A zero-shot anti-interference target detection method and related device

Abstract

The application belongs to the technical field of computer vision, and discloses a zero-sample anti-interference target detection method and a related device, which comprises the following steps: obtaining a to-be-detected image; inputting the to-be-detected image into a target detection model to output a target detection result of the to-be-detected image; wherein the target detection model is constructed based on an improved convolutional neural network architecture, and the improved convolutional neural network is a standard convolutional neural network integrated with a biomimetic framework BioNet; the biomimetic framework BioNet comprises a visual focus simulation module and an olfactory bulb simulation module; the biomimetic framework BioNet is integrated into the standard convolutional neural network to construct the target detection model, the dependence on prior knowledge of interference types is eliminated from the bottom layer of the model, the noise types do not need to be preset in advance, the intrinsic robustness of the detection network is given from the model architecture level, the hardware resource consumption and the system operation complexity of model training are effectively reduced, and the additional calculation overhead is reduced.

Description

Technical Field

[0001] This invention belongs to the field of computer vision technology, and relates to target detection technology, and in particular to a zero-sample anti-interference target detection method and related device. Background Technology

[0002] As one of the core technologies in the field of computer vision, object detection technology is widely used in fields such as autonomous driving, intelligent monitoring, and robot vision. Its core is to accurately identify target objects in images or videos and determine the location and category of the targets. In recent years, with the rapid development of artificial intelligence technology, object detection algorithms based on convolutional neural networks (CNN) can meet the multi-scenario detection needs in laboratory environments, promoting the large-scale application of computer vision technology.

[0003] Although CNN-based object detection algorithms perform well in standard laboratory test scenarios, their performance degrades significantly when deployed in real-world open-world scenarios. Compared to the controlled environment of a laboratory, open-world scenarios often have unpredictable environmental disturbances such as severe weather, low light, motion blur, and sensor noise, which damage the original features of the input image, leading to missed detections, false detections, and decreased accuracy of the CNN model. As a result, it cannot meet the core requirements of industrial and consumer applications for technical stability and robustness.

[0004] To mitigate the impact of environmental disturbances, mainstream anti-interference solutions generally adopt the traditional data-dependent paradigm, which mainly includes data-level augmentation and image-level restoration. Data-level augmentation simulates common disturbances during the training phase, enabling the model to memorize noise distributions to adapt to similar environments. Image restoration involves adding a preprocessing module before detection to clean up the disturbed image before inputting it into the detection model. The traditional data-dependent paradigm can mitigate the effects of specific disturbances and is currently the most widely used anti-interference path.

[0005] However, the aforementioned traditional data-dependent paradigms all rely on prior knowledge of the types of disturbances, failing to fundamentally solve the robustness problem of object detection in open-world scenarios. Specifically, the essence of the traditional data-dependent paradigm depends on prior knowledge of the types of disturbances. Whether it is the disturbances simulated by data-level augmentation or the disturbances targeted by image-level restoration, the specific types and characteristics of the disturbances need to be clearly defined in advance. On the one hand, reliance on prior knowledge increases additional computational overhead. Specifically, data-level augmentation requires generating a large amount of training data with disturbances, extending model training time and consuming hardware resources; image-level restoration requires additional deployment of preprocessing modules, increasing system complexity and operating costs. On the other hand, a more critical problem is insufficient generalization ability. When the model encounters novel disturbances (i.e., disturbance types not included in prior knowledge) that have not been seen during training, domain shift occurs, causing the detection model to be unable to adapt to the new disturbance environment, and the anti-interference performance drops sharply. Summary of the Invention

[0006] To address the technical problems existing in the prior art, this invention provides a zero-sample anti-interference target detection method and related apparatus to solve the technical problem that traditional data-dependent paradigms rely on prior knowledge of interference types and fail to fundamentally solve the robustness problem of target detection in open-world scenarios.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention provides a zero-shot, interference-resistant target detection method, comprising: Acquire the image to be detected; The image to be detected is input into the target detection model, and the target detection result of the image to be detected is output. Among them, the object detection model is built based on the architecture of an improved convolutional neural network, which is a standard convolutional neural network that integrates the bio-inspired framework BioNet. The biomimetic framework BioNet includes a visual focusing simulation module and an olfactory ball simulation module; The visual focusing simulation module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian nuclei of different scales to filter additive noise; the olfactory bulb simulation module is used to reconstruct context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms.

[0008] Furthermore, the visual focusing simulation module is embedded in the feature extraction backbone of the standard convolutional neural network and placed between two adjacent convolutional units in the feature extraction backbone. The snorting ball simulation module is positioned between the feature fusion network of a standard convolutional neural network and the detection head.

[0009] Furthermore, the visual focusing simulation module includes a center-periphery layer unit, a differential response calculation unit, a channel attention unit, and a residual fusion unit; The center-periphery layer unit is used to perform convolution operations on the input feature map of the visual focusing simulation module using two Gaussian kernels of different scales to obtain the center signal and the peripheral signal; The differential response calculation unit is used to calculate the differential signals of the central signal and the peripheral signal respectively, and sum them after nonlinear activation to simulate the responses of ON-type neurons and OFF-type neurons in the retina and obtain the retinal response output. The channel attention unit is used to perform attention weighting on the retinal response output based on pre-calculated channel weighting coefficients to obtain the weighted response; The residual fusion unit is used to multiply the weighted response by a predetermined scaling factor and then add it to the input feature map of the visual focusing simulation module to obtain the output of the visual focusing simulation module.

[0010] Furthermore, the center-periphery layer unit includes parallel center convolutional pathways and peripheral convolutional pathways; wherein the Gaussian kernel scale used in the center convolutional pathways and peripheral convolutional pathways is different.

[0011] Furthermore, the olfactory bulb simulation module includes an information aggregation unit, a dual-pathway splitting unit, a tufted cell pathway unit, a monk's cap cell pathway unit, and a feature fusion and residual output unit. The information aggregation unit is used to aggregate the initial signal of the input feature map of the sniffing ball simulation module to obtain the initial signal aggregated feature map; The dual-path splitting unit is used to divide the initial signal aggregation feature map into two independent subsets along the channel dimension to obtain the clustered path input feature map and the monk's cap path input feature map. The plexiform cell pathway unit is used to capture local high-frequency features using a pre-defined restricted receptive field. Specifically, a pre-determined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of capturing local high-frequency features, so as to obtain the features after plexiform pathway inhibition. The capillary cell pathway unit is used to encode global spatial context using a pre-defined broad receptive field. Specifically, a pre-determined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of encoding global spatial context, so as to obtain the features after capillary pathway inhibition. The feature fusion and residual output unit is used to concatenate the features after the clustered pathway is suppressed with the features after the monk's cap pathway is suppressed. After convolution mapping, the concatenation is added to the input feature map of the olfactory ball simulation module to obtain the output of the olfactory ball simulation module.

[0012] Furthermore, the predetermined lateral inhibition mechanisms include local lateral inhibition mechanisms and global feedback inhibition mechanisms; The local side suppression mechanism is used to calculate the average activation in the local spatial neighborhood of any intermediate feature map as the suppression signal; the suppression signal is applied to any intermediate feature map to obtain the locally suppressed features; A global feedback suppression mechanism is used to perform global average pooling on the features after local suppression and then pass them through a fully connected layer to obtain a global suppression signal. The global suppression signal is then applied to the features after local suppression to obtain features after clustered pathway suppression or monk's cap pathway suppression.

