Soft photosensitive detection method adaptive to environment

Abstract

The invention discloses a soft photosensitive detection method for an adaptive environment, and the method carries out the soft photosensitive detection of the adaptive environment through the continuous multi-frame performance of an EV value in a preset time domain window, thereby guaranteeing the normal switching between day and night, and avoiding the adverse effects. Therefore, compared with the existing hard photosensitive detection scheme, the method has the advantages of low cost, simple hardware structure, better performance, no temperature influence and the like; and compared with an existing soft photosensitive detection scheme, the environment adaptability is higher, the detection can be accurately carried out in various environments, and especially in a light reflection scene, the bad customer experience caused by switching of a day and night mode all the time is avoided.

Description

Technical Field

[0001] The present application relates to the field of intelligent detection, and more specifically, to a soft photosensitivity detection method for adaptive environments. Background Art

[0002] In the field of photosensitivity detection of intelligent imaging devices, existing technologies generally use static threshold segmentation or fixed time window exposure value (EV) monitoring mechanism to achieve day and night mode switching. This type of method uses a preset single brightness threshold to make a binary judgment on the current environment. For example, when the ambient light intensity exceeds the set threshold, it switches to day mode, otherwise it enters night mode.

[0003] However, this type of technology has exposed three key defects in practical applications: first, the static threshold cannot adapt to the gradual change of lighting scenes in dynamic environments (such as the dawn / dusk transition period), resulting in frequent oscillations or delayed responses in mode switching; second, short-term light source interference (such as the momentary illumination of vehicle high beams) can easily trigger the misjudgment mechanism, causing logical disorder in mode switching; third, existing methods lack in-depth mining of the temporal variation characteristics of EV values, and only judge through single frames or short-term averages, ignoring the trend characteristics of light intensity changes and the inertial requirements of mode switching. Especially in dynamic scenes, simply judging the mode based on a few fixed parameters is obviously not enough to cope with rapidly changing lighting conditions, which limits the flexibility and accuracy of the system.

[0004] In order to overcome the above shortcomings, an optimized soft photosensitivity detection method with adaptive environment is expected. Summary of the invention

[0005] In order to solve the above technical problems, the present application is proposed. The embodiment of the present application provides a soft photosensitivity detection method for an adaptive environment, which uses the continuous multi-frame performance of the EV value within a predetermined time domain window to perform soft photosensitivity detection for an adaptive environment, so as to ensure normal day and night switching and avoid adverse effects. In this way, compared with the existing hard photosensitivity detection solution, it has the advantages of low cost, simple hardware structure, better performance and no temperature impact; and compared with the existing soft photosensitivity detection solution, it has a stronger ability to adapt to the environment and can accurately detect it in various environments, especially in reflective scenes, to avoid the day and night mode switching back and forth, resulting in a bad customer experience.

[0006] According to one aspect of the present application, a method for soft photosensitive detection in an adaptive environment is provided, which includes:

[0007] Call the ISP interface to obtain the time queue of the EV value;

[0008] Extracting a current mode, where the current mode is a day mode or a night mode;

[0009] intercepting a target number of EV values ​​from the time queue of EV values ​​based on a predetermined time window to obtain a target time series of EV values;

[0010] Based on the target time series of the EV value and the current mode, it is determined whether to switch the mode.

[0011] Compared with the prior art, the present application provides a method for soft photosensitivity detection in an adaptive environment, which uses the continuous multi-frame performance of the EV value in a predetermined time domain window to perform soft photosensitivity detection in an adaptive environment, so as to ensure normal switching between day and night and avoid adverse effects. In this way, compared with the existing hard photosensitivity detection solution, it has the advantages of low cost, simple hardware structure, better performance and no temperature influence; and compared with the existing soft photosensitivity detection solution, it has stronger adaptability to the environment and can accurately detect in various environments, especially in reflective scenes, to avoid the day and night mode switching back and forth and causing a bad customer experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] By describing the embodiments of the present application in more detail in conjunction with the accompanying drawings, the above and other purposes, features and advantages of the present application will become more apparent. The accompanying drawings are used to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the accompanying drawings, the same reference numerals generally represent the same components or steps.

[0013] Figure 1 is a flow chart of a method for soft photosensitive detection of an adaptive environment according to an embodiment of the present application;

[0014] Figure 2 A schematic diagram of data flow of a soft photosensitivity detection method for an adaptive environment according to an embodiment of the present application;

[0015] Figure 3 is a flowchart of sub-step S4 of the soft photosensitivity detection method for an adaptive environment according to an embodiment of the present application;

[0016] Figure 4 This is a flowchart of sub-step S43 of the soft photosensitivity detection method for an adaptive environment according to an embodiment of the present application. DETAILED DESCRIPTION

[0017] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0018] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0019] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, any number of different modules can be used and run on the user terminal and / or server. The modules are only illustrative, and different aspects of the system and method can use different modules.

[0020] Flowcharts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, various steps may be processed in reverse order or simultaneously as required. Meanwhile, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0021] Below, the exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application, and it should be understood that the present application is not limited to the exemplary embodiments described here.

[0022] In the technical solution of the present application, a soft photosensitivity detection method for adaptive environment is proposed. Figure 1 Flow chart of a method for soft photosensitive detection in an adaptive environment according to an embodiment of the present application. Figure 2 FIG. 1 is a data flow diagram of a soft photosensitive detection method for an adaptive environment according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the soft photosensitivity detection method for an adaptive environment according to an embodiment of the present application includes the steps of: S1, calling an ISP interface to obtain a time queue of EV values; S2, extracting a current mode, wherein the current mode is a day mode or a night mode; S3, intercepting a target number of EV values ​​from the time queue of the EV values ​​based on a predetermined time window to obtain a target time series of EV values; S4, determining whether to switch modes based on the target time series of EV values ​​and the current mode.

