Image recognition method and device, computer device and storage medium

Abstract

This invention discloses an image recognition method, apparatus, computer device, and storage medium. The method includes: acquiring an amplitude map and a corresponding depth map to be recognized; performing feature extraction processing based on the amplitude map and depth map to obtain amplitude features of the amplitude map and depth features of the depth map; performing feature fusion processing on the amplitude features and depth features based on the channel saliency of the amplitude features and depth features to obtain fused features; performing attention enhancement processing based on the fused features to obtain enhanced fused features; and providing the enhanced fused features to a preset target detection network to obtain the corresponding image recognition result. By extracting features from the amplitude map and depth map separately, performing weighted fusion based on channel saliency, and combining attention enhancement processing, the reflection intensity information of the amplitude map and the distance structure information of the depth map can effectively complement each other at the feature level, thereby achieving high recognition accuracy and reliability even in no-light or low-light scenes.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to an image recognition method, apparatus, computer device, and storage medium. Background Technology

[0002] In industrial production lines, warehouses, and nighttime workshops, continuous monitoring of work areas is typically required to prevent accidental entry of personnel, foreign object intrusion, or equipment malfunction. Currently, the commonly used detection method involves acquiring color images using RGB cameras and then performing target recognition. While this method achieves relatively stable recognition results under normal lighting conditions, in low-light or unlit environments, the signal-to-noise ratio of RGB images decreases significantly, detail information is difficult to preserve, target outlines become blurred, and the detection accuracy and stability of existing recognition models decline markedly.

[0003] To compensate for the limitations of visible light imaging in low-light environments, existing technologies attempt to combine infrared or thermal imaging with RGB images for joint recognition to enhance target perception. However, these solutions still rely primarily on RGB images. In the absence of light or extremely low light conditions, RGB images themselves cannot provide the structured details necessary for effective recognition, thus limiting the overall recognition performance. Furthermore, for small targets or areas with complex backgrounds, relying solely on infrared or thermal imaging is insufficient to guarantee detection accuracy.

[0004] Therefore, how to provide an image recognition method that can accurately detect small targets in the absence of light or in low light conditions has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] Therefore, it is necessary to provide an image recognition method, device, computer equipment, and storage medium to address the above-mentioned technical problems, which has the advantage of accurately detecting small-sized targets in the absence of light or in low-light environments.

[0006] An image recognition method, the method comprising:

[0007] Obtain the amplitude map and corresponding depth map of the target data;

[0008] Based on the amplitude map and the depth map, feature extraction processing is performed to obtain the amplitude features of the amplitude map and the depth features of the depth map;

[0009] Based on the channel saliency of the amplitude feature and the depth feature, feature fusion processing is performed on the amplitude feature and the depth feature to obtain the fused feature;

[0010] Based on the fusion features, attention enhancement processing is performed to obtain enhanced fusion features;

[0011] The enhanced fusion features are provided to a preset target detection network to obtain the corresponding image recognition results.

[0012] Optionally, obtaining the amplitude map to be identified and the corresponding depth map includes:

[0013] Obtain the initial amplitude map and initial depth map acquired by the target depth camera;

[0014] The pixel values ​​of each pixel in the initial amplitude map and the initial depth map are mapped to a preset numerical range to obtain the amplitude map and the depth map to be aligned.

[0015] Alignment processing is performed on the amplitude map to be aligned and the depth map to be aligned to obtain the amplitude map to be identified and the corresponding depth map.

[0016] Optionally, the step of performing feature fusion processing on the amplitude feature and the depth feature based on the channel saliency of the amplitude feature and the depth feature to obtain fused features includes:

[0017] Based on the channel saliency, a first fusion weight for the amplitude feature and a second fusion weight for the depth feature are generated;

[0018] The amplitude feature and the depth feature are weighted and fused according to the first fusion weight and the second fusion weight to obtain the fused feature.

[0019] Optionally, generating the first fusion weight of the amplitude feature and the second fusion weight of the depth feature based on the channel saliency includes:

[0020] Based on the channel saliency, a first weight score is calculated to characterize the importance of the amplitude feature and a second weight score is calculated to characterize the importance of the depth feature;

[0021] The first fusion weight is calculated based on the first weight score and the second weight score;

[0022] The second fusion weight is calculated based on the first fusion weight.