[0013] The present invention also provides a zero-shot anti-interference target detection system for implementing the aforementioned zero-shot anti-interference target detection method, comprising: The image acquisition module is used to acquire the image to be detected; The target detection module is used to input the image to be detected into the target detection model and output the target detection result of the image to be detected. Among them, the object detection model is built based on the architecture of an improved convolutional neural network, which is a standard convolutional neural network that integrates the bio-inspired framework BioNet. The biomimetic framework BioNet includes a visual focusing simulation module and an olfactory ball simulation module; The visual focusing simulation module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian nuclei of different scales to filter additive noise; the olfactory bulb simulation module is used to reconstruct context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms.

[0014] The present invention also provides an electronic device, comprising: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the zero-sample anti-interference target detection method.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the zero-sample anti-interference target detection method.

[0016] The present invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the zero-sample anti-interference target detection method.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: The zero-shot anti-interference target detection method provided by this invention integrates the bio-inspired framework BioNet into a standard convolutional neural network to construct a target detection model. Compared with traditional anti-interference schemes that rely on prior knowledge of interference, it eliminates the dependence on prior knowledge of interference types from the model's underlying layer. It does not require pre-setting noise types, simulating feature interference samples, or adding an image restoration preprocessing module before detection. It endows the detection network with inherent robustness at the model architecture level, effectively reducing the hardware resource consumption and system operation complexity of model training, and reducing additional computational overhead. Specifically, the visual focusing simulation module achieves low-frequency noise filtering at the physical level through dual Gaussian kernels, while the snorting ball simulation module completes background interference suppression in the semantic space based on lateral inhibition. The visual focusing simulation module and the snorting ball simulation module work together to significantly reduce the computational overhead and system complexity caused by data generation and preprocessing, and effectively avoid the insufficient generalization ability caused by domain offset in traditional methods. It maintains stable detection accuracy under unknown and harsh environmental disturbances, while not losing detection performance on clean images. It can be plug-and-play adapted to various convolutional neural network detection architectures and meets the actual needs of target detection stability and versatility in open-world scenarios.

[0018] Furthermore, embedding the visual focusing simulation module between adjacent convolutional units of the feature extraction backbone network can filter out physical noise and low-frequency interference in the image at the early stage of feature extraction, effectively preventing interference features from being amplified and propagated layer by layer as the network deepens. At the same time, it does not change the original feature extraction logic of the backbone network, ensuring the complete and effective extraction of basic visual features. Setting the snorting ball simulation module between the feature fusion network of the standard convolutional neural network and the detection head can optimize the semantic features after high-level fusion, accurately suppress background interference and false activation in the semantic space, making the features of the final input detection head purer and the target features more prominent. This achieves hierarchical collaborative processing of physical noise filtering and semantic interference suppression, maximizing the overall anti-interference performance of the model in open-world scenarios.

[0019] The zero-sample anti-interference target detection system, electronic device, computer-readable storage medium, and computer program product provided by this invention possess all the advantages of the aforementioned zero-sample anti-interference target detection method. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1A flowchart of the zero-sample anti-interference target detection method provided in Example 1; Figure 2 This is a schematic diagram illustrating the principle of the traditional data-dependent paradigm. Figure 3 This is a schematic diagram illustrating the principle of the biomimetic framework BioNet in Example 1; Figure 4 This is a structural block diagram of the visual focusing simulation module in Example 1; Figure 5 This is a structural block diagram of the olfactory bulb simulation module in Example 1; Figure 6 This is a structural diagram of the improved convolutional neural network in Example 1; Figure 7 A qualitative comparison of object detection results of different methods on the VOC dataset; Figure 8 A qualitative comparison chart of object detection results of different methods on real-world datasets; Figure 9 A visual comparison of different methods on the feature attention heatmap; Figure 10 This is a block diagram of the zero-sample anti-interference target detection system provided in Example 2; Figure 11 This is a structural block diagram of the electronic device provided in Example 3. Detailed Implementation

[0022] To make the technical problems, technical solutions, and beneficial effects solved by this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0023] Before describing the specific embodiments of this application, some of the technical terms involved in the embodiments of this application are explained as follows: Convolutional Neural Networks (CNNs) are deep learning models designed specifically for data with a grid structure (such as images, speech, and time series). They automatically extract hierarchical spatial features through local connectivity, weight sharing, and pooling, significantly reducing the number of parameters and improving robustness to translation and scaling. They are the mainstream foundational models for computer vision.

[0024] Domain shift refers to the difference between the sample domain where the training data is located and the sample domain where the actual test data is located.

[0025] The center-surround antagonism mechanism of the retina involves the receptive fields of retinal ganglion cells, which consist of a central (excitatory) area and a peripheral (inhibitory) area, and the two areas produce opposite neural responses to light stimulation.

[0026] Zero-sum balance, in game theory, is when the sum of the gains and losses of the participants is zero.

[0027] Zero-shot refers to unseen interference scenarios.

[0028] Lateral inhibition is a mechanism in which adjacent neurons or receptors in the nervous system enhance signal contrast through inhibitory interactions.

[0029] Winner-Take-All occurs in a competitive system where a dominant individual controls all or most of the resources, resulting in minimal gains for other participants.

[0030] Scale-space theory analyzes the multi-scale features of images by continuously varying scale parameters, simulating the human visual system's hierarchical understanding of objects.

[0031] Physical common-mode rejection is the ability of a differential amplifier circuit to suppress common-mode signals (interference with the same amplitude and phase), thereby improving the signal-to-noise ratio.

[0032] This invention provides a zero-shot, interference-resistant target detection method, comprising the following steps: Step 100: Obtain the image to be detected.

[0033] Step 200: Input the image to be detected into the target detection model and output the target detection result of the image to be detected; wherein, the target detection model is constructed based on the architecture of an improved convolutional neural network, and the improved convolutional neural network is a standard convolutional neural network integrating the bio-inspired framework BioNet.

[0034] It should be noted that the biomimetic framework BioNet includes a visual focusing simulation module and an olfactory bulb simulation module. The visual focusing simulation module is used to simulate the center-peripheral antagonism mechanism of the retina by using dual Gaussian kernels of different scales to filter additive noise. The olfactory bulb simulation module is used to reconstruct context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms.

[0035] In the above implementation, the biomimetic framework BioNet is integrated into a standard convolutional neural network to construct an object detection model. By relying on the biomimetic framework BioNet, two biological perception mechanisms, retinal center-peripheral antagonism and olfactory bulb side inhibition, are incorporated into the object detection model. The model can be endowed with inherent robustness by training on clean data. It does not need to rely on interference prior knowledge, simulate damaged samples, or add image restoration preprocessing steps, thus fundamentally avoiding the problems of high computational cost and high system complexity of traditional methods.