[0023] In particular, the S1 calls the ISP interface to obtain the time queue of the EV value. Among them, the ISP (Image Signal Processor) interface is a bridge between the camera sensor and the image processing algorithm. It is not only responsible for converting the raw image data captured by the sensor into a usable image format, but also can provide a variety of key parameters including exposure value (EV), white balance, color correction, etc. When implementing the time queue for calling the ISP interface to obtain the EV value, first, the system needs to initialize the connection with the ISP and ensure that the required API or interface function can be accessed so that the exposure value in the current scene can be extracted from the ISP. This usually means that developers need to be familiar with the SDK or API documentation provided by the ISP supplier to understand how to correctly set parameters and how to initiate query requests for specific data. It is worth noting that in actual operation, considering real-time and efficiency issues, it may also be necessary to optimize the data acquisition process. For example, asynchronous I / O or multi-threading technology is used to avoid blocking the main thread to ensure that the response speed of the system is not affected.

[0024] In particular, S2 extracts the current mode, and the current mode is day mode or night mode. It should be understood that in the process of implementing the extraction of the current mode, the first thing to be clear is that the system needs to be able to recognize and distinguish between the two states of day mode and night mode. In one example, after obtaining the EV value, the system will determine the current working mode based on a preset threshold or by analyzing the trend of EV value changes over a period of time. For example, in a simple application scenario, if the EV value corresponding to the current ambient light intensity is higher than a preset threshold, it can be considered as day mode; conversely, if it is lower than the threshold, it is determined to be night mode.

[0025] In particular, the S3 intercepts a target number of EV values ​​from the time queue of the EV values ​​based on a predetermined time window to obtain a target time series of EV values. Different from the traditional single-frame or short-time average analysis, this method can more accurately reflect the changing trend of the ambient light, so as to perform trend analysis and avoid misjudgment caused by short-term light source interference.

[0026] In particular, the S4 determines whether to switch modes based on the target time series of the EV values ​​and the current mode. In the first embodiment, if the current mode is the day mode, each EV value in the target time series of the EV values ​​is compared with a calibrated daytime threshold value to obtain a target time series of day mode comparison results; based on the target time series of the day mode comparison results, it is determined whether to switch modes; if the current mode is the night mode, each EV value in the target time series of the EV values ​​is compared with a calibrated night threshold value to obtain a target time series of night mode comparison results; based on the target time series of the night mode comparison results, it is determined whether to switch modes. That is, in response to all the day mode comparison results in the target time series of the day mode comparison results showing that the EV value changes are not within the range of the day mode, it is determined to switch to the night mode; in response to all the night mode comparison results in the target time series of the night mode comparison results showing that the EV value changes are not within the range of the night mode, it is determined to switch to the day mode.

[0027] In the first embodiment, for example, during the dusk period, as the sun gradually sets, the ambient light will gradually weaken. If you only rely on the brightness value at a certain moment to determine whether to switch from day mode to night mode, you may make a wrong decision due to the instantaneous brightness fluctuation at a certain moment. On the contrary, by analyzing the EV value changes over a continuous period of time (for example, 12 consecutive frames) and comparing it with the preset threshold range of day or night mode, the real mode switching needs can be identified more accurately.

[0028] In an embodiment of the present application, when the system detects that the EV values ​​of 12 consecutive frames are significantly lower than the set day mode threshold range, it indicates that the current ambient light is no longer suitable for maintaining the working state of the day mode, and the system will automatically switch to the night mode to adapt to the lower lighting conditions. Similarly, if in the night mode, the system detects that the EV values ​​of 12 consecutive frames exceed the upper threshold of the night mode, indicating that the ambient light has increased to a sufficient degree to require switching back to the day mode. Specifically, the system calculates the difference between the EV value of each frame in the 12 frames and the previous frame. If these differences indicate that the EV value continues to decrease, and the EV values ​​of all 12 frames are lower than the threshold range of the day mode, the system will determine that it is necessary to switch from the day mode to the night mode. This approach enables the system to reduce false triggers caused by short-term light sources or other atypical lighting conditions while ensuring high accuracy, thereby providing a smoother and more natural user experience. In addition, in other embodiments, by analyzing the time series of EV values ​​in combination with a deep learning model, the system can further explore the potential patterns in the data and improve the intelligence of prediction and response.

[0029] In a specific embodiment, the complete implementation process of the first embodiment is as follows:

[0030] First, call the ISP interface to obtain the EV value.

[0031] Then, determine whether the current mode is day mode or night mode. If it is day mode, compare the EV value with the calibrated night threshold. If it is night mode, compare the EV value with the calibrated day threshold.

[0032] Then, when the EV value changes are not within the current mode range for 12 consecutive frames, the mode will be switched. That is, if it is day mode, the EV value is greater than the night threshold for 12 consecutive frames, it will switch to night mode. If it is night mode, the EV value is less than the day threshold for 12 consecutive frames, it will switch to day mode. If there are no 12 consecutive frames where the EV value meets the above conditions, the current mode will continue. A sign will be sent to indicate whether day and night switching is required.

[0033] Then, confirm whether lock 2 is turned on. If not, proceed to the next step to determine the day and night switching mark. If it is turned on, record the current time and EV value, and compare them with the time and EV value when lock 2 was marked. If the lock time exceeds 1 hour or the EV value changes by more than 40%, clear all marked data and switch to day mode. If not, continue to maintain night mode.

