A portable multi-modal anti-stealing detection device and an intelligent identification method thereof

Abstract

The application belongs to the technical field of data representation and signal monitoring, and relates to a portable multi-modal anti-stealing detection device and an intelligent identification method thereof, which comprises a heterogeneous sensing module, an intelligent analysis module and an interactive optimization module; the heterogeneous sensing module comprises an infrared sensing unit, a light supplementing unit and an optical sensing unit; the intelligent analysis module comprises a feature extraction unit, a feature matching unit and a lightweight network model, and real-time comparison and matching of heat source features and visible light reflection image features are performed; the matching comprises texture analysis; the texture analysis comprises wavelet transform and Gabor filtering, calculation of texture direction, frequency and contrast; the interactive optimization module comprises a display unit and an alarm unit; the method realizes portable multi-modal anti-stealing detection, and improves the technical defects of low multi-modal detection fusion detection efficiency, poor environmental adaptability, insufficient power consumption control and poor human-computer interaction experience of existing anti-stealing devices and methods.

Description

Technical Field

[0001] This invention belongs to the field of data characterization and signal monitoring technology, and particularly relates to a portable multimodal anti-spy camera detection device and its intelligent identification method. Background Technology

[0002] With the development of technology and the arrival of the digital age, the risk of personal privacy information leakage has increased significantly.

[0003] With increasing awareness of personal privacy protection, the demand for anti-spy camera products has surged. According to market research reports, the global anti-spy camera market is growing at a rate exceeding 10% annually. From a market size perspective, the global anti-spy camera device market has seen continuous growth over the past few years and is projected to maintain a steady upward trend over the next five years. The global anti-spy camera device market was worth approximately $500 million in 2019 and is expected to exceed $800 million by 2026.

[0004] Currently, mainstream anti-spy camera technologies are mainly divided into three categories: infrared filter detection, infrared thermal imaging, and visible light detection. Infrared filter products, such as a certain brand's detector, use the principle of active illumination with 850nm infrared LEDs, identifying cameras by observing the reflection on the filter. While these devices are inexpensive (around 200 yuan per unit), convenient, and compact; they detect the infrared emission source of cameras, triggering an alarm upon detection; and filter detection involves actively emitting infrared light, utilizing the principle of infrared light reflection—the infrared light reflects off the lens, and the location of the camera is determined by observing the reflected light spot through the filter. However, they have three major drawbacks: 1) Users need to continuously observe the target close to the filter, which can cause visual fatigue or dizziness after more than 3 minutes; 2) The recognition rate of reflected light spots decreases by more than 40% when the ambient light intensity exceeds 300 lux; 3) They cannot detect cameras with infrared illumination turned off; experimental data shows that their overall detection rate against common spy cameras on the market is only 62%. Their actual performance is poor, with a high false detection rate. Furthermore, their anti-spy camera function is limited, lacking visual detection, thermal imaging detection, and image processing.

[0005] Infrared thermal imaging, such as a certain thermal imager, locates hidden cameras by detecting temperature differences of 0.5-2℃. Market research shows that high-end models can achieve a detection rate of 90% under ideal conditions, but there are two fatal weaknesses: (1) when there are multiple heat sources in the environment (such as routers, chargers, etc.), the false alarm rate is as high as 35%; (2) it is almost ineffective for miniature cameras with low power consumption design (working temperature and ambient temperature difference < 0.3℃). A laboratory test showed that the detection rate of this technology for pinhole cameras with a diameter of less than 5mm is less than 20%. Product features: convenient and compact, simple to operate, download the APP on your mobile phone, and plug the device into your mobile phone to view. The product technology is thermal imaging detection: using the working principle of thermal radiation, when the camera is working, heat will be generated, and the detector can identify the temperature difference within ±2 degrees in the area. By finding abnormal heat sources, the location of hidden cameras can be determined. Product disadvantages: low resolution products have short recognition distance and poor temperature sensitivity; high resolution products have good recognition effect, but are expensive, ranging from 2000 to 6000 yuan. Interface: The device needs to be plugged into a mobile phone for viewing, and the interface between the phone and the device is prone to loosening, and the device is at risk of being dropped. Functionality: When the camera is working in complex heat source scenarios, such as when there is a TV, power strip, and router in the same small area, it is difficult to identify them and is prone to missing detections. Scene matching: In addition, the function is limited, only able to display heat sources. Visible light detection devices, such as a certain brand's detector, use an 850nm active light source with CMOS imaging, which can achieve real-time image display, but have the following limitations: the effective detection distance is no more than 30cm and the device surface temperature exceeds 50℃ after 15 minutes of continuous operation; the power consumption is as high as 3.5W, which makes it difficult for the product's battery life to exceed 2 hours.

[0006] The dual-light detection device has the following characteristics: Real-time images can be viewed by plugging the device into a laptop and using a video player; it utilizes the visible light reflection principle, with LED beads actively emitting 850nm light, which is reflected back onto a mirror and then imaged after filtering by a camera; the device's disadvantages include a small visible reflection range: due to the influence of reflection distance, light intensity, and angle, the visible range is relatively small. When the device lens and the object being measured are horizontally aligned, the field of view extends approximately 20cm outwards from the object. Power consumption: because it requires active infrared light emission, the LED lights have relatively high power consumption, resulting in rapid temperature rise; after 15 minutes of operation, the device temperature reaches approximately 50 degrees Celsius. Scene matching and anti-peeping functions are limited, relying solely on infrared and visible light viewing.

[0007] Existing hybrid products, such as a certain brand's dual-light detector, which combines thermal imaging and visible light technology, have revealed new problems: the devices are generally larger than 200×100×50mm and weigh over 500g; the user interface requires professional training to operate, making the learning curve too steep for ordinary users; and the price range of 3000-8000 yuan far exceeds the affordability of individual consumers. Market feedback shows that user satisfaction with these products is only 58%, with the main issues concentrated on portability and ease of use.

[0008] Infrared thermal imaging + visible light: Equipped with both visible light and infrared thermal imaging capabilities, providing clear real-time images and heat source displays; Infrared thermal imaging: Offers consistent thermal imaging functionality, utilizing the principle of thermal radiation to locate heat sources. Product features: Relatively large and thick, making it inconvenient to carry. User interface: The interface is overly technical and difficult to use; it lacks a dual-light fusion design.

[0009] In summary, based on the research of the current products, the following conclusions were drawn: 1) The current global anti-spy camera equipment market is still dominated by single-function products (such as infrared filters or thermal imagers). (Infrared thermal imaging + visible light + AI image processing) covers as many types of hidden cameras as possible (including pinhole cameras without infrared emission) and is used in typical scenarios such as hotels and changing rooms. The equipment has a false detection rate as high as 30% to 50%. There is an urgent need to accelerate the development of effective equipment and achieve large-scale penetration in the high-end business travel and government and enterprise security fields.

[0010] 2) According to a Market Research Future report, in 2023, the compound annual growth rate of multi-functional integrated devices in the global privacy protection hardware market reached 18.7%, far exceeding that of single-function devices (9.2%); there is an urgent need to fill the market gap for cost-effective low-end devices (below 200 yuan) and professional-grade devices (above 2000 yuan).