[0036] Specifically, the visual focusing simulation module simulates the center-periphery antagonism mechanism of the retina, employing a dual-scale Gaussian kernel to construct a zero-sum balanced filtering mechanism. This proactively filters out physical interferences such as additive noise and contrast attenuation before feature encoding, while retaining high-frequency key features such as target edges and textures, achieving intrinsic denoising of low-level features. The olfactory bulb simulation module replicates the olfactory bulb neural side-inhibition circuit, using dual-path parallel processing and a two-stage inhibition mechanism to complete feature decorrelation and background interference suppression in the high-level semantic space. This forces feature sparsity and highlights salient targets, effectively solving semantic ambiguity problems in complex scenes. The target detection model only needs to be trained on a clean dataset to endow the network with inherent robustness, enabling it to handle unknown adverse weather conditions and low light levels. The invention features zero-sample generalization capabilities free from motion blur and other disturbances, avoiding performance drops caused by domain shifts. Robustness is significantly improved in scenarios with more severe interference. Furthermore, both the visual focusing simulation module and the olfactory ball simulation module adopt a plug-and-play design, seamlessly adapting to mainstream object detection architectures. This provides stable enhancement of the model's anti-interference capabilities without sacrificing or even slightly improving the detection accuracy of clean data, exhibiting excellent compatibility and scalability. Based on biological perception mechanisms, this invention constructs a robust detection framework, providing a novel technical path for object detection in open-world scenarios. It meets the industrial-grade requirements for detection stability and reliability in practical applications such as autonomous driving, intelligent monitoring, and robot vision, demonstrating high practical value and promising prospects for widespread adoption.

[0037] In this invention, the visual focusing simulation module achieves zero-sum balance through dual Gaussian kernels, which can efficiently filter out low-frequency spatial noise and additive interference; the snorting ball simulation module achieves a winner-takes-all effect through dual-path dual-stage suppression, which can suppress background redundancy interference and reconstruct robust feature representations in the semantic space; the visual focusing simulation module and the snorting ball simulation module work together to ensure that the model maintains stable detection performance under unknown and severe perturbations, effectively overcoming the insufficient generalization ability caused by domain shift; without sacrificing the detection accuracy of clean data, it significantly improves the robustness and applicability of target detection in open-world scenarios, and the module is plug-and-play with good adaptability and scalability.

[0038] The following specific embodiments further explain the zero-shot anti-interference target detection method provided by the present invention: Example 1 As attached Figure 1 As shown, this embodiment 1 provides a zero-shot anti-interference target detection method, including the following steps: Step 1: Obtain the image to be detected. The image to be detected is an image from an open-world scene or a standard test scene in a laboratory; specifically, images from scenarios such as autonomous driving, intelligent monitoring, robot vision, security patrol, or medical image analysis.

[0039] Step 2: Input the image to be detected into the object detection model, and output the object detection result of the image to be detected. The object detection model is built based on an improved convolutional neural network architecture, which is a standard convolutional neural network integrating the bio-inspired framework BioNet.

[0040] In this embodiment 1, the biomimetic framework BioNet includes a Visual Focus Simulation (VFS) module and an Olfactory Bulb Simulation (OBS) module. The VFS module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian nuclei of different scales to filter additive noise. The OBS module is used to reconstruct the context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and using parallel signal processing and lateral inhibition mechanisms.

[0041] As attached Figure 2-3 As shown, attached Figure 2 The document provides a schematic diagram illustrating the principles of the traditional data dependency paradigm, with appendices. Figure 3 The attached diagram illustrates the principle of the bionic framework BioNet in Example 1; from the appendix... Figure 2 As can be seen, traditional data-dependent paradigms, including data augmentation and image-level restoration, typically face problems such as high training costs, overfitting to specific noise priors, or poor generalization ability when dealing with unexposed corruptions; from the appendix Figure 3 As can be seen from this, the biomimetic framework BioNet in this embodiment 1, by integrating anti-interference modules inspired by vision and smell, can endow the object detection network with zero-sample robustness against various unknown perturbations without training with damaged images.

[0042] As attached Figure 4As shown, the VFS module includes a Center-Surround Layer unit, a differential response calculation unit, a channel attention unit, and a residual fusion unit. The Center-Surround Layer unit performs convolution operations on the input feature map of the visual focusing simulation module using two Gaussian kernels of different scales to obtain the center signal and the peripheral signal. The differential response calculation unit calculates the difference between the center signal and the peripheral signal, and sums them after nonlinear activation to simulate the responses of ON-type and OFF-type neurons in the retina, obtaining the retinal response output. The channel attention unit performs attention weighting operations on the retinal response output according to pre-calculated channel weighting coefficients to obtain the weighted response. The residual fusion unit multiplies the weighted response by a pre-determined scaling factor and adds it to the input feature map of the visual focusing simulation module to obtain the output of the visual focusing simulation module.

[0043] The center-periphery layer unit, as the basic computational unit of the module, aims to simulate the receptive field mechanism of the retina. Specifically, the center-periphery layer unit includes parallel central convolutional pathways and peripheral convolutional pathways. The central convolutional pathway simulates the narrow-band excitation center of the retina, while the peripheral convolutional pathway simulates the broadband inhibitory periphery of the retina. The central and peripheral convolutional pathways use different Gaussian kernel sizes. Preferably, the Gaussian kernel size used in the central convolutional pathway is 3×3, and the Gaussian kernel size used in the peripheral convolutional pathway is 5×5. It should be noted that the center-periphery layer unit employs parallel central and peripheral convolutional pathways with different Gaussian kernel sizes. The central convolutional pathway focuses on capturing high-frequency key features of target edges and textures, while the peripheral convolutional pathway is used to accurately estimate local background intensity. The two pathways compute in parallel without interference, stably realizing the biological visual mechanism of retinal center-periphery antagonism, accurately distinguishing effective target signals from uniform noise, and directly filtering interference in the spatial domain. The structure is simple, computationally inexpensive, and can seamlessly adapt to the convolutional unit structures of various backbone networks, exhibiting good compatibility and practicality.

[0044] In the middle-periphery layer unit, convolution weights are initialized using two Gaussian kernels of different scales, and convolution operations are performed on the input feature map of the visual focusing simulation module to obtain the center signal and peripheral signals; wherein, it is assumed that the input feature map of the visual focusing simulation module is the input feature map. The definitions of the center signal and the surrounding signals at this time are as follows:

[0045]

[0046] in, As the central signal; Input feature map; This is a depthwise convolution operation; The Gaussian kernel used in the center convolutional path; The kernel size is the Gaussian kernel used in the central convolutional path; The standard deviation of the Gaussian kernel used in the central convolutional path; For surrounding signals; The Gaussian kernel used in the peripheral convolutional pathway; The kernel size is the Gaussian kernel used in the peripheral convolutional path; is the standard deviation of the Gaussian kernel used in the peripheral convolutional path.

[0047] For the Gaussian kernel used in the central convolutional path Gaussian kernels used in convolutional pathways and surrounding convolutional paths It satisfies the following spatial constraints: and This ensures that the central pathway captures fine-grained local details, while the peripheral pathway estimates the regional background intensity.

[0048] In the differential response calculation unit, by analyzing the center signal... and surrounding signals Perform element-by-element subtraction operations to calculate the center signal. With surrounding signals The difference is used to obtain differential features; these features are then summed after ReLU activation to simulate the responses of ON-type and OFF-type neurons in the retina, yielding the retinal response output. Mathematically, the retinal response output functions as a learnable Difference of Gaussians operator to effectively handle uniform background disturbances (e.g., ...). The retinal response suppresses the region of retinal activity while strongly activating the edge and texture regions; specifically, the mathematical expression of the retinal response output is as follows:

[0049]

[0050]

[0051] in, For retinal response output; For the center signal With surrounding signals The difference and the result after ReLU activation; For surrounding signals With center signal The difference and the result after ReLU activation.