[0023] Optionally, the attention enhancement processing based on the fusion features to obtain enhanced fusion features includes:

[0024] Channel attention processing is performed based on the fusion features to obtain the first enhanced feature;

[0025] Spatial attention processing is performed based on the fusion features to obtain the second enhanced feature;

[0026] The first enhancement feature and the second enhancement feature are multiplied element-wise to obtain the enhanced fusion feature.

[0027] Optionally, the image recognition result includes the coordinates of the recognition box and the recognition confidence score. After providing the enhanced fusion feature to a preset target detection network to obtain the corresponding image recognition result, the method further includes:

[0028] Based on the coordinates, calculate the intersection-union ratio (IUU) between the recognition box and the preset warning area;

[0029] When the cross-union ratio is not less than a preset cross-union ratio threshold and the identification confidence level is not less than a preset confidence threshold, it is determined that an intrusion exists in the warning area.

[0030] Optionally, the method further includes:

[0031] Obtain the recognition results of multiple frames within a preset time window. The recognition results of multiple frames include the coordinates of the recognition box of each frame and the corresponding recognition confidence.

[0032] Based on the changes in the bounding box coordinates and the changes in recognition confidence in the multi-frame image recognition results, a stability index is calculated to characterize the stability of the recognition results.

[0033] The cross-union ratio (CUNR) threshold and the identification confidence threshold are dynamically adjusted based on the stability index. Specifically, when the stability index meets a first preset condition, the CUNR threshold and / or the identification confidence threshold are decreased, and when the stability index meets a second preset condition, the CUNR threshold and / or the identification confidence threshold are increased.

[0034] An image recognition device, the device comprising:

[0035] The first acquisition module is used to acquire the amplitude map to be identified and the corresponding depth map;

[0036] The feature extraction module is used to perform feature extraction processing based on the amplitude map and the depth map to obtain the amplitude features of the amplitude map and the depth features of the depth map;

[0037] The feature fusion module is used to perform feature fusion processing on the amplitude feature and the depth feature based on the channel saliency of the amplitude feature and the depth feature to obtain fused features;

[0038] The feature enhancement module is used to perform attention enhancement processing based on the fused features to obtain enhanced fused features;

[0039] The target detection module is used to provide the enhanced fusion features to a preset target detection network to obtain the corresponding image recognition results.

[0040] A computer device includes a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor, wherein the processor implements the image recognition method described above when executing the computer-readable instructions.

[0041] A readable storage medium having computer-readable instructions stored thereon, which, when executed by a processor, implement the above-described image recognition method.

[0042] The aforementioned image recognition method, apparatus, computer equipment, and storage medium acquire an amplitude map and a corresponding depth map to be recognized; perform feature extraction processing based on the amplitude map and the depth map to obtain amplitude features of the amplitude map and depth features of the depth map; perform feature fusion processing on the amplitude features and the depth features based on the channel saliency of the amplitude features and the depth features to obtain fused features; perform attention enhancement processing on the fused features to obtain enhanced fused features; and provide the enhanced fused features to a preset target detection network to obtain the corresponding image recognition result. By extracting features from the amplitude map and the depth map separately, performing weighted fusion based on channel saliency, and combining attention enhancement processing, the reflection intensity information from the amplitude map and the distance structure information from the depth map can effectively complement each other at the feature level, thereby avoiding the problem of detail loss under low light conditions. This makes the feature representation provided to the target detection network more stable and highlights key areas, achieving high recognition accuracy and reliability even in no-light or low-light scenes. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic flowchart of an image recognition method according to an embodiment of the present invention;

[0045] Figure 2 This is a schematic diagram of the structure of an image recognition device according to an embodiment of the present invention;

[0046] Figure 3 This is a schematic diagram of a computer device according to an embodiment of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] In one embodiment, such as Figure 1 As shown, an image recognition method is provided, including the following steps:

[0049] 101. Obtain the amplitude map and the corresponding depth map to be identified.

