution and cross-stage partial connection structures to enhance global context awareness.

[0013] Furthermore, during image preprocessing: Gaussian filtering and contrast-limited adaptive histogram equalization are applied to infrared images to enhance the thermal differences between the leaked area and the background; dehazing and sharpening are performed on visible light images to enhance edge and texture details, and blurry, overexposed, and invalid frames are removed to ensure the quality of input data.

[0014] Furthermore, the adaptive weighted feature fusion module uses a shared spatial attention mechanism to concatenate the dual-modal features along the channel dimension, and generates a spatial attention map through convolution and Sigmoid activation to suppress the background region and enhance the target region with weights.

[0015] This invention discloses an infrared and visible light dual-modal adaptive fusion liquid medium leakage detection method. By filtering out thermal and electronic noise in infrared images through frequency domain decomposition, it significantly improves the identification capability of weak and minute leaks, solving the problem of high false detection rates in industrial scenarios. It can automatically switch modal weights according to complex environments such as lighting, obstruction, dust, and nighttime, maintaining high detection accuracy even when a single mode fails, achieving stable operation around the clock. The EUCSC module effectively solves the feature misalignment problem, and in conjunction with a multi-scale receptive field structure, it has accurate positioning capabilities for irregular diffusion leaks, narrow strip leaks, and small-area dripping leaks. The overall architecture is lightweight and modular, enabling real-time inference on Ascend AI accelerator cards and embedded edge devices, and is compatible with explosion-proof gimbals and industrial C / S architectures. On self-built industrial datasets and public datasets, mAP50... Its 95% performance index is superior to mainstream YOLO series and multimodal fusion detection models, and it has significant advantages in detecting liquid media leaks in low-contrast, weak-texture, and strong-interference environments. Attached Figure Description

[0016] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of an infrared and visible light dual-modal adaptive fusion liquid medium leakage detection method according to the present invention.

[0018] Figure 2 This is a framework diagram of the dual-branch backbone network in this invention.

[0019] Figure 3 This is a framework diagram of the adaptive weight fusion module in this invention.

[0020] Figure 4 This is a framework diagram of the multi-scale wavelet convolution module in this invention.

[0021] Figure 5 This is a framework diagram of the efficient upsampling channel offset convolution block in this invention.

[0022] Figure 6 This is a schematic diagram illustrating the specific implementation process of the method in the embodiments of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0024] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0025] like Figure 1 The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method shown includes the following steps: Infrared and visible light images from the industrial site are acquired simultaneously and geometrically registered and scale-aligned. Features are extracted from both images using a dual-branch backbone network. Infrared features undergo frequency domain denoising and structural enhancement via a multi-scale wavelet convolution module C3K2_WTConv. The dual-modal features are input to an adaptive weighted feature fusion module IV-AWFM, which dynamically allocates the contribution ratio of the infrared and visible light modes through learnable channel weights, achieving adaptive feature fusion. The fused features are then sequentially input to a multi-scale convolution kernel module CSP_MSK and a high-efficiency upsampling channel offset convolution block EUCSC to achieve multi-scale context enhancement and spatial semantic alignment. The optimized features are then input to a decoupled detection head to locate and classify the liquid medium leakage area, outputting the leakage target bounding box, confidence level, and category.

[0026] Furthermore, this invention employs an end-to-end deep learning detection architecture that integrates dual-modal image synchronous acquisition, dual-branch feature extraction, adaptive weight fusion, multi-scale context enhancement, spatial alignment optimization, and decoupled head detection output to construct the infrared-visible dual-modal adaptive fusion leak detection network IRV-FLDNet.

[0027] Furthermore, this invention is based on a dual-modal image acquisition and preprocessing method: Simultaneously acquired infrared and visible light images suffer from problems such as viewing angle deviation, scale inconsistency, and asynchronous imaging timing. This invention employs the following preprocessing steps: 1) The checkerboard calibration method is used to calibrate the intrinsic and extrinsic parameters of the dual-spectrum camera, achieving pixel-level geometric registration; 2) Scale the images uniformly to 640×640 resolution to maintain consistent target scale; 3) Gaussian filtering and contrast-limited adaptive histogram equalization are applied to the infrared image to enhance the thermal difference between the leak area and the background; 4) Dehaze and sharpen visible light images to enhance edge and texture details; 5) Remove blurry frames, overexposed frames, and invalid frames to ensure the quality of input data.

