Method and apparatus for verifying vision fail-safe recognition algorithm

Abstract

The present invention is characterized in that, in a method for verifying a vision fail-safe recognition algorithm performed by a processor, the method comprises: a step of generating a contamination shape mask using a first artificial intelligence model; a step of generating a contamination image from an uncontaminated original image based on the contamination shape mask using a second artificial intelligence model; and a step of verifying the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image.

Description

Technology Field

[0001] The present invention relates to a method and apparatus for verifying a vision fail-safe recognition algorithm, and more particularly to a method and apparatus for verifying a vision fail-safe recognition algorithm that detects contamination of a camera lens in an autonomous driving system using a generative AI model. Background Technology

[0002] Image recognition technology using cameras plays a core role in autonomous driving systems. Accordingly, image recognition via cameras must necessarily encompass a wide variety of situations. In particular, given the nature of cameras requiring the recognition of various objects or surrounding environments, the process of determining whether the camera is in a state capable of normal operation (clean, without lens contamination) is critical. This is known as Vision Fail-Safe (VFS) technology; it is a technology that detects critical situations caused by camera lens contamination (obscuring, blurring, rainwater, etc.) and is an essential technology for ensuring the reliability of recognition performance and passenger safety.

[0003] To verify the vision fail-safe recognition algorithm of an autonomous driving image recognition camera, contaminated images are typically acquired by driving an actual vehicle on the road. Subsequently, to verify the recognition algorithm, Ground Truth (GT) is manually labeled onto the target contaminant areas in each frame of the acquired images. Finally, the performance of the recognition algorithm is verified by comparing the algorithm results with the GTs. In other words, the conventional method for verifying vision fail-safe recognition algorithms follows the process of collecting contaminated images using an actual vehicle -> manual GT labeling -> vision fail-safe recognition algorithm verification.

[0004] Various contaminant image datasets are essential for developing and evaluating vision fail-safe recognition technology. This is because these datasets must satisfy functional specifications specific to each vehicle model, such as camera mounting position, resolution, lens characteristics, and contaminant implementation. To satisfy these functional specifications with every vehicle model development, the process—from collecting image datasets using actual vehicles to GT labeling, developing vision fail-safe recognition, and evaluating vision fail-safe recognition—consumes a significant amount of time and cost.

[0005] Conventional methods utilize actual vehicles for image acquisition to verify vision fail-safe recognition algorithms, which can always be exposed to the risk of accidents, and contaminant GT data must also be manually labeled for every image frame. Therefore, conventional methods require excessive manpower and time for sourcing actual vehicles, establishing an acquisition (logging) environment, and GT labeling.

[0006] Additionally, collecting datasets utilizing actual vehicles severely limits the feasible contamination cases. Furthermore, since occlusion ratios, contamination levels, and contamination locations are determined manually, it is extremely difficult to precisely set these parameters.

[0007] Therefore, there is a need for a method and apparatus to verify a vision fail-safe recognition algorithm capable of overcoming the limitations of conventional methods. Prior art literature

[0008] Korean Registered Patent No. 10-2663123 The problem to be solved

[0009] The present invention aims to provide a method and apparatus for verifying a vision fail-safe recognition algorithm that utilizes a generative artificial intelligence model to efficiently generate diverse and precisely controllable contamination images and systematically verify the performance of the algorithm. means of solving the problem

[0010] To achieve the above objective, the present invention is characterized in that, in a method for verifying a vision fail-safe recognition algorithm performed by a processor, the method comprises: a step of generating a contamination shape mask using a first artificial intelligence model; a step of generating a contamination image from an uncontaminated original image based on the contamination shape mask using a second artificial intelligence model; and a step of verifying the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image.

[0011] Preferably, the first artificial intelligence model is a Variational Auto-Encoder (VAE), and the variant auto-encoder structure may include an encoder and a decoder based on a Convolutional Neural Network (CNN).

[0012] Preferably, the step of generating the contamination shape mask may, during training, input the original contamination shape mask and the blockage ratio into an encoder to convert them into a low-dimensional latent vector, and input the low-dimensional latent vector, random noise, and the blockage ratio into a decoder to regenerate the original contamination shape mask.