[0034] Next, determine whether a day-night switch is required. If not, keep the current mode unchanged and no subsequent operation is required. If day-night switch is required, if it is night to day, open lock 1, record the time and EV value, and switch to day mode. If it is day to night, proceed to the next step.

[0035] Then, if it turns from day to night and lock 1 is turned on, proceed to the next step. If not, clear the data marks such as lock 1 and switch to night mode.

[0036] Finally, if day turns to night and lock 1 is turned on, and this happens for 10 consecutive times, lock 2 will be turned on; if not, the data marks of lock 1 will be cleared and the system will switch to night mode.

[0037] In particular, in the technical solution of the present application, it can also be implemented in other ways: determining whether to switch modes based on the target time series of the EV value and the current mode.

[0038] Specifically, in the second embodiment of the present application, Figure 3As shown, the S4 includes: S41, performing one-hot encoding on the current mode to obtain a current mode one-hot embedded coding vector; S42, inputting the target time series of the EV value into the EV value temporal change feature extractor based on the GRU model to obtain an EV value temporal association implicit coding feature vector; S43, performing EV value-pattern deep implicit alignment analysis on the EV value temporal association implicit coding feature vector and the current mode one-hot embedded coding vector to obtain an EV temporal feature-pattern semantic feature comparison coding vector; S44, determining whether to switch the mode based on the EV temporal feature-pattern semantic feature comparison coding vector.

[0039] Specifically, the S41 performs one-hot encoding on the current mode to obtain a one-hot embedding encoding vector of the current mode. It should be understood that in the mode switching decision process of the traditional photosensitive detection system, the mode state is usually represented by a simple Boolean scalar (such as 0 / 1). This discretized low-dimensional representation is difficult to carry the semantic difference information between modes, resulting in a spatial mismatch between mode semantics and data features during subsequent feature fusion. Especially in a dynamic lighting environment, when the system is in daytime mode, if it encounters a short-term strong light interference (such as a backlight scene that suddenly appears at the entrance of a tunnel), the existing technology directly marks the current mode as a single symbol of "daytime" and inputs it into the decision model, which cannot effectively distinguish the stable state or abnormal transition state of the mode. One-hot encoding is a commonly used data preprocessing technology that can convert classified data into a form that is easy to process by machine learning algorithms. In this application scenario, one-hot encoding is used to vectorize and embed the current mode, and "daytime mode" and "night mode" are respectively mapped to independent basis vectors in a high-dimensional orthogonal vector space (such as [1,0] and [0,1]), which essentially constructs a mathematical representation system of mode semantics. This encoding method not only retains the mutually exclusive characteristics of pattern categories, but also explicitly expresses the logical opposition between patterns through the geometric relationship of vector space, avoiding mutual influence between numerical values.

[0040] Specifically, in S42, the target time series of the EV value is input into the EV value time series change feature extractor based on the GRU model to obtain the EV value time series associated implicit coding feature vector. Among them, GRU is a kind of gated recurrent neural network, and GRU (Gated Recurrent Unit) is a special recurrent neural network (RNN), which is particularly good at processing sequence data and can effectively learn long-term dependencies. This means that the GRU model can deeply explore the inherent laws of EV value changes over time, identify the changing trend of lighting conditions and their inertial requirements, and thus provide more reliable data support for mode switching. By inputting the target time series of the EV value into the EV value time series change feature extractor based on the GRU model, the system can not only consider the current EV value, but also combine the EV value changes in the past period of time for comprehensive evaluation. This ability enables the system to foresee the moment when mode switching may be needed earlier when the light gradually changes, and it can also remain stable when encountering sudden light source interference to avoid unnecessary mode switching. For example, during dusk, as the sun sets and the ambient light gradually weakens, the GRU model can continuously monitor and analyze the changes in EV values ​​for 12 consecutive frames during this period, extracting useful information to assist in determining whether to switch from day mode to night mode. Based on this, the system can more intelligently decide whether to switch modes, thereby ensuring a smooth and stable user experience.

[0041] Specifically, the S43 performs EV value-mode deep implicit alignment analysis on the EV value time series associated implicit coding feature vector and the current mode unique hot embedding coding vector to obtain an EV time series feature-mode semantic feature comparison coding vector. It should be understood that although analyzing the change trend of the EV value over time alone can provide important information about the change of ambient light, in order to make an accurate mode switching decision, the current working mode state also needs to be considered. This is because there are significant differences in lighting conditions and their changing patterns between day and night modes. Relying solely on changes in EV values ​​may not be able to fully capture these differences. The EV value time series associated implicitly encoded feature vector contains trend information of dynamic lighting changes (such as the acceleration of light intensity changes, fluctuation period, etc.), and the current mode one-hot embedded encoding vector solidifies the discrete semantics of the current system state (such as the binary label of "day mode" or "night mode"). The heterogeneity of the two in the feature space may cause the system to be unable to accurately evaluate "whether the inertia conditions for switching to night mode are met". If the two features are simply spliced ​​together, not only will redundant interference be introduced (such as the discretized labels in the mode encoding do not match the continuous vector dimensions of the EV time series), but it is also likely to ignore key decision clues (such as the weighted impact of the current mode duration on the trend credibility) due to the lack of high-order interaction modeling.

[0042] In order to overcome the above problems and to more comprehensively understand the overall situation of the ambient light and its impact on the current working mode, in the technical solution of the present application, the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector are subjected to EV value-mode deep implicit alignment analysis to obtain the EV temporal feature-mode semantic feature comparison coding vector. By performing deep implicit alignment analysis on the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector, the effective fusion of two different types of data can be achieved, thereby capturing deeper semantic information.