[0050] In this embodiment of the invention, the image recognition method described above can be specifically applied to a depth camera, which can acquire the amplitude map and the depth map. When acquiring the amplitude map and the corresponding depth map to be recognized, the depth camera, through its internal infrared emission and measurement mechanism, can simultaneously output an amplitude map reflecting the reflection intensity and a depth map representing the distance from the object to the camera. In practical applications, the amplitude map is usually generated by the depth camera after emitting infrared light, based on the intensity of the returned light. Its pixel values ​​can reflect the reflection characteristics of different surfaces in the scene. For example, objects with smooth surfaces or high reflectivity usually have larger pixel values ​​in the amplitude map, while objects with strong light absorption or darker materials have lower pixel values. The depth map calculates the distance information of each pixel based on the time-of-flight (ToF) principle or structured light principle, and represents the distance relationship in the form of grayscale values ​​or 16-bit pixel values. For example, areas closer to the camera have smaller pixel values, and areas farther away have larger pixel values.

[0051] By acquiring both types of images simultaneously, relatively stable imaging information can be obtained even in low-light or no-light environments, providing a reliable data foundation for subsequent feature extraction and recognition.

[0052] 102. Based on the amplitude map and depth map, feature extraction processing is performed to obtain the amplitude features of the amplitude map and the depth features of the depth map.

[0053] In this embodiment of the invention, two types of images can be input into different feature extraction pathways. Through operations such as convolution, nonlinear activation, and downsampling, higher-level representation information is gradually extracted from the original pixels. For the amplitude map, its original pixels mainly reflect the distribution of reflection intensity at various locations in the scene. The feature extraction process can transform intensity changes into more abstract feature representations such as edges, textures, and local contrast. For the depth map, its pixel values ​​correspond to the distances between objects in the scene and the camera. The feature extraction process can gradually transform simple distance distributions into depth features related to object contours, shapes, and spatial structures. For example, near the edge of a device or a step, there will be obvious abrupt changes in distance in the depth map. Through convolution operations, the abrupt change information can be aggregated into feature responses reflecting shape boundaries.

[0054] In one possible embodiment, separate convolutional feature extraction networks can be set up for the amplitude map and the depth map, respectively. For example, the amplitude map is input into a first convolutional branch, and the depth map is input into a second convolutional branch. Each branch may include multiple convolutional layers, batch normalization layers, and nonlinear activation layers. Spatial resolution is gradually compressed and the number of channels is increased through stride or pooling operations to extract feature representations at different scales and semantic levels. After the above processing, the intermediate features output by the first convolutional branch are used as amplitude features, and the intermediate features output by the second convolutional branch are used as depth features for subsequent channel saliency evaluation and feature fusion.

[0055] 103. Based on the channel saliency of amplitude and depth features, feature fusion processing is performed on amplitude and depth features to obtain fused features.

[0056] In this embodiment of the invention, the response of the two types of features across each channel dimension can be analyzed first to determine the importance of different channels in the current scene. Since amplitude features and depth features originate from different sources, they may contain different types of valid information at the same location. For example, amplitude features respond more strongly in areas with significant reflection differences, while depth features are more sensitive at locations with abrupt changes in shape and distance. By statistically analyzing and measuring the response values ​​of each channel (i.e., convolutional channels), a set of channel saliency indicators can be obtained to characterize the relative importance of that channel.

[0057] After obtaining the saliency of the channels, corresponding fusion weights can be generated based on their magnitudes, so that channels with higher importance receive larger weights during fusion, while channels with lower importance receive relatively smaller weights. The fusion method can employ a weighted summation approach, where amplitude and depth features are scaled proportionally according to their respective weights, and then the scaled features are summed element-wise to obtain the fused feature representation. This allows for the selective retention of effective information from both types of features, enhancing the more discriminative feature components in the current scene while maintaining overall structural information.

[0058] In one possible implementation, channel saliency can be obtained by calculating the mean, variance, maximum value, or other statistics of each channel feature, or by using a lightweight weight generation network to input the two types of features and output the corresponding channel weights. In this case, the fusion weights under different scenarios can be dynamically adjusted according to changes in the input features, making the fused features more adaptable to the current image content and providing a more discriminative representation for subsequent attention enhancement and object detection.