[0013] Preferably, the step of generating the contamination shape mask may, during inference, use only the decoder of the first artificial intelligence model and input a blocking ratio and random noise into the decoder to generate a contamination shape mask in which the size of the contamination region is controlled according to the blocking ratio.

[0014] Preferably, the step of generating the contamination image can be performed by inputting the original image, the noise ratio, and the generated contamination shape mask into the second artificial intelligence model to generate a contamination image in which the degree of contamination is adjusted according to the noise ratio.

[0015] Preferably, the second artificial intelligence model is a diffusion model that transforms an image through a noising and denoising process, and may include encoder and decoder structures to apply features extracted from the original image to the diffusion model.

[0016] Preferably, the step of generating the contamination image may involve inputting the original image into an encoder to extract a feature vector, progressively adding noise to the extracted feature vector according to the input noise ratio, progressively removing noise from the noise-added feature vector while maintaining the characteristics and structure of the original image by adding the original image's feature vector, adding a contamination shape mask to apply contamination only to the masked area, and inputting the denoised feature vector into a decoder to generate the contamination image.

[0017] Preferably, the verification step may perform a performance evaluation using the F1-Score by comparing the contamination recognition result of the vision fail-safe recognition algorithm for the generated contamination image with the ground truth generated based on the contamination shape mask.

[0018] Preferably, the verification step may generate a sequence of multiple contaminated images while gradually increasing the blocking rate for the same original image, and evaluate the accuracy of the contamination detection threshold of the vision fail-safe recognition algorithm for the generated sequence images.

[0019] In addition, the present invention is further characterized in that it comprises a processor including one or more cores and a memory, as a device for verifying a vision fail-safe recognition algorithm, wherein the processor generates a contamination shape mask using a first artificial intelligence model, generates a contamination image from an uncontaminated original image based on the contamination shape mask using a second artificial intelligence model, and verifies the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image. Effects of the invention

[0020] The present invention has the advantage of significantly reducing the time and cost required for verifying vision fail-safe recognition algorithms, improving the safety of the verification process, and enabling a more systematic and comprehensive algorithm evaluation by efficiently generating images with various contamination patterns and degrees through a generative AI model that precisely controls the blocking ratio and noise ratio without acquiring actual vehicle data. Brief explanation of the drawing

[0021] Figure 1 shows a flowchart of a method for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Figure 2 shows a model structure of a method for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Figure 3 shows the structure of a first artificial intelligence model during learning according to an embodiment of the present invention. Figure 4 shows the structure of a first artificial intelligence model during inference according to an embodiment of the present invention. FIG. 5 shows the structure of a second artificial intelligence model according to an embodiment of the present invention. FIG. 6 shows a flowchart of the verification step according to an embodiment of the present invention. FIG. 7 shows a configuration diagram of a device for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Specific details for implementing the invention

[0022] The present invention will be described in detail below with reference to the contents described in the attached drawings. However, the present invention is not limited or restricted by exemplary embodiments. Identical reference numerals in each drawing indicate components that perform substantially the same function.

[0023] The purpose and effects of the present invention may be naturally understood or become clearer through the following description, and the purpose and effects of the present invention are not limited solely to the description below. Furthermore, in describing the present invention, if it is determined that a detailed description of known technology related to the present invention may unnecessarily obscure the essence of the present invention, such detailed description will be omitted.

[0024] The terms used in this invention are used merely to describe specific embodiments and are not intended to limit the invention. The singular expression includes the plural expression unless the context clearly indicates otherwise. In this application, terms such as "comprising" or "having" are intended to specify the presence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the description of the invention, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0025] Terms such as "first," "second," etc., may be used to describe various components, but said components should not be limited by said terms. These terms are used solely for the purpose of distinguishing one component from another. For example, without departing from the scope of the present invention, the first component may be named the second component, and similarly, the second component may be named the first component.

[0026] Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by those skilled in the art to which this invention pertains. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant technology, and should not be interpreted in an ideal or overly formal sense unless explicitly defined in this invention.

[0027] In interpreting the components, they are interpreted to include a margin of error even without a separate explicit indication. In the case of descriptions regarding temporal relationships, for example, where the temporal sequence is described using 'after,' 'following,' 'next,' 'before,' etc., cases that are not continuous are included unless 'immediately' or 'directly' is used.