[0011] 3) Existing equipment cannot be simultaneously adapted to smart home, public safety, and enterprise services; for smart home and public safety, customized versions are provided for government agencies, supporting linkage with law enforcement recorders for counter-espionage investigation; for enterprise services, a dedicated model for conference rooms is developed, integrating wireless signal detection functions to prevent the leakage of trade secrets.

[0012] The technical bottlenecks are mainly reflected in the following four aspects: the fusion algorithm for multimodal detection is inefficient, and the image processing latency of existing devices generally exceeds 200ms; the environmental adaptability is poor, and the false alarm rate in complex lighting environments (such as hotel rooms) is 3 times higher than that in laboratory conditions; the power consumption control is insufficient, and the battery life in continuous working mode is difficult to exceed 4 hours; the human-computer interaction experience is poor, and 83% of users reported that they needed to repeatedly consult the manual to complete basic operations.

[0013] In conclusion, these shortcomings of existing technologies severely restrict the widespread application of anti-spy camera technology, and breakthroughs are urgently needed through technological innovation. Summary of the Invention

[0014] The purpose of this invention is to improve the technical defects of existing anti-spy camera devices and methods, such as low multimodal detection fusion detection efficiency, poor environmental adaptability, insufficient power consumption control, and poor human-computer interaction experience. This invention proposes a portable multimodal anti-spy camera detection device and its intelligent recognition method.

[0015] To achieve the above objectives, the present invention adopts the following technical solution.

[0016] As a first aspect of the present invention, a portable multimodal anti-spy camera detection device is proposed, comprising a heterogeneous sensing module, an intelligent analysis module, and an interaction optimization module connected in sequence; wherein: The heterogeneous sensing module includes an infrared sensing unit, a supplementary lighting unit, and an optical sensing unit; the infrared sensing unit is used to acquire infrared thermal imaging images, the supplementary lighting unit is used to emit active infrared light, and the optical sensing unit is used to acquire visible light reflection images. The intelligent analysis module includes a feature extraction unit, a feature matching unit, and a lightweight network model connected in sequence. The feature extraction unit is used to receive infrared thermal imaging images from the infrared sensing unit and extract heat source features, and to receive visible light reflection images from the optical sensing unit and extract visible light reflection image features. The feature matching unit is used to compare and match the heat source features with the visible light reflection image features in real time. The lightweight network model includes an input preprocessing layer, a shallow feature extraction layer, a bottleneck inverted residual feature layer, an attention multi-scale residual layer, a lateral inhibition constraint layer, a global pooling factorization layer, and an output discriminant layer. The interactive optimization module includes a display unit and an alarm unit; the display unit is signal-connected to the intelligent analysis module and is used to output detection results; the alarm unit is signal-connected to the output end of the lightweight network model.

[0017] In specific implementation, the infrared sensing unit is an external thermal imaging sensor with a resolution better than 160×120 and a temperature difference sensitivity of ±0.5℃; the supplementary lighting unit is an 850nm active infrared supplementary light with an adjustable power range of 5 to 20W; and the optical sensing unit is a wide-angle optical lens with an FOV greater than or equal to 120°. The feature matching unit performs matching between the heat source features and the visible light reflection image features, including texture analysis; the texture analysis includes wavelet transform and Gabor filtering, and calculates texture direction, frequency and contrast.

[0018] The texture includes gray-level co-occurrence matrix, roughness, contrast, orientation, line density, regularity, and coarsness.

[0019] The heat source characteristics include shape, outline, texture, emissivity, radiation intensity, peak radiation wavelength, pattern and trajectory, and concentric circle heat distribution characteristics. Emissivity refers to the ratio of the radiation intensity emitted by an object at a certain temperature to the radiation intensity emitted by a blackbody at the same temperature. Specifically, the emissivity of metal surfaces ranges from 0.1 to 0.2, while the emissivity of human skin exceeds 0.9. Different objects have different emissivities; objects with higher emissivity emit stronger infrared radiation. The radiation intensity is proportional to the fourth power of temperature. When the temperature of the target object is higher than the background ambient temperature, its radiation intensity is stronger, appearing as a brighter area than the background in the infrared image; conversely, it appears as a darker area. The peak radiation wavelength varies depending on the temperature of the target object. Specifically, the peak radiation wavelength of the human body is approximately 9.35 μm, and the temperature is approximately 310 K. The aforementioned pattern and trajectory refer to the target or object's straight line, curve, acceleration, deceleration, direction, and speed, which appear as a series of continuous points or lines in an infrared image, i.e., the trajectory. In specific implementation, for example, by analyzing the motion trajectory of the target object, the device type, target type, and behavior can be determined. In complex motion modes, the target object's behavior information can be used to assist in target identification. In specific implementation, the shape features are as follows: although the contrast and signal-to-noise ratio of infrared images are low, the shape of the target object can still be identified to some extent; by analyzing the edge and corner features of the target object, target shape information such as circles, rectangles, and triangles can be extracted; the contour features refer to the boundaries of the object's brightness changes; the contour information of the object can be extracted through edge detection, and then the shape and size of the target can be analyzed; for example, in security monitoring, by analyzing the object's contour, it can be determined whether it is a person or a vehicle; The texture features refer to the surface texture of the target object; in specific implementation, the texture is related to the material and surface treatment; generally, metal textures are smooth and wood textures are rough; the texture includes gray-level co-occurrence matrix, roughness, contrast, directionality, line density, regularity, and coarsness.

[0020] The visible light reflection image features include texture, color, shape, spatial relationships, statistics, and specular reflection spot features; the color features include color moments and histograms; the color moments are the image color distribution represented by moments, and the moments include the first moment (mean), the second moment (variance), and the third moment (skewness); the histogram is the global distribution of image colors, that is, the proportion of different colors in the entire image, which is suitable for describing images that are difficult to segment automatically and images that do not need to consider the spatial position of objects; The shape features include boundary, geometric, and spatial relationship features; the boundary features are used to obtain the shape parameters of the image by describing the boundary features; in specific implementation, the boundary features are Hough transform to detect parallel lines and boundary direction histograms; further, the boundary features also include Fourier shape descriptors: by utilizing the closure and periodicity of the region boundary, the two-dimensional problem is transformed into a one-dimensional problem, and the shape expression such as curvature function, centroid distance, and complex coordinate function is derived from the boundary points; The geometric features refer to the features used to describe and match the shape of an image, including but not limited to quantitative measurements of shape, such as moments, area, and perimeter features. The spatial relationship features include inclusion or containment, overlap or overlap, and connection or adjacency relationship features; The inclusion or containment feature reflects whether a target object is contained within another target or object, or whether a target or object contains other targets or objects. The overlapping or overlapping features reflect whether there is an overlap or overlap between targets in space; the connection / adjacency describes the interconnection or adjacency relationship between multiple segmented targets or objects, such as the top, bottom, left, and right relationships. The statistical features include mean, median, gray-level variance, and image gray-level value range. The mean refers to the average pixel value, reflecting the average reflectance intensity. The median refers to the middle value among the pixel gray levels, representing a contrast condition. The gray-level variance reflects the overall dispersion of each pixel's gray-level value from the image's average gray-level value, and is an important measure of the information content of an image. The image gray-level value range is the difference between the maximum and minimum gray-level values ​​of the image, reflecting the degree of variation in image gray-level values ​​and indirectly reflecting the information content of the image. The interactive optimization module includes a display unit and an alarm unit. The display unit supports a dual-light fusion display mode; the alarm unit supports vibration / sound dual-mode alarm.