[0059] 104. Attention enhancement processing is performed based on the fusion features to obtain enhanced fusion features.

[0060] In this embodiment of the invention, the effective regions related to the target can be further emphasized based on the fused features, while suppressing background noise or unimportant feature components. By applying an attention mechanism to the fused features, the response of the features can be analyzed in both the channel and spatial dimensions, enabling the model to more accurately focus on the parts that have a greater impact on the recognition result. For example, in the channel dimension, by statistically analyzing the overall response of each channel, channels that contribute more to the current scene can be highlighted; in the spatial dimension, by weighting the responses at different locations in the feature map, the feature intensity near the target region can be enhanced, while the interference information in the background region can be weakened.

[0061] In one possible embodiment, channel attention processing can be performed on the fused features first. By pooling the average or maximum values ​​of the fused features in the spatial dimension, a descriptive vector reflecting the overall importance of each channel can be obtained. Then, a series of nonlinear transformations are used to generate channel attention weights, and these weights are applied to the corresponding channels of the fused features to obtain the first enhanced feature after channel enhancement.

[0062] Subsequently, spatial attention processing can be applied to the fused features. A spatial feature description map is constructed by converging along the channel dimension, and then a spatial attention map representing the importance of each spatial location is generated through convolution. This attention map is then multiplied positionally with the fused features to obtain the spatially enhanced second feature. Finally, the two enhanced features are multiplied element-wise to superimpose the effects of channel enhancement and spatial enhancement, resulting in the final enhanced fused feature, which possesses stronger expressive and discriminative capabilities in subsequent object detection stages.

[0063] In another possible implementation, when performing attention enhancement processing based on fused features, channel attention and spatial attention can be performed simultaneously without separating them. Instead, a joint attention mechanism can be used to model the fused features concurrently. The joint attention mechanism can uniformly characterize the multi-dimensional relationships of the features by constructing a joint weight matrix of the fused features in both the channel and spatial dimensions. For example, the fused features can be input into a lightweight attention encoding network, generating a weight map with the same size as the fused features through multiple convolutions or pointwise convolutions. This weight map can simultaneously describe the importance of the features in both the channel and spatial dimensions. Subsequently, multiplying this weight map element-wise with the fused features yields the enhanced fused features.

[0064] In another possible implementation, attention enhancement can be achieved through self-attention. Specifically, the fused features can be divided into several feature blocks. By calculating the similarity between feature blocks, weights reflecting the correlation between regions are obtained. Feature values ​​are then redistributed based on these weights, allowing the feature map to focus more on regions with structural correlations. This approach enhances long-range dependency information related to the target, even when the target is small or the feature distribution is sparse, thereby improving the stability of subsequent detection.

[0065] 105. The enhanced fusion features are provided to the preset target detection network to obtain the corresponding image recognition results.

[0066] In this embodiment of the invention, the object detection network can adopt a single-stage detection structure, such as a YOLO-style detection architecture, where feature extraction and candidate box prediction are performed simultaneously within the same network, eliminating the need for an additional candidate region generation process. This reduces inference latency and overall computational complexity. In this structure, a backbone network constructed from depthwise separable convolutions can be used as the feature extraction part. By decomposing standard convolutions into channel-wise and pointwise convolutions, the number of parameters and computational cost of convolution operations is reduced. Furthermore, a channel reduction strategy is employed to appropriately reduce the number of intermediate feature channels during network design, making the network more lightweight and suitable for operation on resource-constrained edge devices.

[0067] Building upon the aforementioned backbone network, a feature pyramid structure can be further introduced to fuse intermediate feature maps at different scales, enabling multi-scale object detection. Specifically, feature maps from different levels of the backbone network can be upsampled, downsampled, and laterally connected to construct top-down or bottom-up feature pyramids. This combines high-level semantic information with low-level spatial detail information, thus addressing the detection needs of both large and small objects. For smaller objects, such as slender tools, small parts, or locally protruding structures, the combined participation of multi-scale features can improve the discriminability of the target in the feature space, thereby increasing the detection rate of small targets. The output of the object detection network can include a set of candidate boxes (i.e., recognition boxes) predicted on feature maps at different scales. Each candidate box is associated with a corresponding target category prediction result and a confidence prediction result.