[0028] As used in this specification, the term "image" encompasses both static images, which are single, still frames, and video, which consists of multiple consecutive frames. That is, in this invention, terms such as "image conversion," "image generation," and "image processing" include not only the processing and conversion of single frames but also the processing of sequences of consecutive frames. Depending on the context, the terms "image" and "video" may be used interchangeably, but they refer to the same technical subject. However, for the sake of clear explanation of this invention, the term "image" is primarily used.

[0029] Hereinafter, the technical configuration of the present invention will be described in detail with reference to the attached drawings.

[0030] FIG. 1 shows a flowchart of a method for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Referring to FIG. 1, the method for verifying a vision fail-safe recognition algorithm may include the step of generating a contamination shape mask (S100), the step of generating a contamination image (S300), and the step of verifying contamination recognition performance (S500).

[0031] FIG. 2 illustrates the model structure of a method for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Referring to FIG. 2, the overall model structure for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention is illustrated. In the method for verifying the vision fail-safe recognition algorithm, two artificial intelligence models can be utilized. First, the first artificial intelligence model can generate a contamination shape mask by receiving a blocking ratio (0–100%) and random noise as inputs. The generated contamination shape mask and the original image are used as inputs to the second artificial intelligence model to finally generate a contamination image. The contamination image generated in this way can be utilized for verifying the vision fail-safe recognition algorithm. A detailed description of the specific structure, operating principle, and verification method of each artificial intelligence model will be provided below.

[0032] A method for verifying vision fail-safe recognition algorithms enables the efficient generation of images with various contamination patterns and degrees by precisely controlling blocking and noise ratios through generative AI models without acquiring actual vehicle data. This significantly reduces the time and cost required for verifying vision fail-safe recognition algorithms, enhances the safety of the verification process, and allows for a more systematic and comprehensive evaluation of the algorithms.

[0033] The step of generating a contamination shape mask (S100) can generate a contamination shape mask using a first artificial intelligence model.

[0034] Here, the contamination shape mask is a two-dimensional mask representing the contaminated area of ​​the camera lens, where the contaminated area is typically represented as white (or a value of 1) and the non-contaminated area as black (or a value of 0). This contamination shape mask can represent various types of contamination patterns, such as lens occlusion, blur, and rainwater, and can be used to simulate various contamination situations that may occur in a real autonomous driving environment.

[0035] The first artificial intelligence model is a Variational Auto-Encoder (VAE), and the variant auto-encoder structure may include an encoder and a decoder based on a Convolutional Neural Network (CNN).

[0036] Variant autoencoders are probabilistic generative models of data that learn the latent representations of input data and can generate new data similar to the original. Convolutional neural network-based encoders effectively extract features from input images and map them to a low-dimensional latent space, while decoders can reconstruct an image similar to the original from these latent representations.

[0037] The encoder and decoder of the first artificial intelligence model may each be composed of multiple convolution layers, pooling layers, activation functions, batch normalization layers, etc., and the specific layer configuration or parameters may be designed in various ways depending on the application environment and requirements. In addition, the loss function of the variant autoencoder is generally composed of a combination of reconstruction error and KL divergence, but the present invention may include an additional loss term to accurately reflect the blocking ratio parameter.

[0038] FIG. 3 shows the structure of a first artificial intelligence model during training according to an embodiment of the present invention. Referring to FIG. 3, the step of generating a contamination shape mask (S100) during training can input the original contamination shape mask and the blockage ratio into an encoder to convert them into a low-dimensional latent vector, and input the low-dimensional latent vector, random noise, and the blockage ratio into a decoder to regenerate the original contamination shape mask.

[0039] Specifically, the step of generating a contamination shape mask (S100) first inputs an original contamination shape mask and a blockage ratio to an encoder. Here, the original contamination shape mask is a 2D binary image representing the pattern and location of lens contamination, and the blockage ratio is a parameter that quantifies the ratio of contamination to the total area, with a value between 0% and 100%. The encoder can process the two inputs to generate a low-dimensional latent vector that compresses high-dimensional information. In this process, various layers of a convolutional neural network (CNN) structure can learn the relationship between the spatial characteristics of the mask and the blockage ratio.