[0021] The lightweight network model includes networks with the following features: 1) replacing traditional convolutions with depthwise separable convolutions; or introducing grouped convolutions, pooling, pointwise convolutions, and skip connections; 2) introducing bottleneck layers and inverted residual blocks, attention-based multi-scale residuals, or searching and automating networks through weak connections; 3) using global average pooling and fully connected layers or expanding coefficients and factorization; 4) based on feedback and reuse or based on feature extraction networks, cascaded memory group convolutional networks, and reconstruction networks; 5) employing lateral suppression networks.

[0022] In specific implementation, the following measures are taken: 1) Use depthwise separable convolution to replace traditional convolution, introduce bottleneck layers and inverted residual blocks, and also use global average pooling and fully connected layers or expansion coefficients and factorization; 2) Use weak connection search and automated networks; 3) Introduce grouped convolution, pooling, pointwise convolution and skip connections; 4) Based on feature extraction networks, cascaded memory group convolutional networks and reconstruction networks; 5) Based on feedback and reuse; 6) Based on attention-based multi-scale residuals; 7) Lateral suppression networks; 8) Multi-scale large receptive field feature distillation.

[0023] The shallow feature extraction layer uses depthwise separable convolution instead of traditional convolution, and the depthwise separable convolution includes channel-wise convolution and point-wise convolution. The bottleneck inverted residual feature layer includes a bottleneck layer and an inverted residual block. The bottleneck layer adopts a three-segment structure of 1×1 convolution dimensionality reduction, 3×3 convolution feature extraction, and 1×1 convolution dimensionality increase. The inverted residual block adopts a structure of first increasing dimensionality, then depthwise convolution, and then pointwise dimensionality reduction, and is connected by short-circuited residuals. The attention-based multi-scale residual layer sets up a multi-scale receptive field, extracts large-scale contours and thermal distribution features from infrared images and small specular reflection spots and fine-grained texture features from visible light images in parallel, and performs adaptive weight allocation on the dual-modal feature channels through a channel attention mechanism.

[0024] In specific implementation, the display unit supports dual-light fusion display mode; the alarm unit supports vibration / sound dual-mode alarm; the threshold of the vibration / sound dual-mode alarm can be customized; the setting principle of the threshold customization setting is: the threshold of the vibration / sound dual-mode alarm can be customized to trigger the alarm when the heat source area is ≥3mm² and the reflected light intensity is ≥200lux. A second aspect of the present invention provides an intelligent identification method for an anti-spy camera detection device, comprising the following steps: S1. Extract infrared thermal imaging images and visible light reflection images through infrared sensing unit and optical sensing unit respectively; S2. Feature Extraction: Extracting heat source features from infrared thermal imaging images and extracting visible light reflection image features from visible light reflection images; S3. Feature Matching: Real-time comparison and matching of extracted heat source features and visible light reflection image features; S4. Lightweight network detection and recognition are achieved using a lightweight network model. S5. Output the detection results through the display unit and provide an alarm prompt through the alarm unit.

[0025] In step S1, the infrared thermal imaging sensor has a resolution better than 160×120 and a temperature difference sensitivity of ±0.5℃; the thermal imaging sensor requires infrared supplementary light, and the power range of the infrared supplementary light is 5W to 20W; the optical sensing is achieved through a wide-angle optical lens with an FOV ≥ 120°.

[0026] The feature matching in step S3 also includes texture analysis, which uses wavelet transform and Gabor filtering to obtain texture direction, frequency and contrast.

[0027] In specific implementation, heterogeneous sensing involves extracting infrared and visible light images; the infrared resolution is better than 160×120, and the temperature difference sensitivity is ±0.5℃; thermal imaging sensing requires infrared supplementary light, and the power range of the infrared supplementary light is 5 to 20W; the optical sensing is achieved through a wide-angle optical lens with an FOV greater than or equal to 120°. The feature extraction process extracts heat source features from infrared images and visible light reflection image features; the feature matching process compares the heat source features and specular reflection spot features obtained from the feature extraction process in real time. The lightweight network recognition includes the following feature networks: 1) replacing traditional convolution with depthwise separable convolution, introducing bottleneck layers and inverted residual blocks, and also using global average pooling and fully connected layers or expanding coefficients and factorization; 2) through weak connection search and automated networks; 3) introducing grouped convolution, pooling, pointwise convolution and skip connections; 4) based on feature extraction networks, cascaded memory group convolutional networks and reconstruction networks; 5) based on feedback and reuse; 6) based on attention-based multi-scale residuals; 7) lateral suppression networks; 8) multi-scale large receptive field feature distillation.

[0028] The heat source characteristics include shape, outline, texture, emissivity, radiation intensity, radiation peak wavelength, pattern and trajectory, and concentric circle heat distribution characteristics.

[0029] The texture includes grayscale co-occurrence matrix, roughness, contrast, orientation, line density, regularity, and coarsness.

[0030] The visible light reflection image features include texture, color, shape, spatial relationships, statistics, and specular reflection spot features; The interaction optimization includes display and alarm; the display is an OLED touchscreen that supports dual-light fusion display mode; the alarm supports vibration / sound dual-mode alarm; the threshold of the vibration / sound dual-mode alarm can be customized.

[0031] Beneficial effects The portable multimodal anti-spy camera detection device and its intelligent identification method proposed in this invention have the following advantages compared with the prior art: 1. The detection device solves the problem of multimodal detection fusion, specifically by integrating infrared light filter detection, thermal imaging detection and visible light reflection, thus solving the problem of high false detection rate of single technologies such as using only infrared filters; 2. The detection device achieves human-computer interaction optimization, specifically through dual-light fusion visualization interface thermal imaging and visible light superposition display and intelligent alarm, so that users do not need to observe the filter with the naked eye for a long time; 3. The detection device can adapt to dynamic environments, specifically by adaptively adjusting detection parameters such as infrared light intensity and thermal imaging sensitivity; 4. The detection method can still distinguish cameras from other heat sources in complex heat source scenarios, such as hotel rooms with electrical appliances, by matching heat source features and image analysis, reducing the false detection rate to below 10%. 5. The detection method described above solves the problems of high false negative rates, such as thermal imaging in complex heat source environments and limited detection range, such as visible light reflection. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the system composition and connection of a portable multimodal anti-spy camera detection device according to the present invention; Figure 2 This is a flowchart of an intelligent identification method for an anti-spy camera detection device according to the present invention. Detailed Implementation

[0033] The following detailed description, in conjunction with the accompanying drawings and specific embodiments, elaborates on the specific implementation of a portable multimodal anti-spy camera detection device and its intelligent identification method according to the present invention.