[0052] In the channel attention unit, the retinal response is output. The formal mapping function is then used; next, channel-dimensional weight coefficients are learned through a multilayer perceptron with global average pooling (GAP) and fully connected layers; subsequently, in order to extract effective information through inter-channel dependencies, the learned channel-dimensional weight coefficients are used to evaluate the retinal response output. The mapping function is subjected to attention weighting to obtain a weighted response. Then, in the residual fusion unit, the weighted response is multiplied by a predetermined scaling factor and added to the input feature map of the visual focusing simulation module to obtain the output of the visual focusing simulation module. Specifically, when the predetermined scaling factor is 2, the output of the visual focusing simulation module is as follows:

[0053] in, This is the output of the visual focusing simulation module; Attention-weighted operations; For retinal response output The mapping function.

[0054] In this embodiment 1, the visual focusing simulation module adopts a combined structure of a center-periphery layer unit, a differential response calculation unit, a channel attention unit, and a residual fusion unit. This not only enhances spatial saliency but also adaptively recalibrates the importance of different feature channels to maximize signal fidelity. Specifically, the center-periphery layer unit can effectively separate high-frequency details of the target from low-frequency background information; the differential response calculation unit can simulate the response of biological visual neurons to achieve noise suppression and edge enhancement; the channel attention unit can adaptively weight and strengthen effective feature channels while weakening noisy channels; and the residual fusion unit can retain original feature information and avoid gradient vanishing. The overall structure of the visual focusing simulation module can learn to implement Gaussian difference filtering, achieving efficient zero-shot denoising without additional image preprocessing. This significantly improves anti-interference capability while maintaining feature fidelity, ensuring the detection accuracy of the model under both clean and interfering data.

[0055] It should also be noted that in the visual focusing simulation module, the difference of Gaussian operator is used to achieve zero-sum balance. That is, in the uniform noise region, the central excitation and the peripheral inhibition cancel each other out, suppressing low-frequency background interference; in the high-frequency region such as the target boundary, the difference is activated to retain the key signal; secondly, the importance of different feature channels is adaptively recalibrated using the channel attention unit to further highlight the effective information related to the target and suppress redundant noise channels; in addition, there is no need to train for specific noise types, and additive noise and contrast attenuation are eliminated from the source through a biomimetic physical filtering mechanism.

[0056] As attached Figure 5As shown, the OBS module includes an information aggregation unit, a dual-pathway splitting unit, a clustered cell pathway unit, a monk's cap cell pathway unit, and a feature fusion and residual output unit.

[0057] The information aggregation unit is used to aggregate the initial signal of the input feature map of the snorting ball simulation module to obtain the initial signal aggregated feature map; the dual-path splitting unit is used to divide the initial signal aggregated feature map into two independent subsets along the channel dimension to obtain the clustered path input feature map and the monk's cap path input feature map.

[0058] The clustered cell pathway unit is used to capture local high-frequency features using a predefined restricted receptive field. Specifically, a predefined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of capturing local high-frequency features to obtain features after clustered pathway inhibition. Preferably, the predefined restricted receptive field is a 3×3 depthwise convolution.

[0059] The monk-hat cell pathway unit is used to encode global spatial context using a predefined broad receptive field. Specifically, a predefined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of encoding global spatial context to obtain the suppressed features of the monk-hat pathway. Preferably, the predefined broad receptive field is a 5×5 depthwise convolution.

[0060] The feature fusion and residual output unit is used to concatenate the features after the clustered pathway is suppressed with the features after the monk's cap pathway is suppressed. After convolution mapping, the concatenation is added to the input feature map of the olfactory ball simulation module to obtain the output of the olfactory ball simulation module.

[0061] In this embodiment 1, a residual learning paradigm is adopted throughout the OBS module to ensure effective gradient backpropagation; the working principle of the sniffing ball simulation module includes: Input feature map of OBS module After aggregating the initial signals, we obtain the aggregated feature map of the initial signals. Then, the initial signal is aggregated into a feature map. The pathway is divided into two independent subsets along the channel dimension. Then, within the plexiform pathway unit, a pre-defined restricted receptive field is used to focus on capturing detailed local high-frequency features in one subset, thereby obtaining the features after plexiform pathway inhibition. Simultaneously, within the capitol cell pathway unit, a pre-defined broad receptive field is utilized to integrate extensive spatial context and encode global dependencies, thereby obtaining features resulting from capitol pathway inhibition. Next, the characteristics of plexiform pathway inhibition. Features after inhibition of the cap pathway The features are concatenated and fused using a 1×1 convolutional layer; then, they are added to the input feature map of the snoring ball simulation module to obtain the output of the snoring ball simulation module; specifically, the mathematical expression for the output of the snoring ball simulation module is as follows:

[0062] in, This is the output of the snorkeling bulb simulation module; This is the input feature map for the OBS module; This is a characteristic of plexiform pathway inhibition; This is a characteristic of the monastic cap pathway after inhibition.

[0063] It should be noted that when visual acuity is impaired, organisms usually rely on alternative sensory modalities (such as olfaction) for compensation; the OBS module is designed to simulate the complex neural interaction network of the olfactory bulb, especially the parallel processing and lateral inhibition mechanisms of mitral cells and tufted cells. Its main goal is to reconstruct context-aware feature relationships, thereby achieving robust semantic discrimination.

[0064] In this embodiment 1, to ensure robust feature representation under complex interference, a predetermined dual-stage inhibition mechanism is introduced into the parallel processing pathways (plexiform cell pathway unit and monk's cap cell pathway unit) of the OBS module. This mechanism is applied to arbitrary intermediate feature maps of the plexiform cell pathway unit and the monk's cap cell pathway unit to obtain features after plexiform pathway inhibition. Features after inhibition of the cap pathway Among them, the predetermined lateral inhibition mechanisms include local lateral inhibition mechanisms and global feedback inhibition mechanisms; the local lateral inhibition mechanism is used to simulate the function of periglobinocytes to achieve local lateral inhibition; the global feedback inhibition mechanism is used to simulate the function of granulocytes to introduce a remote global feedback inhibition mechanism.

[0065] In the local lateral inhibition mechanism, the function of periglobulin cells (PGCs) is mimicked to achieve local lateral inhibition, aiming to enhance local contrast and suppress high-frequency noise artifacts caused by environmental perturbations; specifically, for any intermediate feature map in the pathway... The average activation within its local spatial neighborhood is calculated as the suppression signal. Then, the suppression signal will be... Applicable to any intermediate feature map The features after local suppression are obtained; the formulaic representation of the features after local suppression is as follows:

[0066] in, These are characteristics after local inhibition; For any intermediate feature map; These are learnable parameters initialized to 0, enabling the network to adaptively adjust the inhibition strength on the training device. To suppress the signal.

[0067] In the global feedback inhibition mechanism, the function of granular cells (GCs) is simulated to introduce a remote global feedback inhibition mechanism; specifically, global context information is used to suppress spurious activations that are inconsistent with the overall semantic context of the current scene; that is, for features that have undergone local suppression... Global average pooling is performed followed by two fully connected layers to model the remote dependency and obtain a global suppression signal. Then, the global suppression signal will be... Features applied after local suppression Global suppression signal After subtraction, SiLU activation is performed to decouple the channel dependence and obtain the characteristics after suppression of the clustered pathway or the monk's cap pathway.