[0068] To obtain the aforementioned target detection network, supervised training can be performed offline. During training, the Adam optimizer can be used to update the network parameters. The initial learning rate is set to 0.01, and a cosine annealing strategy is used to dynamically adjust the learning rate, allowing it to gradually decay during training. This results in faster convergence in the early stages of training and more stable detail fitting capabilities in the later stages. The batch size can be set to 8, and iterative training can be performed on a large number of labeled samples. The number of training epochs can be set to 100, allowing the network to fully learn the appearance features and spatial distribution patterns of targets under different working conditions. The nonlinear activation function in the network can be ReLU, introducing nonlinear expressive power into the features during forward propagation while maintaining computational simplicity and ease of implementation on embedded platforms.

[0069] The loss function used during training can be defined as a weighted sum of classification loss and regression loss, i.e., L = L_cls + λL_reg. Here, L_cls can be the classification loss based on cross-entropy, used to constrain the target class prediction result; L_reg can be the regression loss based on CioU (Complete IoU), used to constrain the positional deviation, scale difference, and overlap between the predicted bounding box and the ground truth bounding box; and λ is a weighting coefficient used to balance the influence of classification loss and regression loss. By coordinating the optimizer, learning rate strategy, batch size, number of training epochs, and loss function settings, the object detection network can achieve a good fit on the fused features of the magnitude and depth maps while ensuring training stability. This results in a pre-defined object detection network with fixed parameters that can be directly invoked during the inference phase.

[0070] After training the network and obtaining the initial model, the model can be optimized for edge deployment. In one possible embodiment, the trained object detection network can be pruned first, removing channels or convolutional kernels that contribute little to the final recognition result to reduce model size and computational load. Subsequently, the pruned network can be quantized, converting network weights and some intermediate features from floating-point representation to fixed-point representation with a lower bit width, thereby further reducing storage footprint and computational resource requirements. The pruned and quantized model can be deployed on Jetson Nano or other embedded computing platforms, achieving an inference frame rate of at least 30 frames per second with an input resolution of 640×480, to meet the real-time detection needs of industrial environments.

[0071] During the inference phase, after the enhanced fusion features are processed by a pre-defined target detection network, the output image recognition result can include prediction information for several detected targets. Each detected target can include at least the coordinates of a bounding box, the corresponding recognition confidence score, and the recognition category. The coordinates of the bounding box can be represented as the coordinates of the top-left and bottom-right corners, or as the coordinates of the center point and the width and height. The recognition confidence score characterizes the network's belief that the target belongs to a certain category. The recognition category indicates the category to which the target belongs in a pre-defined category set. In one possible implementation, the recognition result of each frame can also include the number of detected targets and their internal ranking information, so that subsequent modules can filter or further process them based on the confidence score or the target location. The above image recognition results can be used as input data for subsequent warning area judgment and alarm logic to achieve automatic identification and response to intrusion or abnormal targets in industrial scenarios.

[0072] In this embodiment of the invention, an amplitude map and a corresponding depth map to be identified are obtained; feature extraction processing is performed based on the amplitude map and the depth map to obtain amplitude features of the amplitude map and depth features of the depth map; feature fusion processing is performed on the amplitude features and the depth features based on the channel saliency of the amplitude features and the depth features to obtain fused features; attention enhancement processing is performed based on the fused features to obtain enhanced fused features; the enhanced fused features are provided to a preset target detection network to obtain the corresponding image recognition result. By extracting features from the amplitude map and the depth map respectively, performing weighted fusion based on channel saliency, and combining attention enhancement processing, the reflection intensity information from the amplitude map and the distance structure information from the depth map can effectively complement each other at the feature level, thereby avoiding the problem of detail loss under low light conditions. This makes the feature expression finally provided to the target detection network more stable and highlights key areas, achieving high recognition accuracy and reliability even in no-light or low-light scenes.