[0034] Example 1 As a first aspect of the present invention, a portable multimodal anti-spy camera detection device is provided, the composition and connection relationship of which are as follows: Figure 1 As shown, it includes a heterogeneous perception module, an intelligent analysis module, and an interaction optimization module; The heterogeneous sensing module includes an infrared sensing unit, a supplementary lighting unit, and an optical sensing unit; In practice, the infrared sensing unit is an external thermal imaging sensor with a resolution better than 160×120 and a temperature difference sensitivity of ±0.5℃. The detection device solves the problem of multimodal detection fusion through a heterogeneous sensing module. Specifically, it integrates infrared light filter detection, thermal imaging detection, and visible light reflection to solve the problems of high false detection rate of single technology, such as using only infrared filters, and high false detection rate, such as thermal imaging in complex heat source environments and limited detection range, such as visible light reflection. In a specific implementation, the supplementary lighting unit is an 850nm active infrared supplementary light with an adjustable power range of 5 to 20W; the optical sensing unit is a wide-angle optical lens with an FOV greater than or equal to 120°; in a specific implementation, an optical wide-angle lens with an FOV of 120° is used. The intelligent analysis module includes a feature extraction unit, a feature matching unit, and a lightweight network model; The intelligent analysis module compares and matches heat source features with visible light reflection image features in real time; the matching includes, but is not limited to, texture analysis; the texture analysis, in specific implementation, uses wavelet and Gabor filtering to obtain texture direction, frequency and contrast. The feature extraction unit is connected to the feature matching unit, and the feature matching unit is connected to the lightweight network model.

[0035] The feature extraction unit is used to extract the heat source features of the infrared image and the visible light reflection image features; the feature matching unit performs real-time comparison of the heat source features and specular reflection spot features obtained by the feature extraction unit.

[0036] The intelligent analysis module of the detection device can also adapt to dynamic environments. Specifically, it adjusts detection parameters such as infrared light intensity and thermal imaging sensitivity through adaptive algorithms. In complex heat source scenarios, such as hotel rooms with electrical appliances, it can still distinguish cameras from other heat sources through heat source feature matching and image analysis, reducing the false detection rate to below 2%. The heat source features include shape, contour, texture, emissivity, radiation intensity, radiation peak wavelength, pattern and trajectory, and concentric circle heat distribution characteristics. The emissivity refers to the ratio of the radiation intensity emitted by an object at a certain temperature to the radiation intensity emitted by a blackbody at the same temperature. In practice, the emissivity of a metal surface ranges from 0.1 to 0.2, while the emissivity of human skin exceeds 0.9. Different objects have different emissivities, and objects with higher emissivity will emit stronger infrared radiation. The radiation intensity is proportional to the fourth power of the temperature; when the temperature of the target object is higher than the background ambient temperature, its radiation intensity is stronger and it will appear as a brighter area than the background in the infrared image; conversely, it will appear as a darker area; the peak wavelength of the radiation varies depending on the temperature of the target object; in specific implementation, the peak wavelength of human radiation is about 9.35 μm and the temperature is about 310 K. The pattern and trajectory refer to the target or object's straight line, curve, acceleration, deceleration, direction and speed, which are a series of continuous points or lines in the infrared image, i.e., the trajectory. In practical implementation, for example, by analyzing the motion trajectory of the target object, the type of equipment, the type of target, and its behavior can be determined; under complex motion modes, the behavior information of the target object can be used to assist in the identification of the target. In specific implementation, the shape features are as follows: although the contrast and signal-to-noise ratio of infrared images are low, the shape of the target object can still be identified to a certain extent; by analyzing the edge and corner features of the target object, and by analyzing the shape of the target object, target shape information such as circles, rectangles, and triangles can be extracted. The outline feature refers to the boundary of the object's brightness change; the outline information of the object can be extracted through edge detection, and then the shape and size of the target can be analyzed; in specific implementation, such as in security monitoring, by analyzing the object's outline to extract the outline feature, the target can be determined to be a person or a vehicle. The texture feature refers to the surface texture of the target object; in practice, the texture is related to the material and surface treatment; typically, metal textures are smooth, while wood textures are rough. The texture includes grayscale co-occurrence matrix, roughness, contrast, orientation, line density, regularity, and coarsness.

[0037] The visible light reflection image features include texture, color, shape, spatial relationships, statistics, and specular reflection spot features; The color features include color moments and histograms; the color moments are the image color distribution represented by moments, and the moments include the first moment (mean), the second moment (variance), and the third moment (skewness); the histogram is the global distribution of image colors, that is, the proportion of different colors in the whole image, which is suitable for describing images that are difficult to segment automatically and images that do not need to consider the spatial position of objects. The shape features include boundary, geometric, and spatial relationship features; the boundary features are used to obtain the shape parameters of the image by describing the boundary features; in specific implementation, the boundary features are Hough transform to detect parallel lines and boundary direction histograms; the boundary features also include Fourier shape descriptors: by utilizing the closure and periodicity of the region boundary, the two-dimensional problem is transformed into a one-dimensional problem, and the curvature function, centroid distance, and complex coordinate function are derived from the boundary points; The geometric features refer to the features used to describe and match the shape of an image, including quantitative measures of shape, such as moments, area, and perimeter features. The spatial relationship features include features represented by inclusion or containment relationships, overlapping or overlapping relationships, and connection or adjacency relationships; The inclusion or containment feature reflects whether a target object is contained within another target or object, or whether a target or object contains other targets or objects. The overlapping or overlapping features reflect whether there is an overlap or overlap between targets in space; the connection / adjacency describes the interconnection or adjacency relationship between multiple segmented targets or objects, such as the top, bottom, left, and right relationships. The statistical features include mean, median, gray-level variance, and image gray-level value range; The mean value refers to the average pixel value, reflecting the average reflectance intensity; the median value refers to the middle value among the pixel gray levels, representing a contrast condition; the gray level variance reflects the overall dispersion of each pixel's gray level value from the image's average gray level value, and is an important measure of the amount of information in an image; the image gray level value range is the difference between the maximum and minimum gray level values ​​of the image, reflecting the degree of variation in the image's gray level values, and indirectly reflecting the amount of information in the image. The interaction optimization module includes a display unit and an alarm unit.

[0038] The display unit supports dual-light fusion display mode; the alarm unit supports vibration / sound dual-mode alarm.

[0039] The interactive optimization module of the detection device realizes human-computer interaction optimization. Specifically, through dual-light fusion visualization interface thermal imaging and visible light superposition display and intelligent alarm, users do not need to observe the filter with the naked eye for a long time. The lightweight network model includes a network with the following characteristics: 1) A network that uses depthwise separable convolutions instead of traditional convolutions; 2) Introduce a network with grouped convolution, pooling, pointwise convolution, and skip connections; 3) Introduce bottleneck layers and inverted residual blocks, attention-based multi-scale residuals, or networks searched through weak connections; 4) Use global average pooling and fully connected layers or expand the network with coefficient and factorization; 5) Convolutional and reconstruction networks based on feedback, reuse, or feature extraction, as well as cascaded memory; 6) Lateral inhibition network.