[0040] Next, the step of generating a contamination shape mask (S100) can input the generated low-dimensional latent vector, random noise, and blocking ratio to the decoder. It is important that the blocking ratio is provided directly to the decoder, as this enables conditional generation to create a contamination shape that matches a specific ratio. The decoder processes these three inputs (latent vector, random noise, and blocking ratio) to regenerate a mask similar to the original contamination shape mask.

[0041] Random noise is an important factor that enhances the probabilistic characteristics of a variation autoencoder, enabling the generation of various contamination patterns. Random noise can be sampled from various probability distributions, such as a normal distribution or a uniform distribution, and the decoder can generate new contamination shapes by combining the noise, latent vectors, and blocking ratios.

[0042] The step (S100) of generating a contamination shape mask allows the first artificial intelligence model to adjust parameters in a direction that minimizes a loss function including reconstructed loss and KL divergence (Kullback-Leibler divergence) during the learning process. Reconstructed loss measures the difference between the original contamination shape mask and the reconstructed contamination shape mask, and KL divergence can induce the latent distribution generated by the encoder to approach a predefined distribution (typically a standard normal distribution).

[0043] FIG. 4 shows the structure of a first artificial intelligence model during inference according to an embodiment of the present invention. Referring to FIG. 4, the step (S100) of generating a contamination shape mask uses only the decoder of the first artificial intelligence model during inference, and inputs a blocking ratio and random noise into the decoder to generate a contamination shape mask in which the size of the contamination region is controlled according to the blocking ratio.

[0044] Specifically, the step of generating a contamination shape mask (S100) does not perform the original contamination shape mask or encoder processing during inference. Instead, the step of generating a contamination shape mask (S100) can directly input a blocking ratio (0–100%) and random noise into the decoder. The blocking ratio can have a continuous real value between 0% and 100%, enabling precise control of the contamination size. Random noise is a diversity factor that enables the generation of various forms of contamination patterns even for the same blocking ratio.

[0045] The decoder takes a blocking rate and random noise as input, processes them, and outputs a contamination shape mask with a size of contamination region corresponding to a specified blocking rate. Since the decoder has already learned the relationship between various contamination patterns and blocking rates during the training process, it can efficiently generate a contamination mask of the desired size during inference.

[0046] The step of generating a contamination shape mask (S100) can significantly reduce the computational burden by omitting the encoder process during inference. The step of generating a contamination shape mask (S100) can immediately generate contamination patterns of various sizes with various blocking ratios. This enables systematic testing of various contamination levels during the verification process of the vision fail-safe recognition algorithm and is particularly useful for evaluating the limit performance of the algorithm while gradually increasing the contamination size.

[0047] The generated contamination shape mask is utilized as an important input in the next step, the step of generating a contamination image (S300), and can ultimately contribute to comprehensively evaluating the performance of the vision fail-safe recognition algorithm under various contamination conditions.

[0048] The step of generating a contaminated image (S300) can generate a contaminated image from an uncontaminated original image based on a contaminated shape mask using a second artificial intelligence model.

[0049] The step of generating a contamination image (S300) is a core process of generating an image with lens contamination by applying a previously generated contamination shape mask to an actual image. The second AI model can generate a realistic contamination effect by not simply overlaying the mask, but by mimicking the complex visual characteristics that appear when contamination occurs on an actual camera lens. The step of generating a contamination image (S300) is a sophisticated image transformation process that applies a natural contamination effect to a designated area while maintaining the overall structure and information of the original image.

[0050] The original image may be a real road scene captured by the camera of an autonomous vehicle or a clean image generated in a simulated environment. Such original images may include various road conditions, weather conditions, lighting environments, etc., and may encompass all types of visual input that the camera of an autonomous driving system typically encounters.

[0051] The step of generating a contamination image (S300) inputs the original image, the noise ratio, and the generated contamination shape mask into a second artificial intelligence model to generate a contamination image in which the degree of contamination is adjusted according to the noise ratio.