[0068] Global suppression signal The formalized representation is as follows:

[0069] in, This is a global suppression signal; A linear layer that restores the channel dimension to its original state; To reduce the channel dimension to half that of a linear layer.

[0070] The formulaic representation of the characteristics after inhibition of the plexus pathway or the monk's cap pathway is as follows:

[0071] in, These are characteristics of the plexiform pathway after inhibition or the capillary pathway after inhibition. This is a learnable scaling factor.

[0072] In this embodiment 1, the olfactory bulb simulation module works collaboratively through an information aggregation unit, a dual-pathway splitting unit, a clustered cell pathway unit, a monk's cap cell pathway unit, and a feature fusion and residual output unit. It first completes the initial signal aggregation of the input features, and then splits the local details and global context along the channel dimension into two processing pathways. Features are extracted through the adapted receptive field and suppression processing is applied. Finally, the features are fused and output in a residual manner. It can simultaneously take into account the capture of local high-frequency details and the encoding of global spatial context, effectively solve the semantic ambiguity and false activation problems in high-level features, force the sparsity of feature representation, highlight the real target features, and greatly improve the feature discrimination ability and detection stability in complex perturbation scenarios.

[0073] In the olfactory bulb simulation module, a two-stage inhibition mechanism is formed by combining local side inhibition and global feedback inhibition. The local side inhibition mechanism simulates the function of periglobulin cells, sharpening the local topological structure of features and eliminating spatially related interference artifacts. The global feedback inhibition mechanism simulates the function of granule cells, introducing a remote global feedback inhibition mechanism to suppress spurious activations that are inconsistent with the overall semantic context of the current scene using global context information. The local side inhibition mechanism and the global feedback inhibition mechanism work together to form a winner-takes-all feature selection effect, forcing the model to focus on real target features and weakening background redundancy interference. This can effectively cope with unpredictable complex semantic perturbations in open scenes, further enhance the model's generalization and adaptability to unknown interference, and improve the accuracy and reliability of target detection results.

[0074] It is worth noting that the olfactory ball simulation module achieves a winner-take-all effect through lateral inhibition, which makes feature channels with similar structures inhibit each other, forces the feature representation to be sparsified, highlights salient targets, and suppresses the activation of diffuse background noise. The dual-path design takes into account both local details and global semantics, avoids the information loss caused by a single receptive field, and improves the target discrimination ability in complex scenes. Local inhibition eliminates spatial correlation, and global inhibition decouples channel correlation, maximizes the signal-to-noise ratio, and enhances the model's ability to generalize to unknown semantic perturbations.

[0075] As attached Figure 6 As shown, the improved convolutional neural network includes a feature extraction backbone network, a feature fusion network, and a detection head connected in sequence. The feature extraction backbone network is used to perform multi-level convolution and downsampling operations to generate feature maps of different scales. The feature extraction backbone network includes several levels of convolutional units. The feature fusion network is used to perform cross-scale feature aggregation on feature maps of different scales to form multi-scale feature maps. The feature fusion network adopts a feature pyramid structure, which fully integrates the precise location information of shallow feature maps and the high-level semantic information of deep feature maps through lateral connection paths, bottom-up paths, and bottom-up path operations. The detection head is used to complete target classification and localization.

[0076] The visual focusing simulation module is embedded in the feature extraction backbone of the standard convolutional neural network and positioned between two adjacent convolutional units. Features enter the visual focusing simulation module, where physical noise is filtered out using dual-scale Gaussian difference filtering, and after channel attention enhancement, the denoised features are output. Multiple snorting ball simulation modules are positioned between the feature fusion network of the standard convolutional neural network and the detection head. The multi-scale feature maps formed in the feature fusion network enter the snorting ball simulation modules, where they undergo dual-pathway, dual-stage suppression to suppress semantic background interference and reconstruct robust feature representations. The features output from the snorting ball simulation modules enter the detection head to complete target classification and localization.

[0077] Principle Analysis and Explanation: The zero-shot anti-interference target detection method described in Embodiment 1 integrates the biomimetic framework BioNet into a standard convolutional neural network to construct a target detection model. Inspired by the collaborative interaction between visual perception and olfactory cognition, it incorporates two biological perception mechanisms—retinal center-periphery antagonism and olfactory bulb lateral inhibition—into the target detection model based on the biomimetic framework BioNet to address signal interference at different levels, thereby achieving zero-shot anti-interference target detection in complex dynamic environments. Specifically, to combat the noise interference and contrast attenuation commonly found in the original image, a visual focusing simulation module simulating the retinal center-periphery antagonism mechanism is embedded into the feature extraction backbone of the standard convolutional neural network and placed between two adjacent convolutional units in the feature extraction backbone. To overcome semantic ambiguity in high-level features, an olfactory bulb simulation module is introduced between the feature fusion network and the detection head. By utilizing the complex neural circuits of the olfactory bulb, especially its unique parallel signal processing and lateral inhibition mechanisms, the olfactory bulb simulation module can effectively reconstruct contextual information in the high-level semantic space.

[0078] In this embodiment 1, the biomimetic framework BioNet is a physics-based denoising paradigm inspired by the inherent topological robustness of biological sensing systems. By combining the retinal preprocessing mechanism with the network architecture of the olfactory bulb, a general framework independent of noise priors for specific datasets is formed.

[0079] To mitigate the sensitivity of standard CNNs to low-frequency environmental degradation, it is necessary to rigorously decouple high-frequency structural features from spatially correlated background noise. Theoretically, a bandpass filtering mechanism approximating the Laplacian operator is required to suppress local DC components. Inspired by the center-periphery antagonism of retinal ganglion cells, a visual focusing simulation module is proposed as a zero-shot active denoiser, essentially eliminating uniform environmental interference before latent feature encoding. Specifically, in the visual focusing simulation module, physical instantiation is performed using scale-space theory, employing a dual-scale Gaussian kernel to calculate the differential response between the narrow-band excitation center and the broadband suppression periphery, achieving a numerical zero-sum balance. The dual-scale Gaussian kernel, as the only smoothing operator capable of preventing artificial structural artifacts, captures high-frequency edge details by calculating the narrow-band excitation center and estimates local background intensity by calculating the broadband suppression periphery. In the presence of uniform noise input, the central excitation and peripheral suppression cancel each other out, effectively filtering additive noise through physical common-mode suppression. Conversely, high-frequency target boundaries disrupt the balance, thus preserving key signal information.

[0080] While visual focusing simulation modules can mitigate physical noise, handling non-stationary semantic interference requires implementing higher-order signal cleansing to force sparsity in the representation. Theoretical neuroscience suggests that perceptual systems maximize information capacity through redundancy reduction achieved by lateral inhibition. Based on this principle, an olfactory ball simulation module is proposed to force CNN networks to converge to a winner-takes-all solution, where structurally similar feature channels mutually inhibit each other, thereby maximizing the signal-to-noise ratio through physical feature decorrelation. Specifically, mathematically, the basis for implementing lateral inhibition lies in solving the feature covariance problem, where background noise typically induces correlated, diffuse activations in channels and space, while salient targets are inherently sparse. To rigorously separate the background noise components, the olfactory ball simulation module instantiates a subtractive feedback mechanism mediated by two different dynamics: a local inhibition mechanism and a global feedback inhibition mechanism. Theoretically, the local inhibition mechanism is used to simulate periglobulin cells, acting as a spatial Laplacian operator to sharpen the topology by suppressing diffuse spatial correlations; the global feedback inhibition mechanism is used to simulate granular cells, acting as a Gram-Schmidt-like orthogonalization process to decouple channel dependencies.