[0040] In specific implementation, the following measures are taken: 1) Use depthwise separable convolution to replace traditional convolution, introduce bottleneck layers and inverted residual blocks, or use global average pooling and fully connected layers or expand coefficients and factorization; 2) Use weak connection search and automated networks; 3) Introduce grouped convolution, pooling, pointwise convolution and skip connections; 4) Based on feature extraction networks, cascaded memory group convolutional networks and reconstruction networks; 5) Based on feedback and reuse; 6) Based on attention-based multi-scale residuals; 7) Lateral suppression networks; 8) Multi-scale large receptive field feature distillation.

[0041] In a specific implementation, the display unit is an OLED touchscreen that supports dual-light fusion display mode; the alarm unit, in a specific implementation, supports vibration / sound dual-mode alarm; the threshold of the vibration / sound dual-mode alarm can be customized. The setting principle for the threshold customization is as follows: the threshold of the vibration / sound dual-mode alarm can be customized to trigger the alarm when the heat source area is ≥3mm² and the reflected light intensity is ≥200lux. Example 2 In a second aspect, the present invention provides an intelligent identification method for an anti-spy camera detection device, the process of which is as follows: Figure 2 As shown, it includes the following steps: S201. Heterogeneous sensing is achieved through infrared and optical modules, specifically including two parts: thermal imaging sensing and optical sensing. In specific implementation, the infrared thermal imaging sensor has a resolution better than 160×120 and a temperature difference sensitivity of ±0.5℃; the thermal imaging sensor requires infrared supplementary lighting, and the power range of the infrared supplementary lighting is 5 to 20W; the optical sensing is achieved through a wide-angle optical lens with an FOV greater than or equal to 120°. S202, Intelligent Analysis; The intelligent analysis includes feature extraction, feature matching, and lightweight network recognition, which compares and matches heat source features with visible light reflection image features in real time. The matching includes, but is not limited to, texture analysis; In the specific implementation of the texture analysis, wavelet transform and Gabor filtering are used to obtain texture direction, frequency and contrast. The feature extraction specifically extracts the heat source features of the infrared image and the visible light reflection image features; the feature matching compares the heat source features and specular reflection spot features obtained from the feature extraction in real time. The lightweight network recognition method, in its specific implementation, includes the following measures: 1) replacing traditional convolution with depthwise separable convolution, introducing bottleneck layers and inverted residual blocks, or using global average pooling and fully connected layers, or expanding coefficients and factorization; 2) using weak connection search and automated networks; 3) introducing grouped convolution, pooling, pointwise convolution and skip connections; 4) based on feature extraction networks, cascaded memory group convolutional networks and reconstruction networks; 5) based on feedback and reuse; 6) based on attention-based multi-scale residuals; 7) lateral suppression networks; 8) multi-scale large receptive field feature distillation.

[0042] The heat source characteristics include shape, contour, texture, emissivity, radiation intensity, radiation peak wavelength, pattern and trajectory, and concentric circle heat distribution characteristics.

[0043] The texture includes grayscale co-occurrence matrix, roughness, contrast, orientation, line density, regularity, and coarsness.

[0044] The visible light reflection image features include, but are not limited to, texture, color, shape, spatial relationships, statistics, and specular reflection spot features; S203, Interaction optimization; The interaction optimization includes display and alarm; the display is an OLED touchscreen that supports dual-light fusion display mode; the alarm supports vibration / sound dual-mode alarm; the threshold of the vibration / sound dual-mode alarm can be customized.

[0045] The device achieves high-precision, low-false-alarm-rate detection of concealed cameras through the synergy of infrared thermal imaging, visible light reflection detection, and AI image recognition. Specifically, it employs modular design, such as a shared processor and optimized sensor layout, and uses intermittent infrared illumination and a low-power thermal imaging chip. The power consumption is less than 10 watts during continuous operation for one hour, and the surface temperature of the device remains stable below 50°C, solving the problem of severe overheating in similar products. The hardware cost is reduced by more than 50% compared to high-end thermal imaging devices.

[0046] The intelligent recognition method addresses the problem of multimodal detection fusion. By integrating infrared filter detection, thermal imaging detection, and visible light reflection, it solves issues such as high false negative rates when using only infrared filters, high detection range in complex heat source environments, and limited detection range in visible light reflection. Operation time is reduced by more than 50%, and the dizziness problem is eliminated. Experimental data shows that the overall detection accuracy of the fusion algorithm is improved to over 95%, an average improvement of more than 40% compared to single technologies.

[0047] Enhanced privacy protection, verified by a third-party testing organization, the device covers over 80% of common hidden camera devices (including pinhole cameras, disguised Wi-Fi devices, etc.), with a detection response time of ≤1 minute in public places, significantly improving user security. The device's output detection logs comply with the ISO / IEC 29100 privacy framework, providing technical support for legal evidence collection and contributing to the establishment of industry standards. In a detection experiment conducted in a 200㎡ enclosed space, the detection device achieved a detection rate of over 95% for hidden cameras: 52.3% with traditional infrared filters, 78.5% with single thermal imaging, and a false alarm rate of <1 time / hour.

[0048] Example 3 This embodiment provides a portable anti-spy camera detection device based on multimodal fusion, combining infrared thermal imaging, visible light reflectance detection, and AI image recognition to achieve high-precision, low-power concealed camera detection. The specific implementation steps and technical details are as follows: In the implementation of the heterogeneous sensing module, the hardware architecture uses an ARM Cortex-M7 processor as the main control chip and integrates a low-power Bluetooth 5.2 module (such as Nordic nRF52840) to achieve wireless connection with the mobile APP, avoiding loose physical interfaces; The sensor group includes: 1) Infrared thermal imaging module: FLIR Lepton 3.5 microbolometer is selected, with a resolution of 160×120, temperature measurement accuracy of ±0.5℃, working wavelength of 8-14μm, and power consumption of only 150mW; 2) Visible light detection module: Equipped with an 850nm infrared LED array: 10 LEDs, each with a power of 0.5W and an OV2740 CMOS sensor, supporting HDR mode to enhance the recognition of reflected light spots in low-light environments; Environmental sensors: integrates a BH1750 light sensor and a BME280 temperature and humidity sensor for dynamic adjustment of detection parameters; The power supply uses the TI BQ25601 charging IC, paired with a 5000mAh lithium polymer battery, supports PD fast charging, and provides up to 4 hours of continuous operation.

[0049] The intelligent analysis module employs multimodal detection, including: thermal imaging initial screening, visible light precise positioning, and dual-light fusion display; The thermal imaging coarse screening, in specific implementation, uses the Lepton module to scan the environment and generate a thermal distribution map, and automatically marks abnormal heat sources with a temperature difference ≥2℃ based on the YOLOv5s lightweight model, excluding interference sources such as routers and televisions; The visible light precise positioning, in specific implementation, involves emitting 850nm infrared light to the suspicious area, with the CMOS sensor capturing the reflected light spot. Combined with OpenCV's SIFT feature matching, the unique circular halo feature of the lens is identified, reducing the false detection rate to below 10%. The dual-light fusion display is specifically implemented by overlaying thermal imaging and visible light images on a mobile APP. The dual-light synchronous visualization is achieved by adjusting the transparency, and users can intuitively compare and confirm the camera position. The detection device measures 80×50×15mm and weighs 120g. It uses a magnetic clip to attach to the back of a mobile phone for collaborative operation. Within a 20㎡ area, it completes a full-room scan in 10 minutes with a detection rate of 98.7%, compared to 96.2% for the professional anti-spy camera device FLIR ONE Pro. Power consumption is controlled: after one hour of continuous operation, the surface temperature is ≤40℃, significantly better than similar products such as a certain brand's visible light detector which reaches 50℃. The device's extended functions include cloud database construction, allowing users to upload suspicious device characteristics to the cloud for continuous AI model optimization through federated learning. It also features electromagnetic wave-assisted detection and can reserve interfaces for radio frequency detection modules such as NFC / RFID, enabling future expansion to identify wireless transmission spy cameras.