[0052] In the step of generating a contamination image (S300), the noise ratio is an important parameter that controls the intensity or transparency of the contamination. Unlike the blocking ratio, which controls the size of the contamination, the noise ratio can determine the intensity of the contamination. The noise ratio can have a value between 0% and 100%, and a low value can indicate fine contamination (e.g., light raindrops, faint fingerprints), while a high value can indicate severe contamination (e.g., dark stains, opaque occlusion). The step of generating a contamination image (S300) can generate contamination effects of varying intensities even for the same contamination location and pattern through the noise ratio.

[0053] The second AI model can generate a realistic contamination image by processing these three inputs (original image, noise ratio, and contamination feature mask). In this case, the contamination feature mask specifies which areas to apply contamination to, while the noise ratio determines the intensity of the contamination. The second AI model can selectively apply contamination only to the areas designated by the contamination feature mask while preserving the overall context and characteristics of the original image.

[0054] The contamination effects generated by the second AI model can encompass various types of lens contamination, such as lens occlusion (opaque contamination that partially blocks the camera's field of view), blur (an effect that blurs specific areas of an image), rain effect (a phenomenon where the image is distorted by raindrops accumulating on the lens), fogging (a blurring effect caused by water vapor condensing on the lens surface), dirt or dust (a scattering effect caused by small particles attached to the lens surface), and various other types of contamination (fingerprints, scratches, stains):

[0055] The contamination image generated in the step of generating a contamination image (S300) can be differentiated from existing contamination image generation methods in that it realistically simulates various camera contamination situations that may occur in an actual driving environment and allows for precise control of the location, size, and intensity of the contamination. This controllability is an essential element for systematically and comprehensively verifying the performance of the vision fail-safe recognition algorithm.

[0056] FIG. 5 shows the structure of a second artificial intelligence model according to an embodiment of the present invention. Referring to FIG. 5, the second artificial intelligence model is a diffusion model that transforms an image through a noising and denoising process, and includes an encoder and a decoder structure to apply features extracted from the original image to the diffusion model.

[0057] The second AI model may be a structure based on a Latent Diffusion Model (LDM). Unlike conventional diffusion models that operate in pixel space, the Latent Diffusion Model first transforms the image into a low-dimensional feature space and then applies the diffusion process. This approach enables high-quality image transformation while significantly improving computational efficiency.

[0058] The core principle of the diffusion model consists of two stages: a 'forward process' that gradually adds noise to the data, and a 'reverse process' that reconstructs the original data from the noise-added data. In the second AI model, this process takes place in the feature space extracted by the encoder, rather than in the pixel space of the original image.

[0059] The step of generating a contaminated image (S300) involves inputting an original image into an encoder to extract a feature vector, progressively adding noise to the extracted feature vector according to the input noise ratio, progressively removing noise from the feature vector with added noise, while maintaining the characteristics and structure of the original image by adding a feature vector of the original image, applying the contaminant only to the masked area by adding a contaminant shape mask, and inputting the denoised feature vector into a decoder to generate a contaminated image.

[0060] Specifically, the step of generating a contaminated image (S300) can input the original image into an encoder to convert it into a feature vector. The encoder is typically composed of a Convolutional Neural Network (CNN) and can effectively extract the visual characteristics of the image and compress them into a low-dimensional representation. The feature vector contains the core visual information of the image while having a much lower dimension than the original image.

[0061] The step of generating a contaminated image (S300) can gradually add noise to the extracted feature vector according to the noise ratio. The higher the noise ratio, the more noise is added, which may result in greater loss of information from the original features. This process corresponds to the forward process of a diffusion model, and various forms of noise, such as Gaussian noise, can be used.

[0062] The step of generating a contaminated image (S300) can perform a denoising process that progressively removes noise from a feature vector with added noise. In this step, two important pieces of additional information (the feature vector of the original image and the contaminated shape mask) can be utilized. By referencing the feature vector of the original image during the denoising process, the step of generating the contaminated image (S300) can ensure that the transformed image maintains the structure and context of the original. This can be implemented through a skip connection or an attention mechanism, allowing the information of the original image to be preserved in parts outside the contaminated area. The step of generating the contaminated image (S300) can utilize the contaminated shape mask information during the denoising process to selectively apply the contamination effect only to the area designated by the mask. This is a form of conditional generation, whereby different denoising processes are applied to the masked area and the non-masked area, allowing the contamination to appear only in the designated area.