[0081] Experimental verification and result analysis: In this embodiment 1, the performance of the bionic framework BioNet (denoted as VO) is verified by integrating it into different standard convolutional neural network architectures and conducting multi-dimensional evaluations covering synthetic impairments, diverse real-world harsh conditions, and large-scale scalability tests, as follows: (1) To assess the robustness of synthetic damage, the theoretical robustness was evaluated by validation on the PASCAL VOC dataset under synthetic weather damage conditions. Specifically, the bionic framework BioNet was integrated into the architectures of four different standard convolutional neural networks, and the four different standard convolutional neural networks were used as baseline models. The quantitative results of the baseline models and the models integrated with the bionic framework BioNet (denoted as VO-*) were compared under clean and damaged data settings, as shown in Table 1.

[0082] Table 1. Quantization results comparing the baseline model and the model integrated with the biomimetic framework BioNet in clean and damaged settings.

[0083] As shown in Table 1, the table presents the quantitative results comparing the baseline model and the model integrated with the biomimetic framework BioNet under clean and damaged settings, including performance metrics for five corrosion types at severity levels S1, S3, and S5. Table 1 shows that integrating the biomimetic framework BioNet into four different standard convolutional neural network architectures achieved continuous performance improvements under varying environmental conditions, without compromising or even enhancing performance on clean data. Analysis revealed a positive correlation between damage severity and performance improvement. Specifically, when the severity increased from S1 to 5, the performance of the baseline model dropped sharply, while the model integrated with the biomimetic framework BioNet exhibited significantly higher stability. For example, under foggy conditions, the improvement of VO-YOLOv11 relative to the baseline increased from 3.1% at severity level S1 to 5.7% at severity level S5.

[0084] (2) The generalization ability of the real-world scene was evaluated on the RTTS dataset, ExDark dataset and DAWN dataset. Specifically, the bionic framework BioNet was integrated into the architecture of three different standard convolutional neural networks and three different standard convolutional neural networks were used as the baseline model. In order to bridge the gap between synthetic simulation and real-world complexity, the zero-shot protocol was followed. The evaluation results of the generalization ability of the real-world scene are shown in Table 2.

[0085] Table 2 shows the generalization performance test results on a real-world severe weather dataset.

[0086] As shown in Table 2, the generalization performance test results on real-world severe weather datasets are presented. Table 2 shows that the biomimetic framework BioNet significantly enhances generalization ability. Notably, compared to the baseline model, VO-YOLOv11 achieves a significant 4.4% mAP improvement on the ExDark dataset. Although VO-YOLOv8 exhibits slight performance fluctuations on the DAWN dataset, it maintains superior accuracy on both the ExDark and RTTS datasets. Overall, the continuous improvement across heterogeneous domains demonstrates that the biomimetic framework BioNet successfully learns domain-invariant representations rather than overfitting to specific noise distributions.

[0087] (3) For large-scale scalability, the evaluation was conducted on the RTTS dataset, ExDark dataset and DAWN dataset; specifically, the bionic framework BioNet was integrated into YOLOv3 and compared with other state-of-the-art domain adaptation methods (such as GDIP-YOLO and ERUP-YOLO); the evaluation results for large-scale scalability are shown in Table 3.

[0088] Table 3 shows the comparison results of large-scale scalability on real severe weather datasets.

[0089] As shown in Table 3, the table presents the comparison results of large-scale scalability on real severe weather datasets, where Data Aug. indicates whether data augmentation was used during training. Table 3 shows that YOLOv3, integrated with the biomimetic framework BioNet, achieved the best performance on all three datasets, achieving mAP of 51.3%, 56.2%, and 40.5% on the RTTS, ExDark, and DAWN datasets, respectively. It is worth emphasizing that, unlike competing methods that rely on extensive data augmentation, the biomimetic framework BioNet achieves superior generalization performance without such requirements. This demonstrates that the visual optimization mechanism effectively extracts domain-invariant features, enabling the model to handle unseen real-world distortions better than explicit augmentation strategies.

[0090] (4) Scalability verification on the MS COCO dataset; Specifically, the bionic framework BioNet was integrated into two different standard convolutional neural network architectures, and the two different standard convolutional neural networks were used as baseline models. The performance saturation of the bionic framework BioNet on large-scale data was tested on the MS COCO dataset. The test results are shown in Table 4.

[0091] Table 4. Overall performance comparison on real-world datasets and synthetic contaminated data (AP50)

[0092] As shown in Table 4, the overall performance comparison (AP50) on real-world datasets and synthetic contaminated data is presented. For synthetic contaminated data, the results are averaged across three severity levels. Table 4 shows that extending the experiments to the MS COCO dataset tests performance saturation on large-scale data. Although training on the MS COCO dataset established a high baseline, the biomimetic framework BioNet continues to provide small but consistent gains. Specifically, VO-YOLOv11 achieved a 2.2% improvement on the DAWN dataset and continued to improve on synthetic contaminated benchmarks, while VO-YOLOv8 achieved similar improvements. These results confirm the scalability of the biomimetic framework BioNet, demonstrating its effectiveness even when applied to models trained on a wide range of highly diverse datasets.

[0093] Visual analysis: (1) Detection results in damaged scenarios: To intuitively demonstrate the effectiveness of the bionic framework BioNet, BioNet was integrated into YOLOv11. On the VOC dataset, the target detection results under synthetic fog and snow interference (severity level 3) were qualitatively compared; see attached. Figure 7 As shown, attached Figure 7 The figure presents a qualitative comparison of the object detection results of different methods on the VOC dataset; among them, the appendix... Figure 7 The first row of images shows the object detection results of YOLOv11 on the VOC dataset, with appended images. Figure 7 The second row of images shows the object detection results of YOLOv11, which integrates the bionic framework BioNet, on the VOC dataset.

[0094] From the appendix Figure 7 As can be seen, in the heavily noisy environment of the VOC dataset, YOLOv11 suffers from severe feature degradation and often has difficulty distinguishing the target object from the background interference. In contrast, the bionic framework BioNet can effectively suppress the simulation distortion of YOLOv11, maintain stable detection capabilities and accurate localization, and further confirm that the bionic framework BioNet can not only generalize to natural domain styles, but also has inherent robustness to significant statistical perturbations.

[0095] (2) Real-world scene detection results: As a supplement to the analysis results of the synthetic damaged dataset, the bionic framework BioNet was integrated into YOLOv11, and the zero-shot target detection results were qualitatively compared on the real-world datasets ExDark (low light) and DAWN (severe weather); see attached. Figure 8 As shown, attached Figure 8The figure presents a qualitative comparison of object detection results of different methods on real-world datasets; among them, appendix Figure 8 The first row of images shows the zero-shot object detection results of YOLOv11 on the real-world datasets ExDark (low light) and DAWN (severe weather). Figure 8 The second row of images shows the zero-shot target detection results of YOLOv11, which integrates the bionic framework BioNet, on the real-world datasets ExDark (low light) and DAWN (severe weather).