[0050] The interaction optimization module also includes an intelligent alarm and detection report generation unit: when a suspicious device is detected, the APP provides vibration and sound alerts and generates a risk level report, such as "High risk: Pinhole camera detected, confidence level higher than 90%"; it automatically switches detection strategies based on ambient light intensity: thermal imaging is prioritized in strong light environments, and infrared supplementary lighting is activated in low light environments. This multimodal portable anti-spy camera detection device has broad application prospects in privacy protection.

[0051] Example 4: Specific Implementation of the Lightweight Network Model Based on the hardware architecture, perception parameters, and feature definitions of the aforementioned embodiments 1 to 3, this embodiment provides a detailed description of the specific architecture of the lightweight network model, the model training process, and the entire process of multimodal feature extraction and feature matching.

[0052] The lightweight network model in this embodiment uses depthwise separable convolutions instead of traditional convolutions, combining grouped convolutions, pooling, pointwise convolutions, and skip connections; it introduces bottleneck layers and inverted residual blocks, attention-based multi-scale residual structures, and employs weak connection search and automated network architecture design; it uses global average pooling in conjunction with fully connected layers, and achieves model compression through expansion coefficients and factorization; it introduces feature feedback and reuse mechanisms, building feature extraction networks, cascaded memory group convolutional networks, and reconstruction networks to complete feature iterative optimization; and it embeds a lateral suppression network to suppress background redundant feature interference.

[0053] The lightweight network model consists of, in sequence, an input preprocessing layer, a shallow feature extraction layer, a first bottleneck inverted residual feature layer, a second bottleneck inverted residual feature layer, an attention multi-scale residual layer, a lateral inhibition constraint layer, a global pooling factorization layer, a cascaded memory group convolutional network, a reconstruction network, and an output discriminant layer.

[0054] The input preprocessing layer (Layer 0) takes a dual-channel image as input: Channel 0 is an infrared thermal imaging image (160×120 resolution, pixel values ​​normalized to [0,1]), and Channel 1 is a visible light reflectance image (acquired synchronously, 160×120 resolution after geometric registration). The preprocessing operations specifically include: Grayscale correction: Histogram equalization is used, clip limit=2.0, tile grid size=8×8; Illumination equalization: Gamma correction is applied to the visible light channel, γ=0.8; Noise filtering: A 5×5 Gaussian kernel with σ=1.0 is used to filter the two channels separately; Texture pre-extraction: A 24-dimensional texture feature vector is generated by wavelet transform (db4 wavelet, 2-layer decomposition) + Gabor filtering (4 scales, 6 directions, 24 filter kernels in total), which is used as auxiliary input to Layer 1.

[0055] The shallow feature extraction layer (Layer 1) uses depthwise separable convolutions instead of traditional convolutions, specifically broken down as follows: Channel-wise convolution: 3×3 kernel, 2 input channels, 2 output channels, 1 channel per group is operated independently, no cross-channel fusion, number of parameters = 2×3×3×1=18; Pointwise convolution: 1×1 kernel, 2 input channels, 16 output channels, to achieve weighted fusion of features between channels, with a parameter count of 2×1×1×16=32; Downsampling: Max pooling is used, with a pooling kernel of 2×2, a step size of 2, and an output size of 80×60. Skip-connect: The original input of Layer 0 is reduced to 16 channels through a 1×1 convolution, and then added to the output of Layer 1 to form a residual connection.

[0056] The bottleneck inverted residual feature layer (Layer 2-3) has the following structure: 1×1 convolution for dimensionality reduction → 3×3 depthwise convolution for feature extraction → 1×1 convolution for dimensionality increase. Taking Layer 2 as an example: Dimensionality increase: 1×1 convolution, 16 input channels, 64 output channels (expansion factor 4); Depthwise convolution: 3×3 convolution, 64 input channels, 64 groups (channel-by-channel), 64 output channels; Dimensionality reduction: 1×1 convolution, 64 input channels, 24 output channels; Short-circuit connection: When the number of input and output channels is different, a 1×1 convolution is used to project the input, and then added to the output. The inverted residual block differs from the standard residual block: the standard residual block is "dimensionality reduction → convolution → dimensionality increase", while the inverted residual block is "dimensionality increase → depthwise convolution → dimensionality reduction". The intermediate layers have more channels, which is beneficial for feature extraction, and the output is more compact after dimensionality reduction.

[0057] The attention-based multi-scale residual layer (Layer 4) sets up three parallel receptive field branches: Branch A: 3×3 convolution to extract large-scale contour features of the infrared heat source; Branch B: 5×5 convolution to extract features of the heat distribution transition region; Branch C: 7×3 asymmetric convolution to extract visible light micro-mirror reflection spots and fine-grained texture features; each branch outputs 16 channels, resulting in 48 channels after splicing.

[0058] In the SE (Squeeze-and-Excitation) channel attention mechanism, Squeeze uses global average pooling to compress 48×40×30 to 48×1×1; Excitation uses two fully connected layers, with the intermediate dimension reduced to 48 / r (r=4, i.e., 12 channels), ReLU activation, and output sigmoid gating; Scale multiplies the gating vector with the original features channel by channel, adaptively enhancing the effective features of the hidden camera target and suppressing the features of interfering heat sources such as home appliances and routers.

[0059] The lateral suppression constraint layer (Layer 5) performs lateral suppression on the feature responses of adjacent pixels in the same frame image. Specifically, it uses a 3×3 convolution kernel with weights initialized to a center of 1.0, a neighborhood of -0.125 (8-neighborhood), and a bias of 0, forming a Laplacian suppression operator. The suppression strength is controlled by a Sigmoid gate, with the gate input being the global statistics (mean + standard deviation) of the current layer's features, and the output suppression coefficient α∈(0,1). The final output is: F_out = α·F_in + (1-α)·(F_in -F_laplacian), where F_laplacian is the Laplacian filtering result. By suppressing the uniform response of large-area continuous background heat sources (such as televisions and radiators), it highlights small-sized local abnormal heat sources and specular reflections from pinhole cameras.

[0060] In the global pooling factorization layer (Layer 6), global average pooling compresses the 40×30×48 feature map into 1×1×48, eliminating spatial dimensions, replacing the traditional fully connected layer, and reducing the number of parameters; extended coefficients and factorization: low-rank decomposition is performed on the 48-dimensional feature vector. If the target rank k=24, SVD decomposition is used to reconstruct the feature, compressing the feature dimensions while retaining the main information.