[0063] The step of generating a contaminated image (S300) can finally input the noised feature vector into a decoder to convert it into a contaminated image with the same resolution as the original image. The decoder performs the inverse process of the encoder and can expand the low-dimensional feature representation back into a high-resolution image.

[0064] The step of generating a contamination image (S300) can accurately apply contamination only to areas designated by conditional generation using a mask through the generation of a contamination image using a second artificial intelligence model, fully preserve visual information of uncontaminated areas by utilizing the features of the original image in the denoising process, finely control the intensity of contamination through a noise ratio parameter, significantly improve computational efficiency by performing a diffusion process in feature space rather than pixel space, and naturally simulate various types of lens contamination (occlusion, blur, rainwater, etc.).

[0065] The step of verifying contamination recognition performance (S500) can verify the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image.

[0066] The step of verifying contamination recognition performance (S500) is a core process for verifying the performance of the vision fail-safe recognition algorithm, which is the ultimate objective of the present invention. The contamination images generated through the first and second artificial intelligence models can be input into the vision fail-safe recognition algorithm and utilized to evaluate the algorithm's contamination detection capability. The step of verifying contamination recognition performance (S500) can comprehensively evaluate various performance indicators, such as the algorithm's robustness, accuracy, and sensitivity, rather than simply measuring the recognition rate.

[0067] FIG. 6 shows a flowchart of a verification step according to an embodiment of the present invention. Referring to FIG. 5, the step of verifying contamination recognition performance (S500) can perform a performance evaluation using an F1-Score by comparing the contamination recognition result of a vision fail-safe recognition algorithm for a generated contamination image with a ground truth generated based on a contamination shape mask.

[0068] An important feature of the step (S500) for verifying contamination recognition performance is that accurate ground truth can be automatically obtained. In conventional vehicle-based verification methods, ground truth for contamination areas had to be manually labeled, but in the present invention, the contamination shape mask used to generate the contamination image can naturally serve as ground truth. This can significantly improve the accuracy and efficiency of the verification process.

[0069] The F1-Score is the harmonic mean of Precision and Recall, serving as a metric that provides a balanced evaluation of a classification model's performance. In this context, Precision represents the ratio of actual contaminated pixels among those classified as contaminated by the algorithm, while Recall represents the ratio of actual contaminated pixels among those correctly classified as contaminated by the algorithm. The F1-Score can be calculated using the following Equation 1.

[0070] [Mathematical Formula 1]

[0071] F1 = 2 Х (Precision Х Recall) / (Precision + Recall)

[0072] Here, Precision is TP / (TP + FP), Recall is TP / (TP + FN), TP (True Positive) is a pixel correctly classified as contaminated by the algorithm, FP (False Positive) is a pixel incorrectly classified as contaminated by the algorithm, and FN (False Negative) is a pixel incorrectly classified as non-contaminated by the algorithm.

[0073] This F1-Score-based evaluation provides a more balanced performance assessment than simple accuracy measurements and can offer valid performance indicators, particularly when the proportion of contaminated areas in the overall image is small (class imbalance problem).

[0074] The step (S500) for verifying pollution recognition performance can individually evaluate the performance of the algorithm for various types of pollution such as occlusion, blur, and rainwater, measure changes in the algorithm's performance while gradually increasing the blocking ratio for the same pollution pattern, evaluate the sensitivity of the algorithm while changing the noise ratio for the same pollution location and size, analyze performance differences by location by applying pollution to various locations of the image (center, periphery, etc.), and evaluate performance differences by environment by applying pollution to the original image including various lighting conditions, weather environments, road conditions, etc.

[0075] The step (S500) for verifying contamination recognition performance can be analyzed from various angles using diverse evaluation indicators such as the ROC curve, Precision-Recall curve, and IoU (Intersection over Union), in addition to the F1-Score.

[0076] The step of verifying the contamination recognition performance (S500) can generate a sequence of multiple contamination images while gradually increasing the blocking ratio for the same original image, and evaluate the accuracy of the contamination detection threshold of the vision fail-safe recognition algorithm for the generated sequence images.