[0043] In this process, the core anchor point information in the feature representation is extracted through the autocorrelation decoupling technology, and the interaction between features is accurately quantified and characterized by means of the two-level interactive modeling of feature granularity and eigenvalue granularity. In this process, the redundant noise is first stripped off by autocorrelation decoupling, and the core trend anchor points in the EV sequence (such as the continuously decreasing light intensity gradient) and the semantic anchor points in the mode vector (such as the mode switching hysteresis coefficient) are extracted. Subsequently, at the feature granularity interaction level, the system semantically couples the macro constraints of the mode semantics (such as "the daytime mode should have light intensity stability") with the micro fluctuation characteristics of the EV sequence (such as the sudden change of light intensity but the overall high level), and identifies the potential conflict between the sudden drop of EV at the tunnel entrance and the semantics of the daytime mode; at the same time, at the eigenvalue granularity level, the nonlinear dependence between the light intensity value jump and the mode state value (such as the disturbance intensity of the EV spike caused by the instantaneous irradiation of the high beam on the [1,0] mode vector) is captured through element-by-element response modeling. This dual-level alignment mechanism is particularly critical in neon strobe scenes: the feature granularity interaction identifies that the EV fluctuation caused by the strobe does not change the overall light intensity distribution, while the eigenvalue granularity analysis quantifies the weight of the impact of a single pulse on the pattern vector, and finally suppresses false triggering through disturbance correction. In this way, the adaptability and accuracy of the system in complex dynamic environments are improved.

[0044] More specifically, in a specific example of the present application, Figure 4As shown, the S43 includes: S431, constructing a semantic autocorrelation association matrix of the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector to obtain the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode one-hot embedding coding feature semantic autocorrelation association matrix; S432, performing fine-grained response interaction encoding based on EV value-pattern feature anchoring on the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode one-hot embedding coding feature semantic autocorrelation association matrix to obtain the EV temporal feature-pattern semantic feature granular response interaction coding vector and the EV temporal feature-pattern semantic feature value. Granularity response interaction coding vector; S433, perform feature perturbation correction on the EV timing feature-pattern semantic feature granularity response interaction coding vector and the EV timing feature-pattern semantic feature value granularity response interaction coding vector to obtain the corrected EV timing feature-pattern semantic feature granularity response interaction coding vector and the corrected EV timing feature-pattern semantic feature value granularity response interaction coding vector; S434, fuse the corrected EV timing feature-pattern semantic feature granularity response interaction coding vector and the corrected EV timing feature-pattern semantic feature value granularity response interaction coding vector to obtain the EV timing feature-pattern semantic feature comparison coding vector.

[0045] More specifically, the S431 constructs a semantic autocorrelation matrix of the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector to obtain the EV value temporal association implicit feature semantic autocorrelation matrix and the current mode one-hot embedding coding feature semantic autocorrelation matrix. By calculating the autocorrelation matrix of the EV temporal feature (such as a 128×128 correlation intensity map), the system can capture the temporal dependence of illumination changes—for example, the high weight values ​​on the diagonal of the matrix may correspond to the persistence of the illumination attenuation trend, while the peaks on the non-diagonal reflect the local fluctuation association caused by sudden interference events (such as flashing lights). At the same time, when constructing an autocorrelation matrix (such as a 2×2 semantic association map) for the mode one-hot encoding, although its dimension is low, the implicit inertia of the mode state (such as the accumulation of autocorrelation intensity corresponding to the "day mode" label over the duration) can be parsed through matrix operations. This structured expression enables the feature vector to be upgraded from a flattened numerical sequence to an information carrier carrying spatiotemporal semantics, significantly improving the robustness and interpretability of feature expression.

[0046] In one example of the present application, the semantic autocorrelation association matrix of the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector is constructed by the following semantic autocorrelation association calculation formula to obtain the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode one-hot embedding coding feature semantic autocorrelation association matrix; wherein the semantic autocorrelation association calculation formula is:

[0047]

[0048]

[0049] in, represents the EV value temporal association implicit coding feature vector, represents the one-hot embedding encoding vector of the current mode, and is a feature mapping function, such as a linear mapping or a nonlinear kernel function, Represents the EV value temporal association implicit feature semantic autocorrelation association matrix, Represents the semantic autocorrelation association matrix of the one-hot embedding encoding feature of the current mode.

[0050] More specifically, the S432 performs fine-grained response interaction encoding based on EV value-pattern feature anchoring on the EV value temporal association implicit feature semantic autocorrelation association matrix and the current pattern one-hot embedded coding feature semantic autocorrelation association matrix to obtain the EV temporal feature-pattern semantic feature granular response interaction encoding vector and the EV temporal feature-pattern semantic feature value granular response interaction encoding vector.

[0051] In an embodiment of the present application, first, the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode unique hot embedding coding feature semantic autocorrelation association matrix are respectively input into the EV value-mode feature anchoring network based on autocorrelation decoupling to obtain the EV value temporal association implicit feature core information anchoring coding vector and the current mode unique hot embedding coding feature core information anchoring coding vector. That is, the core information anchor point of each feature vector is extracted through the core information anchoring network based on autocorrelation decoupling, highlighting those parts that are highly related to the core semantics and removing redundant information.