[0061] In the Cascaded Memory Cluster Convolutional Network and Reconstruction Network (Layer 7-8), the Cascaded Memory Cluster Convolutional Network maintains a 3-frame historical feature cache (FIFO queue). The current frame features and historical frame features are fused by cluster convolution, with 4 groups and 12 channels per group, to achieve temporal information aggregation. Feature feedback reuse mechanism: The high-confidence features output by Layer 7 (judged by Softmax probability > 0.7) are reduced in dimensionality by 1×1 convolution and then fed back to the skip connection of Layer 1 to enhance the stability of continuous detection in dynamic scenes. Reconstruction Network: For feature loss caused by low light and low resolution, 1×1 convolution is used for feature reconstruction, and residual connections avoid information loss.

[0062] The dataset was constructed by collecting images from 50 hotel rooms, 20 public changing rooms, 15 offices, and 30 restrooms, covering typical private scenarios. The samples consist of three categories: positive samples include various pinhole cameras, miniature hidden cameras, and ordinary cameras. Pinhole cameras include button-type, socket-type, smoke alarm disguised, and router-disguised types. Miniature hidden cameras range in diameter from 3 to 8 millimeters. The positive samples comprise 2800 paired infrared and visible light images. Negative samples include common interference sources such as routers, chargers, televisions, lamps, human heat sources, and air conditioner vents, totaling 4200 paired images. Background samples consist of 1500 images from targetless scenes. The total sample size is 8500 images, divided in an 8:1:1 ratio into a training set of 6800 images, a validation set of 850 images, and a test set of 850 images. The annotation standard adopts a two-level annotation: pixel-level annotation includes polygon vertex annotation of the target outline of the hidden camera lens, manual verification of the heat source area after temperature threshold segmentation, and the judgment standard of grayscale threshold greater than 180 and circularity greater than 0.8 for the specular reflection spot area; the category annotation is divided into three categories: category 0 is suspected hidden camera device, category 1 is ordinary heat source interference, and category 2 is ordinary background.

[0063] The hyperparameter configuration includes: a uniform input image size of 160×120, dual-channel. The optimizer used is the Adam optimizer, with parameters β1 of 0.9 and β2 of 0.999. The initial learning rate is set to 1 x 10^-3. The learning rate decays in a stepwise manner, multiplying by 0.1 at rounds 30, 60, and 80. The weight decay coefficient is 1 x 10^-4. The loss function consists of three weighted components: the total loss equals the classification loss plus 0.3 times the texture matching loss plus 0.1 times the center loss. The classification loss is the multi-class cross-entropy loss. The texture matching loss is the L2 loss for texture feature matching, i.e., the L2 norm of the difference between the infrared and visible light channel texture vectors. The center loss is the Center Loss, used to constrain intra-class compactness. The batch size is 16 during pre-training on a PC and 4 during fine-tuning on a Jetson Nano. The number of iterations is 100 pre-training rounds plus 20 fine-tuning rounds. The early stopping mechanism is set to terminate training and save the optimal weight model if the validation set mAP shows no improvement for 10 consecutive rounds. Data augmentation methods include: random rotation of ±15 degrees, random scaling from 0.9 to 1.1 times, random cropping, Gaussian noise with a standard deviation ranging from 0 to 0.05, and a Mosaic augmentation application probability of 0.5.

[0064] (1) For infrared heat source feature extraction processing; Shape features are obtained by using Canny edge detection combined with Hough transform to fit the circumscribed rectangle and extract 5-dimensional features, including the x-coordinate of the center point, the y-coordinate of the center point, the width, the height, and the aspect ratio, which are used to characterize the geometric shape of the target.

[0065] Contour features are derived using Freeman chain codes combined with Fourier descriptors, taking the first 16 coefficients (16 dimensions in total) to characterize the frequency domain features of the boundary shape. Texture features are derived using a gray-level co-occurrence matrix, calculating four indices (energy, entropy, contrast, and correlation) in four directions with a distance d of 1, totaling 16 dimensions, to characterize surface roughness and regularity.

[0066] Emissivity characteristics are based on material database matching, with emissivity ranging from 0.1 to 0.2 for metal surfaces and from 0.8 to 0.95 for plastic surfaces, totaling one dimension, characterizing the ratio of radiation capacity to a blackbody.

[0067] The radiation intensity characteristic is calculated according to the Stefan-Boltzmann law. Radiation intensity equals emissivity multiplied by the Stefan-Boltzmann constant multiplied by the fourth power of the thermodynamic temperature, which is one-dimensional and characterizes the absolute radiation power. The radiation peak wavelength characteristic is calculated according to Wien's displacement law. The peak wavelength equals Wien's displacement constant divided by the thermodynamic temperature, which is one-dimensional and characterizes the characteristic wavelength corresponding to the temperature.

[0068] The pattern and trajectory features were extracted using the inter-frame differential optical flow method, and the velocity vector was extracted in four dimensions, including horizontal velocity, vertical velocity, acceleration and orientation angle, to characterize dynamic behavior features.

[0069] The concentric circle heat distribution characteristics were analyzed using radial temperature gradient analysis and fitted with Gaussian distribution parameters, totaling 3 dimensions, including center temperature, half width at half maximum (FWHM), and skewness, to characterize the spatial decay mode of the heat source.

[0070] (2) Feature extraction of visible light reflectance images; Texture features were extracted using the same method as the infrared channel: Gabor filtering was used to extract 24 dimensions, plus gray-level co-occurrence matrix extraction for 16 dimensions, for a total of 40 dimensions, which were used to characterize the surface microstructure.

[0071] Color features are extracted from moment features in both RGB and HSV spaces. Each space contains a first moment (mean), a second moment (variance), and a third moment (skewness), each with 3 channels, for a total of 18 dimensions, which are used to characterize color distribution statistics.

[0072] The shape features are represented using the same method as the infrared channel, with an additional 7 dimensions of Hu invariant moments, for a total of 12 dimensions, to characterize the rotation- and scaling-invariant shape features.

[0073] Spatial relationship feature encoding encodes the topological relationship between the target and scene elements, including top-bottom, left-right, and containment relationships, totaling 8 dimensions, used to represent contextual location information.

[0074] The statistical features include global mean, median, gray-level variance, and gray-level range, totaling four dimensions, used to characterize overall brightness and contrast. The specular reflection spot features employ adaptive threshold segmentation, using an improved version of the Otsu algorithm to filter regions with a circularity greater than 0.85, extracting the center x-coordinate, center y-coordinate, radius, and average gray-level, totaling four dimensions, used to characterize the lens's specular reflection features.

[0075] (3) Feature matching strategy; The real-time feature comparison process consists of five steps: The first step is spatial registration. Based on the calibration parameters, the infrared thermal imaging image with a resolution of 160×120 is downsampled to 160×120 with the visible light image with an original resolution of 640×480 and then aligned at the pixel level. The reprojection error is less than 1 pixel.

[0076] The second step is candidate region association, which maps the abnormal heat source regions detected by the infrared channel, with a temperature threshold of ambient temperature plus 2 degrees Celsius, to the corresponding regions in the visible light channel.