[0077] The step of verifying contamination recognition performance (S500) can evaluate the performance of the algorithm in dynamic and progressive contamination situations, going beyond static evaluation. In actual autonomous driving environments, camera lens contamination often intensifies gradually over time rather than occurring suddenly. For example, this applies to situations where raindrops gradually accumulate on the lens or dust builds up over time. To simulate these real-world situations, the step of verifying contamination recognition performance (S500) may provide a sequence-based verification method that systematically increases the degree of contamination on the same original image. The specific verification process is as follows.

[0078] First, in the sequence image generation step, a series of contamination images can be generated starting from the same original image, while maintaining the same contamination pattern and location, with only the blocking ratio gradually increasing. For example, a sequence can be constructed with blocking ratios increasing in steps, such as 0%, 10%, 20%, and 30%, and the intensity of contamination can also be gradually increased by adjusting the noise ratio as needed. Next, in the threshold evaluation step, the generated sequence images are sequentially input into a vision fail-safe recognition algorithm, and the algorithm's output (contamination detection result) can be recorded at each step. Through this, the precise threshold at which the algorithm makes a judgment of "contamination detected" or "camera unusable" can be identified, and whether this threshold is appropriate for the safe operation of an actual autonomous driving system can be evaluated.

[0079] Regarding the criteria for evaluating threshold accuracy, detecting contamination too early (oversensitivity) leads to functional limitations due to unnecessary camera interruptions, while detecting it too late (insensitivity) results in safety risks as it fails to detect contamination even in severe conditions. The ideal threshold is to accurately detect contamination at the point where the camera's recognition performance begins to substantially degrade, and threshold consistency across various environmental conditions and contamination types can also be evaluated.

[0080] FIG. 7 shows a configuration diagram of a device (100) for verifying a vision fail-safe recognition algorithm according to an embodiment of the present invention. Referring to FIG. 4, the configuration of the device (100) for verifying a vision fail-safe recognition algorithm illustrated is merely an example shown in a simplified manner. In one embodiment of the present invention, the device (100) for verifying a vision fail-safe recognition algorithm may include other configurations for performing the computing environment of the device (100), and only some of the disclosed configurations may constitute the device (100).

[0081] A device (100) for verifying a vision fail-safe recognition algorithm may include a processor (110) including one or more cores, memory (120), and a network (130).

[0082] The processor (110) may be composed of one or more cores and may include processors for data analysis and deep learning, such as a central processing unit (CPU), a general purpose graphics processing unit (GPGPU), and a tensor processing unit (TPU) of a computing device. The processor (110) may read a computer program stored in memory (120) and perform data processing for machine learning according to one embodiment of the present disclosure. According to one embodiment of the present disclosure, the processor (110) may perform operations for learning a neural network. The processor (110) may perform calculations for learning a neural network, such as processing input data for learning in deep learning (DL), extracting features from input data, calculating errors, and updating the weights of the neural network using backpropagation. At least one of the CPU, GPGPU, and TPU of the processor (110) may process the learning of a network function. For example, a CPU and a GPGPU can work together to process the learning of a network function and data classification using the network function. Additionally, in one embodiment of the present disclosure, processors of a plurality of computing devices can be used together to process the learning of a network function and data classification using the network function. Furthermore, a computer program executed on a computing device according to one embodiment of the present disclosure may be a CPU, GPGPU, or TPU executable program.

[0083] The processor (110) can generate a contamination shape mask using a first artificial intelligence model. The processor (110) can perform the step (S100) of generating the aforementioned contamination shape mask.

[0084] The processor (110) can generate a contaminated image from an uncontaminated original image based on a contaminated shape mask using a second artificial intelligence model. The processor (110) can perform the step (S300) of generating the aforementioned contaminated image.

[0085] The processor (110) can verify the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image. The processor (110) can perform the aforementioned verification step (S500).

[0086] The memory (120) can store any form of information generated or determined by the processor (110) and any form of information received by the network (130).

[0087] The memory (120) may include at least one type of storage medium among flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory, etc.), RAM (Random Access Memory, RAM), SRAM (Static Random Access Memory), ROM (Read-Only Memory, ROM), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk. The computing device (100) may operate in conjunction with web storage that performs the storage function of the memory (120) on the internet. The description of the memory described above is merely an example, and the present disclosure is not limited thereto.