[0073] It is understood that in the specific implementation of this application, data such as amplitude maps, depth maps, and image recognition results are involved. When the embodiments in this application are applied to specific products or technologies, user permission or consent is required, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0074] Optionally, in the step of obtaining the amplitude map to be identified and the corresponding depth map, the initial amplitude map and the initial depth map captured by the target depth camera can also be obtained; the pixel value of each pixel in the initial amplitude map and the initial depth map is mapped to a preset numerical range to obtain the amplitude map to be aligned and the depth map to be aligned; and alignment processing is performed based on the amplitude map to be aligned and the depth map to be aligned to obtain the amplitude map to be identified and the corresponding depth map.

[0075] In this embodiment of the invention, initial amplitude and initial depth maps synchronously acquired by a depth camera can be obtained first. During actual acquisition, the depth camera simultaneously outputs an amplitude map reflecting infrared return intensity and a depth map reflecting scene distance information. The pixel range and numerical distribution of the two types of images are usually inconsistent. Directly using them for subsequent feature extraction may lead to large differences in feature scale between different modalities. Therefore, pixel value normalization processing can be performed on the two types of initial images, mapping each pixel value to a preset numerical range, such as the [0,1] range. Through normalization processing, the numerical scale between the amplitude map and the depth map can be kept consistent, avoiding the dominance of features from one image due to excessively large scale during subsequent fusion, thereby improving the stability of feature extraction and fusion.

[0076] After normalization, spatial registration can be further performed on the amplitude and depth maps to be aligned, compensating for potential viewpoint deviations or geometric distortions that may exist during the acquisition of the two types of images. Specifically, an affine transformation matrix H can be used to perform pixel-level alignment of the amplitude and depth maps. Through operations such as rotation, translation, and scaling, the pixel positions of the amplitude and depth maps on the two-dimensional plane are made to correspond one-to-one. The affine matrix can be obtained through a calibration process or a robust estimation method (such as RANSAC). After applying the transformation, it can ensure, to the greatest extent possible, that the physical positions of the amplitude and depth maps are completely aligned in the same scene. After the above normalization and spatial alignment processing, a dual-channel input tensor containing the two types of images can be formed. As input to the subsequent feature extraction stage, the data collected by the depth camera can enter the neural network in a more consistent and standardized form, providing a more reliable input data foundation for subsequent feature extraction and recognition.

[0077] Optionally, in the step of performing feature fusion processing on amplitude features and depth features based on channel saliency to obtain fused features, a first fusion weight for amplitude features and a second fusion weight for depth features can be generated according to channel saliency; and the amplitude features and depth features can be weighted and fused according to the first fusion weight and the second fusion weight to obtain fused features.

[0078] In this embodiment of the invention, the response of amplitude and depth features across each channel dimension can be used as the criterion. By calculating the statistics of each channel (such as mean, variance, entropy, or other indicators that reflect the importance of the channel), a channel saliency distribution is obtained to measure the relative contribution of each channel in the amplitude and depth features. When the channel saliency of a certain channel is high, it indicates that the channel has a stronger discriminative ability in the current scene, and therefore a higher weight can be given to the channel during the fusion process; conversely, a lower weight is given.

[0079] When generating the fusion weights, based on the aforementioned channel saliency, a first fusion weight α for adjusting the amplitude feature and a second fusion weight β for adjusting the depth feature can be calculated, ensuring that their sum satisfies a certain constraint, such as α + β = 1. In a typical implementation, a weight generation module can be used to adjust the amplitude feature... with depth features The input image is processed, and the values ​​of α and β are dynamically obtained through convergence and nonlinear transformation of the feature channels. This dynamic weight generation mechanism can adaptively adjust the weights according to the actual feature distribution each time a different image is input, so that the fusion process does not depend on a fixed ratio but can automatically adapt to changes in the scene. The fused features can then be calculated as follows:

[0080]

[0081] in, For amplitude characteristics, For depth features, α and β are the dynamically generated fusion weights mentioned above. This weighted fusion method allows the reflection intensity information from the amplitude map and the spatial structure information from the depth map to each play their respective advantages during fusion, resulting in more comprehensive and stable representational capabilities in the final fused features. Especially in scenes with unstable lighting or high image noise, it can effectively improve the performance of subsequent attention enhancement and object detection.