[0096] From the appendix Figure 8 As can be seen, under conditions of dense fog and low light, YOLOv11 is prone to missing small objects at a distance or misclassifying background noise as positive samples. In contrast, YOLOv11, which integrates the bionic framework BioNet, can accurately locate challenging samples and generate significantly more compact and accurate bounding boxes.

[0097] (3) Heatmap results: The bionic framework BioNet was integrated into YOLOv11, and visual comparison was performed on the feature attention heatmap; as shown in the attached figure. Figure 9 As shown, attached Figure 9 The image presents a visual comparison of different methods on feature attention heatmaps; among them, appendix... Figure 9 The first row of images shows the visual results of YOLOv11's feature attention heatmap, with accompanying images. Figure 9 The second row of images shows the visual results of the feature attention heatmap of YOLOv11 integrated with the bio-inspired framework BioNet.

[0098] From the appendix Figure 9 As can be seen, YOLOv11, which integrates the bionic framework BioNet, exhibits stronger anti-interference capabilities, generating activation maps that are more precisely focused on the target object while effectively suppressing background clutter. This is because by fusing the visual focusing simulation module and the olfactory ball simulation module into the bionic framework BioNet, background noise is effectively suppressed and feature saliency is enhanced. Comparative visualization results confirm that the activations generated by the bionic framework BioNet are significantly more focused on the target of interest, avoiding cluttered background responses, thus verifying its superior detection robustness.

[0099] The zero-shot anti-interference target detection method described in Embodiment 1 aims to mitigate the severe performance degradation of traditional detectors under interference conditions. It integrates the biomimetic framework BioNet into a standard convolutional neural network to construct a target detection model, improving the inherent robustness by refining the detector's network topology design. Specifically, the visual focusing simulation module simulates the retinal center-periphery antagonism mechanism, using a zero-sum balance to suppress spatially correlated low-frequency noise while preserving high-frequency cues. The olfactory ball simulation module integrates the olfactory ball side-inhibition mechanism, utilizing a winner-take-all strategy to decorrelate features in the semantic space and suppress background. The target detection model is trained only on clean data, requiring neither prior knowledge of damaged data nor additional restoration preprocessing, demonstrating strong zero-shot generalization capabilities against unknown weather changes and sensor interference in open-world scenarios.

[0100] Example 2 As attached Figure 10 As shown in the figure, this embodiment 2 provides a zero-sample anti-interference target detection system, including an image acquisition module and a target detection module.

[0101] The image acquisition module is used to acquire the image to be detected; the target detection module is used to input the image to be detected into the target detection model and output the target detection result of the image to be detected.

[0102] In this embodiment 2, the target detection model is constructed based on an improved convolutional neural network architecture. The improved convolutional neural network is a standard convolutional neural network that integrates the biomimetic framework BioNet. The biomimetic framework BioNet includes a visual focusing simulation module and a olfactory ball simulation module.

[0103] The visual focusing simulation module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian nuclei of different scales to filter additive noise; the olfactory bulb simulation module is used to reconstruct context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms.

[0104] Example 3 As attached Figure 11 As shown, this embodiment 3 provides an electronic device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the zero-sample anti-interference target detection method; or, the processor executing the computer program to implement the functions of each module in the above-mentioned zero-sample anti-interference target detection system.

[0105] For example, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing a preset function, the instruction segments describing the execution process of the computer program in the electronic device.

[0106] The electronic device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the above are examples of electronic devices and do not constitute a limitation on the electronic device. It may include more components than described above, or combine certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, etc.

[0107] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor, or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device through various interfaces and lines.

[0108] The memory can be used to store the computer program and / or module. The processor implements various functions of the electronic device by running or executing the computer program and / or module stored in the memory and by calling the data stored in the memory.

[0109] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards, secure digital cards, flash memory cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0110] Example 4 This embodiment 4 also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the zero-sample anti-interference target detection method.

[0111] If the modules / units integrated in the zero-sample anti-interference target detection system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0112] Based on this understanding, the present invention can implement all or part of the processes in the above-described zero-sample anti-interference target detection method, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described zero-sample anti-interference target detection method. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or a preset intermediate form, etc.

[0113] The computer-readable storage medium may include any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0114] Example 5 This embodiment 5 provides a computer product, which includes a computer program stored in a computer-readable storage medium. The processor of the electronic device reads the computer program from the computer-readable storage medium and executes the computer program, so that the electronic device can perform the zero-sample anti-interference target detection method described in embodiment 1, which will not be described again here.

[0115] It should be noted that those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above methods.

[0116] The zero-sample anti-interference target detection method of this invention integrates the biomimetic framework BioNet into a standard convolutional neural network to construct a target detection model. Relying on the biomimetic framework BioNet, it incorporates two biological perception mechanisms, retinal center-peripheral antagonism and olfactory bulb side inhibition, into the target detection model. It eliminates the need to pre-set noise types, simulate specific interference samples, or add an image restoration preprocessing module before detection, effectively reducing the hardware resource consumption and system operation complexity of model training, and reducing additional computational overhead.

[0117] In this invention, a visual focusing simulation module is used to simulate the central-peripheral antagonistic mechanism of the retina. A zero-sum balanced filtering mechanism is constructed using a dual-scale Gaussian kernel. Before feature encoding, physical interference such as additive noise and contrast attenuation is actively filtered out while retaining high-frequency key features such as target edges and textures, thus achieving intrinsic denoising of low-level features. The olfactory bulb simulation module replicates the olfactory bulb neural side inhibition circuit. With dual-path parallel processing and a dual-stage inhibition mechanism, feature decorrelation and background interference suppression are completed in the high-level semantic space. This forces feature sparsity and highlights salient targets, effectively solving the semantic ambiguity problem in complex scenarios.

[0118] In this invention, the object detection model only needs to be trained on a clean dataset to endow the network with inherent robustness, achieving zero-sample generalization capability for unobserved disturbances such as unknown severe weather, low light, and motion blur, avoiding performance drops caused by domain shift, and the robustness improvement is more significant in scenarios with more severe interference. In addition, both the visual focusing simulation module and the olfactory ball simulation module adopt a plug-and-play design, which can be seamlessly adapted to mainstream object detection architectures such as YOLO series and RT-DETR, and stably enhance the model's anti-interference ability without sacrificing or even slightly improving the detection accuracy of clean data, with good compatibility and scalability. This invention constructs a robust detection framework based on biological perception mechanisms, providing a new technical path for object detection in open-world scenarios, which can meet the industrial-grade requirements for detection stability and reliability in practical applications such as autonomous driving, intelligent monitoring, and robot vision, and has high practical value and promotion prospects.

[0119] It should be noted that in open-world scenarios, the performance of object detection often degrades due to environmental disturbances such as severe weather. Existing robustness enhancement methods typically suffer from insufficient generalization ability due to over-reliance on specific noise prior knowledge. This invention provides the biomimetic framework BioNet, which aims to achieve robust object detection with generalization capabilities. The biomimetic framework BioNet mainly comprises two plug-and-play visual focusing simulation modules and an olfactory ball simulation module. The visual focusing simulation module utilizes the center-periphery antagonism mechanism of the retina to filter out low-frequency noise, while the olfactory ball simulation module suppresses background interference in the semantic space by mimicking the lateral inhibition neural circuit of the olfactory ball. The biomimetic framework BioNet can endow the model with inherent robustness by training only on clean data, thus eliminating the need to use damaged samples. Extensive experiments on synthetic and real-world benchmarks demonstrate that the biomimetic framework BioNet achieves stable performance improvements across various network architectures, and its generalization ability significantly outperforms existing state-of-the-art methods.