[0052] In one example, through means such as feature distillation, the model only needs to interact with a small number of globally representative feature anchors, thereby improving computational efficiency and the accuracy of feature modeling. For example, during dusk, as the sun gradually sets and the ambient light gradually weakens, the system not only monitors the changing trend of the EV value, but also comprehensively evaluates whether it needs to switch to night mode based on the fact that it is currently in daytime mode. Through the above process, the system can better understand the deep semantic relationship between the EV value time series features and the current mode. Through the core information anchoring network based on autocorrelation decoupling, the system can effectively extract the most representative core information from complex feature data to form a structurally refined core anchor representation. This representation not only helps to constrain the semantic consistency of feature representation in compressed space, but also improves the generalization ability of the model by reducing redundant information. The ultimate goal of this is to ensure that the system can respond promptly and accurately under complex dynamic lighting conditions, avoiding the risk of misjudgment caused by single parameter judgment. By processing the EV value-mode feature anchoring network based on autocorrelation decoupling of the semantic autocorrelation matrix of the implicit features of the EV value temporal association and the semantic autocorrelation matrix of the unique hot embedding coding features of the current mode, the system can generate a more accurate core information anchor coding vector. This not only solves the problem that the traditional joint coding method is difficult to distinguish between important features and redundant features, but also improves the generalization ability of the model by constraining information entropy.

[0053] In one example of the present application, the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode unique hot embedding coding feature semantic autocorrelation association matrix are respectively input into the EV value-mode feature anchoring network based on autocorrelation decoupling to obtain the EV value temporal association implicit feature core information anchoring coding vector and the current mode unique hot embedding coding feature core information anchoring coding vector; wherein, the feature anchoring formula is:

[0054]

[0055]

[0056] in, Indicates feature decoupling, They represent the first, second, and third semantic autocorrelation matrix of the EV value temporal association implicit feature. and The row vectors are used as the EV value temporal association implicit feature semantic autocorrelation association vectors, and are the EV value temporal association latent feature semantic weight matrix and the EV value temporal association latent feature semantic bias vector, is the semantic modulation vector of the implicit feature associated with the temporal sequence of EV values, is the anchoring factor of the implicit feature semantic core information of the EV value temporal association, express function, The anchor weight of the implicit feature semantic core information associated with the EV value temporal sequence is is the number of row vectors in the semantic autocorrelation matrix of the EV value temporal association implicit feature, represents the anchor coding vector of the implicit feature core information associated with the EV value time series, They represent the first, second, and third semantic autocorrelation matrix of the current mode one-hot embedding encoding feature respectively. and The row vector is used as the current mode one-hot embedding encoding feature semantic autocorrelation vector, and are the semantic weight matrix of the current mode one-hot embedding encoding feature and the semantic bias vector of the current mode one-hot embedding encoding feature, is the semantic modulation vector of the current mode one-hot embedding encoding feature, The anchor factor of the semantic core information of the current mode is the single-hot embedding encoding feature. express function, Anchor weights for the semantic core information of the current mode’s one-hot embedding encoding feature, is the number of row vectors in the semantic autocorrelation matrix of the current mode one-hot embedding encoding feature, Represents the core information anchor encoding vector of the current mode one-hot embedding encoding feature.

[0057] Next, the EV value temporal association implicit feature core information anchor coding vector and the current mode one-hot embedding coding feature core information anchor coding vector are input into the feature granularity response interactive encoder to obtain the EV temporal feature-pattern semantic feature granularity response interactive coding vector; further, the EV value temporal association implicit feature core information anchor coding vector and the current mode one-hot embedding coding feature core information anchor coding vector are input into the feature value granularity response interactive encoder to obtain the EV temporal feature-pattern semantic feature value granularity response interactive coding vector.

[0058] It should be understandable that the EV time series feature core information anchor coding vector (such as the compressed light intensity change gradient and fluctuation period) and the mode semantic core anchor coding vector (such as the mode switching hysteresis coefficient) carry information at different abstract levels respectively - the former reveals the overall inertia and potential trend of illumination changes, and the latter encapsulates the dependency of mode states on historical decisions. If the two are directly fused, semantic confusion may occur due to mismatch of information granularity.

[0059] In the technical solution of the present application, the EV value temporal association implicit feature core information anchor coding vector and the current mode one-hot embedded coding feature core information anchor coding vector are input into the feature granularity response interactive encoder to obtain the EV temporal feature-pattern semantic feature granularity response interactive coding vector; and the EV value temporal association implicit feature core information anchor coding vector and the current mode one-hot embedded coding feature core information anchor coding vector are input into the feature value granularity response interactive encoder to obtain the EV temporal feature-pattern semantic feature value granularity response interactive coding vector.

[0060] In one example, the feature-granularity response interactive encoder focuses on the coupling relationship at the semantic unit level, and associates the macro constraints of mode semantics (such as "day mode must meet the light intensity mean threshold") with the overall behavior pattern of EV time series features (such as the light intensity attenuation slope) for modeling. For example, under the interference of car light reflection in rainstorm weather, the encoder compares the mode vector [1,0] (day mode) with the global stability index of the EV sequence, and identifies that despite the existence of EV spikes, the overall light intensity is still higher than the night threshold, thereby suppressing false switching. The feature value granularity response interactive encoder goes deep into the numerical level and analyzes the disturbance intensity of a single EV mutation on the mode vector. For example, in the neon light strobe scene, the encoder analyzes the nonlinear influence of pulse light on each dimension of the [1,0] mode vector element by element, quantifies the dynamic dependence between the flickering light intensity and the mode state, and avoids misjudging high-frequency pulses as mode switching signals. This multi-level interactive modeling method enables the system to explore high-order feature dependencies in complex feature spaces, which not only solves the problem that traditional joint encoding methods are difficult to distinguish between important features and redundant features, but also improves the generalization ability of the model by constraining information entropy.