[0028] Furthermore, such as Figure 2 As shown, this invention proposes an infrared-visible light dual-modal adaptive fusion liquid medium leakage detection network (IRV-FLDNet) based on a dual-branch backbone feature extraction method. This network aims to improve the robustness, cross-modal complementarity, and multi-scale feature representation capabilities of multimodal target detection in complex industrial scenarios. The network employs a dual-branch feature fusion detection framework. In the backbone stage, features are extracted from both the infrared and visible light modes. The infrared branch focuses on perceiving thermal radiation distribution, temperature difference regions, and abnormal high and low temperatures. It also combines the multi-scale wavelet convolution module C3K2_WTConv to perform frequency domain feature decoupling and multi-scale information modeling, obtaining multi-level feature representations including low-level texture, edge details, and high-level semantic information. Each mode has completed preliminary feature encoding before entering the fusion stage, thus enabling more thorough mining of complementary information between modes during subsequent fusion, thereby enhancing the overall feature representation capability.

[0029] This invention employs an intermediate fusion strategy in the backbone network. Compared to the problem of modal information aliasing caused by direct splicing at the input stage in front-end fusion methods, and the problem of back-end fusion methods that only perform fusion at the high-level semantic stage and are difficult to take into account low-level detailed modeling, intermediate fusion can realize information interaction between modalities during feature extraction, while taking into account the coordinated fusion of low-level detailed information and high-level semantic information, thereby improving the network's ability to represent complex leakage targets.

[0030] To address the issues of large scale variations, mismatched receptive fields, and spatial misalignment during upsampling of leaking liquid media, this invention introduces the CSP_MSK module at different feature levels to enhance the network's multi-scale context modeling capabilities. Simultaneously, it combines the EUCSC module to achieve feature space alignment during upsampling, effectively mitigating semantic offset issues in feature fusion and further improving the detection accuracy and stability of leaking targets in complex industrial environments.

[0031] like Figure 4 As shown, the dual-branch feature extraction architecture is based on an improved feature extraction network. The infrared branch focuses on extracting thermal features such as thermal radiation distribution, temperature difference regions, abnormally high temperatures, and low temperatures. Frequency domain decoupling is achieved through a multi-scale wavelet convolution module C3K2_WTConv. 1) The features are decomposed into low-frequency structural subbands and high-frequency noise subbands using cascaded Haar wavelet transform;

[0032]

[0033] Cascaded wavelet decomposition ( Level: The number of levels in wavelet decomposition) Only the low-frequency subband LL is recursively decomposed to obtain deeper low-frequency and high-frequency representations:

[0034] in

[0035] 2) Threshold suppression is applied to high-frequency subbands to filter out industrial environment thermal noise, electronic noise, and texture jitter;

[0036] 3) By reconstructing the feature map through inverse wavelet transform, infrared features with lower noise, clearer structure, and more prominent target contours are obtained; 4) Residual connections preserve the original feature information, avoiding the loss of small-area leakage features caused by downsampling and filtering.

[0037]

[0038] In C3K2_WTConv, WTConv separates different frequency components in the wavelet domain, thus blocking the problem of high-frequency noise being amplified layer by layer in traditional convolution and effectively suppressing common noise interference in infrared images.

[0039] like Figure 3 As shown, the adaptive weighted feature fusion module designed in this method is the core of achieving efficient dual-modal fusion. Its specific implementation includes... 1) Infrared feature enhancement branch: FLEEModule edge enhancement and local saliency enhancement are used to highlight the outline of the leakage area and the thermal anomaly area; Specifically: First, the Sobel operator is used to extract gradient information in the horizontal and vertical directions of the infrared image:

[0040] in and represents the Sobel convolution kernels in the horizontal and vertical directions, respectively, and G represents the edge features.

[0041] Then, channel compression is performed on the concatenated features using a 1×1 convolution:

[0042] To further enhance directional texture structure, vertical 3×1 and horizontal 1×3 asymmetric row and column convolution sequences are used to enhance vertical and horizontal texture details respectively, effectively enhancing the linear structures commonly found in infrared images;

[0043] Simultaneously, locally saliency branches are constructed, and high-response regions are highlighted and the original resolution is restored through max pooling:

[0044] Finally, the features from the two branches are fused and further integrated with edge and saliency information through convolution:

[0045] Subsequently, the RepDW module was also introduced for semantic enhancement:

[0046]

[0047] 2) Visible light feature enhancement branch: Employs a RepDW depthwise separable convolutional structure to enhance texture details and reduce computational cost; its computation process can be represented as:

[0048] in Indicates the characteristics of visible light. These are the output features of RepDW.