[0077] The third step is feature vector concatenation, which combines the 8-dimensional semantic features of the infrared channel, including shape, contour, texture, emissivity, radiation intensity, radiation peak wavelength, mode trajectory, and concentric circle distribution, with the 6-dimensional semantic features of the visible light channel, including texture, color, shape, spatial relationship, statistics, and specular reflection, into a 14-dimensional joint feature.

[0078] The fourth step is texture-specific matching. The 40-dimensional texture features extracted from the infrared and visible light channels are used to calculate the cosine similarity. The threshold is set to 0.85, and the surfaces are determined to be of the same material.

[0079] The fifth step is lightweight network inference, which involves inputting the registered dual-channel image and 14-dimensional joint features into the network and outputting 3-class classification probabilities.

[0080] The lightweight network model and its intelligent recognition method in this embodiment achieve significant technological advancements in the field of anti-spy camera detection through multimodal fusion and depth optimization. The model parameters are compressed using techniques such as depthwise separable convolution, bottleneck-inverted residual structure, and global pooling factor decomposition. By fusing infrared thermal imaging and visible light reflection in a dual-modal manner, combined with texture feature matching and channel attention mechanisms, three major technical bottlenecks are effectively addressed: missed detections by low-power cameras with temperature differences less than 0.3 degrees Celsius, failures in low-light environments and without infrared supplementary lighting, and high false alarm rates in complex heat source scenarios. It operates stably in typical private scenarios such as hotel rooms, public changing rooms, offices, and restrooms. The cascaded memory group convolutional network and feature feedback reuse mechanism enable caching and iterative updates of 3 frames of historical features, maintaining high-confidence output stability in continuous frame detection and reducing detection volatility to less than 5% in dynamic scenarios.

[0081] The parts of this invention not described in detail are well-known to those skilled in the art. Although illustrative specific embodiments of the invention have been described above to facilitate understanding by those skilled in the art, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they fall within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the inventive concept are protected. Although the invention has been described herein in conjunction with various embodiments, those skilled in the art can understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the description of the drawings, etc., in carrying out the claimed invention. In the specification, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude multiple instances. A single processor or other unit can implement several functions listed in the specification. While certain measures are described in different embodiments, this does not mean that these measures cannot be combined to produce good results.

[0082] Although the invention has been described in conjunction with specific features and embodiments, it is obvious that various modifications and combinations can be made therein without departing from the spirit and scope of the invention. Accordingly, this specification and drawings are merely illustrative of the invention and are considered to cover any and all modifications, variations, combinations, or equivalents within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if such modifications and modifications fall within the scope of the invention and its equivalents, the invention is also intended to include such modifications and modifications.

Claims

1. A portable multimodal anti-spy camera detection device, characterized in that, It includes a heterogeneous sensing module, an intelligent analysis module, and an interaction optimization module connected in sequence; wherein: The heterogeneous sensing module includes an infrared sensing unit, a supplementary lighting unit, and an optical sensing unit; the infrared sensing unit is used to acquire infrared thermal imaging images, the supplementary lighting unit is used to emit active infrared light, and the optical sensing unit is used to acquire visible light reflection images. The intelligent analysis module includes a feature extraction unit, a feature matching unit, and a lightweight network model connected in sequence. The feature extraction unit is used to receive infrared thermal imaging images from the infrared sensing unit and extract heat source features, and to receive visible light reflection images from the optical sensing unit and extract visible light reflection image features. The feature matching unit is used to compare and match the heat source features with the visible light reflection image features in real time. The lightweight network model includes an input preprocessing layer, a shallow feature extraction layer, a bottleneck inverted residual feature layer, an attention multi-scale residual layer, a lateral inhibition constraint layer, a global pooling factorization layer, a cascaded memory group convolutional network, a reconstruction network, and an output discriminant layer. The interactive optimization module includes a display unit and an alarm unit; the display unit is signal-connected to the intelligent analysis module and is used to output detection results; the alarm unit is signal-connected to the output end of the lightweight network model.

2. The portable multimodal anti-spy camera detection device according to claim 1, characterized in that, The feature matching unit performs matching between the heat source features and the visible light reflection image features, including texture analysis; the texture analysis includes wavelet transform and Gabor filtering, and calculates texture direction, frequency and contrast.

3. The portable multimodal anti-spy camera detection device according to claim 2, characterized in that, The texture includes gray-level co-occurrence matrix, roughness, contrast, orientation, line density, regularity, and coarsness.

4. The portable multimodal anti-spy camera detection device according to claim 1, characterized in that, The heat source characteristics include shape, outline, texture, emissivity, radiation intensity, radiation peak wavelength, pattern and trajectory, and concentric circle heat distribution characteristics.

5. The portable multimodal anti-spy camera detection device according to claim 1, characterized in that, The visible light reflection image features include texture, color, shape, spatial relationships, statistics, and specular reflection spot features.

6. The portable multimodal anti-spy camera detection device according to claim 1, characterized in that, The shallow feature extraction layer uses depthwise separable convolution instead of traditional convolution, and the depthwise separable convolution includes channel-wise convolution and point-wise convolution. The bottleneck inverted residual feature layer includes a bottleneck layer and an inverted residual block. The bottleneck layer adopts a three-segment structure of 1×1 convolution dimensionality reduction, 3×3 convolution feature extraction, and 1×1 convolution dimensionality increase. The inverted residual block adopts a structure of first increasing dimensionality, then depthwise convolution, and then pointwise dimensionality reduction, and is connected by short-circuited residuals. The attention-based multi-scale residual layer sets up a multi-scale receptive field, extracts large-scale contours and thermal distribution features from infrared images and small specular reflection spots and fine-grained texture features from visible light images in parallel, and performs adaptive weight allocation on the dual-modal feature channels through the SE channel attention mechanism.

7. The portable multimodal anti-spy camera detection device according to claim 1, characterized in that, The display unit supports dual-light fusion display mode; The alarm unit supports both vibration and sound alarms, and the threshold for the vibration and sound alarms is set to trigger when the heat source area is ≥3mm² and the reflected light intensity is ≥200lux.

8. An intelligent identification method for an anti-spy camera detection device, implemented using the portable multimodal anti-spy camera detection device according to any one of claims 1-7, characterized in that, Includes the following steps: S1. Extract infrared thermal imaging images and visible light reflection images through infrared sensing unit and optical sensing unit respectively; S2. Feature Extraction: Extracting heat source features from infrared thermal imaging images and extracting visible light reflection image features from visible light reflection images; S3. Feature Matching: Real-time comparison and matching of extracted heat source features and visible light reflection image features; S4. Lightweight network detection and recognition are achieved using a lightweight network model. S5. Output the detection results through the display unit and provide an alarm prompt through the alarm unit.

9. The method according to claim 8, characterized in that, In step S1, the infrared thermal imaging sensor has a resolution better than 160×120 and a temperature difference sensitivity of ±0.5℃; the thermal imaging sensor requires infrared supplementary light, and the power range of the infrared supplementary light is 5W to 20W; the optical sensing is achieved through a wide-angle optical lens with an FOV ≥ 120°.

10. The method according to claim 8, characterized in that, The feature matching in step S3 also includes texture analysis, which uses wavelet transform and Gabor filtering to obtain texture direction, frequency and contrast.

Images