[0061] In one example of the present application, the EV value temporal association implicit feature core information anchor coding vector and the current mode unique hot embedding coding feature core information anchor coding vector are respectively input into the feature granularity response interaction encoder and the feature value granularity response interaction encoder to obtain the EV temporal feature-mode semantic feature granularity response interaction coding vector and the EV temporal feature-mode semantic feature value granularity response interaction coding vector; wherein the response interaction coding formula is:

[0062]

[0063]

[0064] in, For positional addition, For positional addition, and are the feature interaction weight matrix and feature interaction bias vector respectively, for function, represents the EV temporal feature-pattern semantic feature granularity response interaction encoding vector, Represents the EV temporal feature-pattern semantic feature value granularity response interaction encoding vector.

[0065] More specifically, the S433 performs feature perturbation correction on the EV time series feature-pattern semantic feature granularity response interaction coding vector and the EV time series feature-pattern semantic feature value granularity response interaction coding vector to obtain the corrected EV time series feature-pattern semantic feature granularity response interaction coding vector and the corrected EV time series feature-pattern semantic feature value granularity response interaction coding vector. In particular, it should be understood that under complex lighting conditions, the interaction between different features may cause unstable perturbations in the feature manifold interface. For example, in a specific scenario, assuming that during the gradual transition period of dusk, the continuous light attenuation trend captured by the feature granularity interaction may produce semantic deviations due to local EV value mutations caused by cloud cover, and the uncorrected feature value granularity analysis is prone to amplify such short-term fluctuations into decision noise, resulting in irrational oscillations in the manifold interface, just like the scattering interference caused by light passing through a prism, destroying the system's overall cognitive consistency of the laws of environmental evolution. In the technical solution of the present application, the EV time series feature-pattern semantic feature granularity response interaction coding vector and the EV time series feature-pattern semantic feature value granularity response interaction coding vector are feature-corrected to obtain the corrected EV time series feature-pattern semantic feature granularity response interaction coding vector and the corrected EV time series feature-pattern semantic feature value granularity response interaction coding vector. In this process, the feature value granularity coding is used as a dynamic anchor point to gradient-domesticate the semantic diffusion path of the feature granularity: for example, in a high-noise environment at dusk during a rainstorm, the high-frequency fluctuation of the EV value caused by the reflection of raindrops is calculated through the disturbance contribution factor and quantified as a tolerable local disturbance weight, and then the overall trend coding of the feature granularity is modulated by a growth exponential, which not only retains the core trend expression of "continuous weakening of light intensity", but also suppresses the abnormal diffusion of random fluctuations on the manifold interface. Through the above correction process, the system can more accurately determine when is the best time to switch modes and avoid misjudgments caused by light fluctuations or short-term light source interference.

[0066] In other words, in this specific example, the EV time series feature-pattern semantic feature granularity response interaction coding vector and the EV time series feature-pattern semantic feature value granularity response interaction coding vector are subjected to feature perturbation correction by the following correction formula to obtain the corrected EV time series feature-pattern semantic feature granularity response interaction coding vector and the corrected EV time series feature-pattern semantic feature value granularity response interaction coding vector; wherein the correction formula is:

[0067]

[0068]

[0069] in, and They are and The corresponding eigenvalues ​​of and They are and The corresponding eigenvalues ​​of Represents the cosine function.

[0070] More specifically, the S434 fuses the corrected EV timing feature-pattern semantic feature granularity response interaction coding vector and the corrected EV timing feature-pattern semantic feature value granularity response interaction coding vector to obtain the EV timing feature-pattern semantic feature comparison coding vector. It should be understood that in order to achieve more comprehensive and accurate pattern recognition and switching decisions, it is necessary to effectively integrate the two different levels of information, the corrected EV timing feature-pattern semantic feature granularity and the corrected EV timing feature-pattern semantic feature value granularity, to generate an EV timing feature-pattern semantic feature comparison coding vector that can reflect both macro trends and capture micro details. The EV timing feature-pattern semantic feature comparison coding vector can not only capture the trend characteristics of the EV value changing over time, but also reflect the specific state in the current mode, thereby providing more reliable data support for mode switching. In one example of the present application, the corrected EV timing feature-pattern semantic feature granularity response interaction coding vector and the corrected EV timing feature-pattern semantic feature value granularity response interaction coding vector are fused by the following fusion formula to obtain the EV timing feature-pattern semantic feature comparison coding vector; wherein the fusion formula is:

[0071]

[0072] in, represents feature fusion, Represents the EV temporal feature-pattern semantic feature comparison encoding vector.

[0073] Specifically, the S44 determines whether to switch modes based on the EV timing feature-mode semantic feature comparison coding vector. That is, in the technical solution of the present application, the EV timing feature-mode semantic feature comparison coding vector is input into the classifier-based mode switching intelligent controller to obtain a control instruction, and the control instruction is used to indicate whether to switch modes. In this way, the system can make more accurate judgments based on the actual changes in the current ambient light and the working state of the device. For example, in the evening period, as the sun gradually sets, the ambient light gradually weakens. The system not only needs to monitor the changing trend of the EV value, but also needs to combine the fact that it is currently in daytime mode to comprehensively evaluate whether it needs to switch to night mode. In this process, the input coding vector can be analyzed according to the pre-trained model, and the corresponding control instructions can be output accordingly. In this way, not only the accuracy of the mode switching decision is improved, but also the flexibility and adaptability of the system are enhanced. In the face of complex and changeable lighting conditions, such as a rapid sunset process or a sudden short-term strong light source interference, the system can respond quickly based on the real-time acquired data, avoiding the problem of misjudgment or slow response caused by simple reliance on fixed parameters.