[0120] In this invention, the biomimetic framework BioNet draws inspiration from the fundamental mechanisms of biological sensory systems, specifically from the robust signal processing mechanisms inherent in biological visual and olfactory systems, which can maintain perceptual stability in variable environments. In the visual pathway, the center-periphery antagonism mechanism of retinal ganglion cells plays a crucial role. By constructing dual Gaussian nuclei at different scales, the biological visual system establishes a numerical zero-sum balance, canceling out spatially consistent background noise while preserving high-frequency features. Secondly, to cope with more complex and deeper semantic interference beyond simple noise, a lateral inhibition mechanism observed in the olfactory bulb is fused. Similar to how the olfactory system inhibits neighboring neurons in complex background odors to achieve a winner-take-all effect, this mechanism helps to highlight salient targets while suppressing redundant background noise.

[0121] Based on the aforementioned biological principles, a robust target detection framework independent of training on damaged samples is proposed. The biomimetic framework BioNet incorporates two plug-and-play biomimetic modules: a visual focusing simulation module and an olfactory ball simulation module. The visual focusing simulation module is placed between layers of the feature extraction backbone network, utilizing dual Gaussian kernels of different scales to simulate retinal antagonism and explicitly filter out additive noise. The olfactory ball simulation module is integrated into the deep layers of the feature pyramid, reconstructing contextual information through neural lateral inhibition. Unlike existing methods that rely on recovery or adversarial training, the biomimetic framework BioNet provides zero-shot generalization capability, training only on clean images without using any damaged samples or introducing targeted denoising steps. Experimental results demonstrate that, solely based on the design of the biomimetic framework BioNet, the target detection model exhibits excellent robustness against a variety of unseen complex interferences, providing a new perspective for improving the reliability of target detectors in real-world open environments.

[0122] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

Claims

1. A zero-shot, interference-resistant target detection method, characterized in that, include: Acquire the image to be detected; The image to be detected is input into the target detection model, and the target detection result of the image to be detected is output. Among them, the object detection model is built based on the architecture of an improved convolutional neural network, which is a standard convolutional neural network that integrates the bio-inspired framework BioNet. The biomimetic framework BioNet includes a visual focusing simulation module and an olfactory ball simulation module; The visual focusing simulation module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian kernels of different scales to filter additive noise; the olfactory bulb simulation module is used to reconstruct the context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms. The visual focusing simulation module is embedded in the feature extraction backbone of the standard convolutional neural network and placed between two adjacent convolutional units in the feature extraction backbone. The snorting ball simulation module is positioned between the feature fusion network of a standard convolutional neural network and the detection head; The visual focusing simulation module includes a center-periphery layer unit, a differential response calculation unit, a channel attention unit, and a residual fusion unit; The center-periphery layer unit is used to perform convolution operations on the input feature map of the visual focusing simulation module using two Gaussian kernels of different scales to obtain the center signal and the peripheral signal; The differential response calculation unit is used to calculate the differential signals of the central signal and the peripheral signal respectively, and sum them after nonlinear activation to simulate the responses of ON-type neurons and OFF-type neurons in the retina and obtain the retinal response output. The channel attention unit is used to perform attention weighting on the retinal response output based on pre-calculated channel weighting coefficients to obtain the weighted response; The residual fusion unit is used to multiply the weighted response by a predetermined scaling factor and add it to the input feature map of the visual focusing simulation module to obtain the output of the visual focusing simulation module. The olfactory bulb simulation module includes an information aggregation unit, a dual-pathway splitting unit, a tufted cell pathway unit, a monk's cap cell pathway unit, and a feature fusion and residual output unit. The information aggregation unit is used to aggregate the initial signal of the input feature map of the sniffing ball simulation module to obtain the initial signal aggregated feature map; The dual-path splitting unit is used to divide the initial signal aggregation feature map into two independent subsets along the channel dimension to obtain the clustered path input feature map and the monk's cap path input feature map. The plexiform cell pathway unit is used to capture local high-frequency features using a pre-defined restricted receptive field. Specifically, a pre-determined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of capturing local high-frequency features, so as to obtain the features after plexiform pathway inhibition. The capillary cell pathway unit is used to encode global spatial context using a pre-defined broad receptive field. Specifically, a pre-determined lateral inhibition mechanism is used to perform local lateral inhibition and global feedback inhibition on any intermediate feature map in the process of encoding global spatial context, so as to obtain the features after capillary pathway inhibition. The feature fusion and residual output unit is used to concatenate the features after the clustered pathway is suppressed with the features after the monk's cap pathway is suppressed. After convolution mapping, the concatenation is added to the input feature map of the olfactory ball simulation module to obtain the output of the olfactory ball simulation module.

2. The zero-sample anti-interference target detection method according to claim 1, characterized in that, The center-periphery layer unit includes parallel center convolutional pathways and peripheral convolutional pathways; the Gaussian kernel scale used in the center convolutional pathways and peripheral convolutional pathways is different.

3. The zero-sample anti-interference target detection method according to claim 1, characterized in that, Predetermined lateral inhibition mechanisms include local lateral inhibition mechanisms and global feedback inhibition mechanisms; The local side suppression mechanism is used to calculate the average activation in the local spatial neighborhood of any intermediate feature map as the suppression signal; the suppression signal is applied to any intermediate feature map to obtain the locally suppressed features; A global feedback suppression mechanism is used to perform global average pooling on the features after local suppression and then pass them through a fully connected layer to obtain a global suppression signal. The global suppression signal is then applied to the features after local suppression to obtain features after clustered pathway suppression or monk's cap pathway suppression.

4. A zero-sample anti-interference target detection system, characterized in that, To implement the zero-sample anti-interference target detection method as described in any one of claims 1-3, comprising: The image acquisition module is used to acquire the image to be detected; The target detection module is used to input the image to be detected into the target detection model and output the target detection result of the image to be detected. Among them, the object detection model is built based on the architecture of an improved convolutional neural network, which is a standard convolutional neural network that integrates the bio-inspired framework BioNet. The biomimetic framework BioNet includes a visual focusing simulation module and an olfactory ball simulation module; The visual focusing simulation module is used to simulate the center-periphery antagonism mechanism of the retina by using dual Gaussian nuclei of different scales to filter additive noise; the olfactory bulb simulation module is used to reconstruct context-aware feature relationships by simulating the neural interaction network of the olfactory bulb and utilizing parallel signal processing and lateral inhibition mechanisms.

5. The present invention also provides an electronic device, characterized in that, include: A processor is used to execute computer programs; A computer-readable storage medium storing a computer program, which, when executed by the processor, performs the zero-sample anti-interference target detection method as described in any one of claims 1-3.

6. The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the zero-sample anti-interference target detection method as described in any one of claims 1-3.

7. The present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the zero-sample anti-interference target detection method as described in any one of claims 1-3.

Images