[0049] The coding features of the visible light branch can be represented as:

[0050] 3) Shared spatial attention mechanism: The dual-modal features are concatenated along the channel dimension, and a spatial attention map is generated by convolution and sigmoid activation. This map suppresses the background region and enhances the target region with weights. First, the two modality encoded features are concatenated by channels and then normalized to the [0,1] interval through convolution and the Sigmoid function to form attention weights.

[0051]

[0052] in Let represent the Sigmoid function, and A represent the shared spatial attention map. Then, element-wise multiplication is used to weight the features of the two modalities:

[0053]

[0054] in This indicates element-wise multiplication.

[0055] 4) Learnable channel weighting mechanism: Set two sets of channel-by-channel learnable parameters and The initial values ​​are all 0.5. During training, the model is constrained to the interval [-1,1] by the clamp function, so that the model can automatically enhance the dominant mode and suppress the inferior mode according to the scene quality.

[0056] in The parameter is initialized to 0.5 as a channel-by-channel learnable scalar parameter. Explicitly apply absolute value constraints during training: the calculation formula is as follows:

[0057]

[0058] in This means truncating x to the interval [a, b]. This constraint guarantees... At the same time, the original signs of the parameters are preserved, thereby achieving non-complementary competitive dynamic weighting.

[0059] Bundle Projected to The interval is defined, but the gradient is not blocked, allowing the optimizer to continue pushing the gradient in the correct direction. The calculation formula is as follows:

[0060] This dynamic adjustment mechanism enables the model to autonomously determine the relative importance of each feature layer, and can amplify a certain modality (e.g., in extreme scenarios) ) or suppress a certain mode (e.g. This enables reverse suppression, allowing for more flexible modal importance assessment. This avoids the model's over-reliance on a single modal feature, thus ensuring the balance and robustness of multi-feature fusion.

[0061] 5) Weighted fusion output: The feature fusion formula below is used to achieve modal advantage complementarity.

[0062]

[0063] in , and These are learnable channel weight parameters.

[0064] The adaptive weighted feature fusion module achieves efficient fusion of infrared and visible light features through four key mechanisms. It can fully exploit the complementary information between infrared and visible light modes, improve the multimodal feature representation capability, and enhance the model's detection robustness and environmental adaptability in complex scenes.

[0065] Furthermore, to address issues such as large variations in the scale of leaked targets, mismatch in feature receptive fields, and spatial misalignment caused by upsampling, this invention employs a two-stage optimization structure.

[0066] 1) Multi-scale convolution kernel module CSP_MSK: The CSP_MSK module, based on the Cross-Stage Partial Connection (CSP) architecture, incorporates convolutional kernels of three scales in parallel: 1×1, 3×3, and 5×5. Specifically, the 1×1 kernel reassembles channel information to adapt to local features in small leak areas; the 3×3 kernel extracts medium-scale structural information to match the morphological features of typical leak areas; and the 5×5 kernel expands the receptive field to capture global contextual information in large leak areas. Through this parallel setup of multi-scale convolutional kernels, the module can simultaneously capture local details, medium-scale structural features, and global contextual information, effectively mitigating the receptive field mismatch problem caused by differences in leak area scale. Furthermore, the CSP_MSK module introduces a channel shuffle mechanism during the feature fusion stage, rearranging feature maps from convolutional kernels of different scales along the channel dimension to promote information interaction and fusion between multi-scale features. This mechanism helps alleviate spatial feature misalignment problems that may be caused by upsampling operations, enhancing the network's adaptability to the geometric structures of leak areas of different sizes and shapes.

[0067] 2) High-efficiency upsampling channel-offset convolutional block EUCSC: like Figure 5 As shown, the features are divided into four groups along the channels, and upward, downward, leftward, and rightward offset operations are performed respectively. Then, through channel shuffling and convolution reconstruction, spatial position correction, semantic alignment, and detail enhancement are achieved, fundamentally alleviating the problems of blurring, misalignment, and inaccurate positioning caused by traditional upsampling.

[0068] Specifically, given an input feature map It is divided into four sub-feature maps evenly along the channel dimension:

[0069] Apply positive and negative circular shifts in both the horizontal (width) and vertical (height) directions. and The offsets are made in the positive and negative directions in the height direction, respectively. and Then, positive and negative offsets are performed in the width direction. This spatial offset process can be represented as:

[0070] Where δ is the cyclic displacement step size (usually 1 or 2 pixels).

[0071] By applying spatial displacement in different directions, complementary information distributions can be formed among the sub-feature maps within their spatial neighborhoods, enabling features to acquire richer contextual relationships within local regions. Unlike traditional convolution, which relies on convolutional kernels for spatial modeling, this operation, through the rearrangement and recombination of feature maps in the spatial dimension, allows for closer spatial relationships between different channels, thereby enhancing feature expressive power.