[0074] In the technical solution of the present application, the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector represent the local temporal depth implicit association features of the EV value and the low-dimensional one-hot embedding coding features of the current mode, respectively. When performing feature interaction response analysis based on core information anchoring, the difference in feature modality and feature information expression order will cause the obtained EV temporal feature-pattern semantic feature comparison coding vector to have fine-grained regression imbalance, that is, the fine-grained feature expression of the EV temporal feature-pattern semantic feature comparison coding vector has target domain classification regression disorder, which affects the classification judgment accuracy of the control instructions obtained by the mode switching intelligent controller based on the classifier.

[0075] Based on this, in a preferred embodiment of the present application, the EV time series feature-mode semantic feature comparison encoding vector is input into a classifier-based mode switching intelligent controller to obtain a control instruction, including:

[0076] The feature values ​​of any two positions in the EV temporal feature-pattern semantic feature comparison encoding vector are quantized in metric space to obtain the EV temporal feature-pattern semantic feature comparison difference quantization matrix, which is expressed as:

[0077]

[0078]

[0079]

[0080] in, and represents the feature values ​​of any two positions in the EV temporal feature-pattern semantic feature comparison encoding vector, and represents the first weight hyperparameter and the second weight hyperparameter, such as , Of course, the above is just an example. The specific weight hyperparameters can be initially set and tuned through grid search or other methods. Represents the first EV temporal feature-pattern semantic feature measurement matrix The value of the position, Represents the second EV temporal feature-pattern semantic feature measurement matrix The value of the position, In the matrix representing the difference between EV temporal features and pattern semantic features, The value of the position.

[0081] The EV temporal feature-pattern semantic feature comparison coding vector is subjected to autocorrelation significant correlation coding to obtain the EV temporal feature-pattern semantic feature comparison intrinsic correlation matrix, which is expressed as:

[0082]

[0083] in, represents the EV temporal feature-pattern semantic feature comparison encoding vector, represents the length of the EV temporal feature-pattern semantic feature comparison encoding vector, represents matrix multiplication, represents the transpose of a vector, Represents the EV temporal features-pattern semantic features contrast intrinsic correlation matrix.

[0084] Based on the EV temporal feature-pattern semantic feature comparison difference quantization matrix, the EV temporal feature-pattern semantic feature comparison encoding vector is subjected to primary projection modulation to obtain the EV temporal feature-pattern semantic feature comparison primary projection modulation vector, which is expressed as:

[0085]

[0086] in, represents the Hadamard product, represents the activation function, represents the quantified matrix of the difference between EV temporal features and pattern semantic features, Represents the primary projection modulation vector comparing EV temporal features and pattern semantic features.

[0087] Based on the EV temporal feature-pattern semantic feature comparison intrinsic association matrix, the EV temporal feature-pattern semantic feature comparison primary projection modulation vector is subjected to secondary projection modulation to obtain the EV temporal feature-pattern semantic feature comparison hierarchical aggregation projection modulation vector, which is expressed as:

[0088]

[0089] in, Represents the hierarchical aggregation projection modulation vector of EV temporal features-pattern semantic features.

[0090] The EV temporal feature-pattern semantic feature comparison hierarchical aggregation projection modulation vector and the EV temporal feature-pattern semantic feature comparison encoding vector are dynamically fused to obtain an optimized EV temporal feature-pattern semantic feature comparison encoding vector, which is expressed as:

[0091]

[0092] in, and represents the third and fourth weight hyperparameters, for example , Of course, the above is just an example. The specific weight hyperparameters can be initially set and tuned through grid search or other methods. Represents the optimized EV temporal feature-pattern semantic feature comparison encoding vector.

[0093] The optimized EV timing feature-mode semantic feature comparison encoding vector is input into the classifier-based mode switching intelligent controller to obtain the control instruction.

[0094] Accordingly, in this preferred embodiment, the first-order projection representation of the difference quantization tensor based on the EV time series feature-pattern semantic feature comparison coding vector is used to perform a multi-scale hierarchical second-order projection representation of the complete isomorphic generation of the intrinsic association of the EV time series feature-pattern semantic feature comparison coding vector, and the coupled aggregation kernel offset is used to compensate for the synergistic mismatch negative interference factor to enhance the intrinsic mode judgment strength of the EV time series feature-pattern semantic feature comparison coding vector within the constrained similarity framework, that is, the significance of the intrinsic mode as an example for the generative regression judgment. In this way, the accuracy of the control instruction obtained by the input of the classifier-based mode switching intelligent controller is improved.

[0095] In summary, the soft photosensitivity detection method for an adaptive environment according to the embodiment of the present application is explained, which uses the continuous multi-frame performance of the EV value in a predetermined time domain window to perform soft photosensitivity detection for an adaptive environment to ensure normal switching between day and night and avoid adverse effects. In this way, compared with the existing hard photosensitivity detection solution, it has the advantages of low cost, simple hardware structure, better performance and no temperature influence; and compared with the existing soft photosensitivity detection solution, it has a stronger ability to adapt to the environment and can accurately detect in various environments, especially in reflective scenes, to avoid the day and night mode switching back and forth and causing a bad customer experience.

[0096] The embodiments of the present disclosure have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A soft photosensitive detection method for adaptive environment, characterized in that: include: Call the ISP interface to obtain the time queue of the EV value; Extracting a current mode, where the current mode is a day mode or a night mode; intercepting a target number of EV values ​​from the time queue of EV values ​​based on a predetermined time window to obtain a target time series of EV values; Based on the target time series of the EV value and the current mode, it is determined whether to switch the mode.

2. The method for soft photosensitive detection of an adaptive environment according to claim 1, characterized in that: The predetermined time window is 12 consecutive frames.