[0088] The network (130) may use any type of known wired or wireless communication system. The network (130) may receive original images, etc. from a related device or related system.

[0089] The network (130) can transmit and receive information, user interfaces, etc., processed by the processor (110) through communication with other terminals. For example, the network (130) can provide a user interface generated by the processor (100) to a client (e.g., a user terminal). In addition, the network (130) can receive external input from a user authorized as a client and transmit it to the processor (110). At this time, the processor (110) can process operations such as outputting, modifying, changing, or adding information provided through the user interface based on the external input from the user received from the network (130).

[0090] Meanwhile, the device (100) for verifying the vision fail-safe recognition algorithm may include a server as a computing system that transmits and receives information through communication with a client. At this time, the client may be any type of terminal capable of accessing the server.

[0091] In an additional embodiment, the device (100) for verifying the vision fail-safe recognition algorithm may include any type of terminal that receives data resources generated from any server and performs additional information processing.

[0092] Although the present invention has been described in detail above through representative embodiments, those skilled in the art will understand that various modifications can be made to the above-described embodiments within the scope of the present invention. Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined by the claims set forth below as well as all modifications or variations derived from the claims and equivalent concepts. Explanation of the symbols

[0093] 100: Device for verifying vision fail-safe recognition algorithms 110: Processor 120: Memory 130: Network

Claims

Claim 1 A method for verifying a vision fail-safe recognition algorithm, performed by a processor, comprising: a step of generating a contamination shape mask using a first artificial intelligence model; a step of generating a contamination image from an uncontaminated original image based on the contamination shape mask using a second artificial intelligence model; and a step of verifying the contamination recognition performance of the vision fail-safe recognition algorithm using the generated contamination image. Claim 2 A method according to claim 1, wherein the first artificial intelligence model is a Variational Auto-Encoder (VAE), and the variational auto-encoder structure includes an encoder and a decoder based on a Convolutional Neural Network (CNN). Claim 3 In paragraph 2, the step of generating the contamination shape mask comprises, during training, inputting the original contamination shape mask and the blockage ratio into an encoder to convert them into a low-dimensional latent vector, and inputting the low-dimensional latent vector, random noise, and the blockage ratio into a decoder to regenerate the original contamination shape mask. Claim 4 A method according to paragraph 2, wherein the step of generating the contamination shape mask is, during inference, to use only the decoder of the first artificial intelligence model, input a blocking ratio and random noise into the decoder to generate a contamination shape mask in which the size of the contamination region is controlled according to the blocking ratio. Claim 5 A method according to claim 1, wherein the step of generating the contamination image comprises inputting an original image, a noise ratio, and a generated contamination shape mask into the second artificial intelligence model to generate a contamination image in which the degree of contamination is adjusted according to the noise ratio. Claim 6 A method according to claim 1, wherein the second artificial intelligence model is a diffusion model that transforms an image through a noising and denoising process, and includes an encoder and a decoder structure to apply features extracted from an original image to the diffusion model. Claim 7 In claim 6, the step of generating the contaminated image comprises inputting an original image into an encoder to extract a feature vector, progressively adding noise to the extracted feature vector according to the input noise ratio, progressively removing noise from the feature vector with added noise, while adding a feature vector of the original image to maintain the characteristics and structure of the original image, adding a contaminated shape mask to apply the contaminant only to the masked area, and inputting the denoised feature vector into a decoder to generate the contaminated image. Claim 8 A method according to claim 1, wherein the verification step is to compare the contamination recognition result of a vision fail-safe recognition algorithm for a generated contamination image with the ground truth generated based on the contamination shape mask, and to perform a performance evaluation using the F1-Score. Claim 9 A method according to claim 1, wherein the verification step comprises generating a sequence of multiple contaminated images while progressively increasing the blocking rate for the same original image, and evaluating the accuracy of the contamination detection threshold of a vision fail-safe recognition algorithm for the generated sequence images. Claim 10 A device for verifying a vision fail-safe recognition algorithm, comprising: a processor including one or more cores; and memory; wherein the processor performs a method for verifying a vision fail-safe recognition algorithm according to claim 1.

Images