[0072] The final offset feature blocks are recombined through channel concatenation to achieve cross-channel information exchange and local feature enhancement. The calculation formula is as follows:

[0073] The EUCSC module effectively improves the interaction efficiency between multi-scale features, enabling more accurate alignment between high-level semantic information and low-level spatial details, thereby further enhancing the target recognition capability of the detection network in complex industrial scenarios.

[0074] Furthermore, this invention sets up a detection output and industrial early warning logic process, in which decoupled detection heads are used to complete bounding box regression, confidence prediction, and category judgment respectively, reducing interference between tasks; dynamic confidence threshold and leak area threshold are set to filter noise interference and false leaks; when a leak target is detected in multiple consecutive frames, an alarm is triggered, and the coordinates of the leak area, the leak confidence, the size of the leak area, the camera number and the timestamp are output, and image storage, log recording and audible and visual alarms are completed simultaneously.

[0075] In practical implementation, this invention addresses the demands of industrial sites for real-time data processing, system reliability, and business continuity. It employs a client / server (C / S) architecture, utilizing a server for data management and computation to ensure data security. The client provides a graphical monitoring interface, deployed in control rooms and workstations, enabling human-machine interaction and real-time control, thus meeting some of the rigid requirements of industrial scenarios. For example... Figure 6As shown: The architecture consists of a client and a server. The client connects to the hard disk recorder via a local area network (LAN), receives relevant image data, performs frame segmentation, encodes the data, and sends it to the server. The client has a complete graphical interface for easy user operation. The encoded image is sent to the server via the LAN. The server runs a multimodal infrared image target detection algorithm. After receiving the data from the client, the algorithm sends it to the multimodal detection algorithm for detection and returns the detected result to the client.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dual-modal adaptive fusion method for detecting leakage in liquid media using infrared and visible light, characterized in that, include: Infrared thermal imaging images and visible light images were acquired, and the images were preprocessed using geometric registration and scale normalization. A dual-branch backbone network is constructed to extract infrared features and dual-modal features from visible light images and infrared images independently. The dual-branch backbone network includes a multi-scale wavelet convolution module, an adaptive weight feature fusion module, a multi-scale convolution kernel module, an efficient upsampling channel offset convolution block, and a decoupled detection head. Infrared features are input into a multi-scale wavelet convolution module for frequency domain noise reduction and structural enhancement; dual-modal features are input into an adaptive weighted feature fusion module, which dynamically allocates the contribution ratio of infrared and visible light modes through learnable channel weights and performs adaptive feature fusion to obtain fused features. The fused features are sequentially input into a multi-scale convolutional kernel module and an efficient upsampling channel offset convolutional block for multi-scale context enhancement and spatial semantic alignment. The optimized features are then input into the decoupled detection head to complete the localization and classification of the liquid medium leakage area, and output the leakage target box, confidence level, and category.

2. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 1, characterized in that: The multi-scale wavelet convolution module employs cascaded Haar wavelet transform to decompose infrared features into low-frequency and high-frequency sub-bands, suppressing high-frequency noise while preserving target structural information.

3. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 1, characterized in that: The adaptive weighted feature fusion module includes an infrared feature enhancement branch and a visible light feature enhancement branch. The infrared feature enhancement branch is used for edge and saliency enhancement through the FLEE module, while the visible light feature enhancement branch is used to generate a spatial attention map and perform weighted processing on the bimodal features, and can learn channel weight parameters. and ,satisfy Used to dynamically adjust modal contribution.

4. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 1, characterized in that: The efficient upsampling channel offset convolutional block aligns the feature space through channel grouping, spatial cyclic displacement, and channel shuffling, thus mitigating the semantic offset phenomenon caused by upsampling.

5. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 1, characterized in that: The multi-scale convolution kernel module employs parallel multi-scale convolution and cross-stage partial connection structures to enhance global context awareness.

6. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 1, characterized in that: When preprocessing the images: Gaussian filtering and contrast-limited adaptive histogram equalization are applied to the infrared images to enhance the thermal difference between the leaked area and the background. The visible light images are dehazed and sharpened to enhance edge and texture details, and blurry, overexposed and invalid frames are removed to ensure the quality of the input data.

7. The infrared-visible dual-mode adaptive fusion liquid medium leakage detection method according to claim 3, characterized in that: The adaptive weighted feature fusion module uses a shared spatial attention mechanism to concatenate the dual-modal features along the channel dimension, and generates a spatial attention map through convolution and Sigmoid activation. This map suppresses the background region and enhances the target region with weights.

Images