3. The method for soft photosensitive detection of an adaptive environment according to claim 2, characterized in that: Determining whether to switch modes based on the target time series of the EV value and the current mode includes: If the current mode is a day mode, comparing each EV value in the target time series of EV values ​​with a calibrated daytime threshold value to obtain a target time series of daytime mode comparison results; Determining whether to switch modes based on the target time sequence of the daytime mode comparison result; If the current mode is a night mode, each EV value in the target time series of EV values ​​is compared with a calibrated night threshold value to obtain a target time series of night mode comparison results; Based on the target time series of the night mode comparison result, it is determined whether to switch the mode.

4. The method for soft photosensitive detection of an adaptive environment according to claim 3, characterized in that: In response to all day mode comparison results in the target time sequence of the day mode comparison results showing that the EV value change is not within the range of the day mode, it is determined to switch to the night mode; in response to all night mode comparison results in the target time sequence of the night mode comparison results showing that the EV value change is not within the range of the night mode, it is determined to switch to the day mode.

5. The method for soft photosensitive detection of an adaptive environment according to claim 1, characterized in that: Determining whether to switch modes based on the target time series of the EV value and the current mode further includes: One-hot encoding the current mode to obtain a current mode one-hot embedding encoding vector; Inputting the target time series of the EV value into the EV value time series change feature extractor based on the GRU model to obtain the EV value time series associated implicit coding feature vector; Performing EV value-mode deep implicit alignment analysis on the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector to obtain an EV temporal feature-mode semantic feature comparison coding vector; Based on the EV temporal feature-mode semantic feature comparison coding vector, determine whether to switch the mode.

6. The method for soft photosensitive detection of an adaptive environment according to claim 5, characterized in that: Performing EV value-mode deep implicit alignment analysis on the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector to obtain an EV temporal feature-mode semantic feature comparison coding vector, including: Constructing a semantic autocorrelation association matrix of the EV value temporal association implicit coding feature vector and the current mode one-hot embedding coding vector to obtain an EV value temporal association implicit feature semantic autocorrelation association matrix and a current mode one-hot embedding coding feature semantic autocorrelation association matrix; Perform fine-grained response interaction encoding based on EV value-pattern feature anchoring on the EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode unique hot embedding coding feature semantic autocorrelation association matrix to obtain EV temporal feature-pattern semantic feature granular response interaction encoding vector and EV temporal feature-pattern semantic feature value granular response interaction encoding vector; Performing feature perturbation correction on the EV temporal feature-pattern semantic feature granularity response interaction coding vector and the EV temporal feature-pattern semantic feature value granularity response interaction coding vector to obtain a corrected EV temporal feature-pattern semantic feature granularity response interaction coding vector and a corrected EV temporal feature-pattern semantic feature value granularity response interaction coding vector; The corrected EV temporal feature-pattern semantic feature granularity response interaction coding vector and the corrected EV temporal feature-pattern semantic feature value granularity response interaction coding vector are fused to obtain the EV temporal feature-pattern semantic feature comparison coding vector.

7. The method for soft photosensitive detection of an adaptive environment according to claim 6, characterized in that: The EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode unique hot embedding coding feature semantic autocorrelation association matrix are subjected to fine-grained response interaction encoding based on EV value-mode feature anchoring to obtain an EV temporal feature-mode semantic feature granular response interaction encoding vector and an EV temporal feature-mode semantic feature value granular response interaction encoding vector, including: The EV value temporal association implicit feature semantic autocorrelation association matrix and the current mode one-hot embedding coding feature semantic autocorrelation association matrix are respectively input into the EV value-mode feature anchoring network based on autocorrelation decoupling to obtain the EV value temporal association implicit feature core information anchoring coding vector and the current mode one-hot embedding coding feature core information anchoring coding vector; Input the EV value temporal association implicit feature core information anchor coding vector and the current mode unique hot embedding coding feature core information anchor coding vector into the feature granularity response interactive encoder to obtain the EV temporal feature-mode semantic feature granularity response interactive coding vector; The EV value temporal association implicit feature core information anchor coding vector and the current mode unique hot embedding coding feature core information anchor coding vector are input into the feature value granularity response interactive encoder to obtain the EV temporal feature-mode semantic feature value granularity response interactive coding vector.

8. The method for soft photosensitive detection of an adaptive environment according to claim 7, characterized in that: The EV time series feature-pattern semantic feature granularity response interaction coding vector and the EV time series feature-pattern semantic feature value granularity response interaction coding vector are subjected to feature disturbance correction to obtain a corrected EV time series feature-pattern semantic feature granularity response interaction coding vector and a corrected EV time series feature-pattern semantic feature value granularity response interaction coding vector, including: The EV temporal feature-pattern semantic feature granularity response interaction coding vector and the EV temporal feature-pattern semantic feature value granularity response interaction coding vector are respectively adjusted based on the expansibility constraints of features and eigenvalues ​​to adjust the growth exponents of features and eigenvalues ​​in the diffusion process, and based on this, the EV temporal feature-pattern semantic feature granularity response interaction coding vector and the EV temporal feature-pattern semantic feature value granularity response interaction coding vector are perturbation-corrected to obtain the corrected EV temporal feature-pattern semantic feature granularity response interaction coding vector and the corrected EV temporal feature-pattern semantic feature value granularity response interaction coding vector.

9. The method for soft photosensitive detection of an adaptive environment according to claim 8, characterized in that: Determining whether to switch modes based on the EV temporal feature-mode semantic feature comparison coding vector includes: The EV timing feature-mode semantic feature comparison encoding vector is input into a classifier-based mode switching intelligent controller to obtain a control instruction, and the control instruction is used to indicate whether to switch the mode.

Images