[0082] In one possible embodiment, the weight generation module can employ a lightweight neural network structure, including one or more fully connected layers, channel-wise convolutional layers, or gated units, to learn the dynamic generation rules of the fusion weights from the input features. Alternatively, Softmax, Sigmoid, or other normalized activation methods can be used to ensure that the generated weights satisfy non-negativity and proportionality constraints, thereby making the fusion process more stable and reliable. All of the above methods can serve as alternative implementations of the feature fusion module for achieving dynamic weighted fusion of amplitude and depth features.

[0083] Finally, it should be noted that the aforementioned channels can be feature channels used to characterize the convolutional feature map in different semantic dimensions, that is, the channel dimensions in the multidimensional feature tensor formed after processing by the convolutional network. Each channel typically corresponds to a specific pattern or response type extracted by the convolutional kernel, such as edge texture, local structural changes, differences in reflection intensity distribution, depth gradient changes, etc. Therefore, the importance of different channels may not be consistent in different scenarios. In this embodiment of the invention, the calculation of channel saliency is based on the response of the convolutional feature map channels. By distinguishing which channels have more discriminative responses, the weights that amplitude features and depth features should bear during fusion are determined.

[0084] Optionally, in the step of generating the first fusion weight of amplitude features and the second fusion weight of depth features based on channel saliency, a first weight score for characterizing the importance of amplitude features and a second weight score for characterizing the importance of depth features can also be calculated based on channel saliency; the first fusion weight is calculated based on the first weight score and the second weight score; and the second fusion weight is calculated based on the first fusion weight.

[0085] In this embodiment of the invention, when generating the first fusion weight of amplitude feature and the second fusion weight of depth feature based on channel saliency, normalization calculation can be performed using intermediate weight scores to make the weight distribution between the two features more stable and controllable.

[0086] Specifically, amplitude features can be analyzed based on channel saliency. with depth features By performing a weighted summation or linear transformation, we obtain a first weight score to characterize the importance of amplitude features and a second weight score to characterize the importance of depth features.

[0087] For example, weighting parameters for amplitude features can be preset. and weight parameters used for deep features , and amplitude characteristics Multiplying them yields the first weight score, which is then used to combine the deep features with... The second weighted score is obtained by multiplying the scores, thereby introducing a learnable parameter based on the channel significance to finely characterize the contribution of different channels or different modes.

[0088] After obtaining the first and second weight scores, the first fusion weight can be calculated using exponential normalization, ensuring good numerical stability and interpretability in the weight allocation between the two modes. In one implementation, the first fusion weight α can be calculated as follows:

[0089]

[0090] in, and These represent the weighted results after applying the corresponding weight parameters to the amplitude feature and the depth feature, respectively. This represents exponential operations. Through the above normalization process, it can be ensured that the first fusion weight α is always between 0 and 1, and increases accordingly as the magnitude feature weight score increases.

[0091] Subsequently, the second fusion weight β can be determined based on the first fusion weight, for example, β can be set to 1. α is used to make the sum of the two constants, thus forming a set of complementary fusion weights.

[0092] By using the above methods, the proportions of amplitude features and depth features in the fusion process can be adaptively adjusted under the combined effect of channel saliency and learnable parameters, so that the fused features can maintain a suitable modal balance in different scenarios, thereby improving the stability and discriminative ability of the overall feature representation.

[0093] Optionally, in the step of performing attention enhancement processing based on the fusion features to obtain enhanced fusion features, channel attention processing can be performed based on the fusion features to obtain a first enhanced feature; spatial attention processing can be performed based on the fusion features to obtain a second enhanced feature; and element-wise multiplication can be performed on the first enhanced feature and the second enhanced feature to obtain the enhanced fusion feature.

[0094] In this embodiment of the invention, channel attention processing can be performed on the fused features first, and the fused features can be converged in the spatial dimension to obtain a channel description vector that reflects the overall response strength of each channel, and channel attention weights can be generated based on the channel description vector; then, the channel attention weights are applied to each channel corresponding to the fused features, the channels with stronger responses and more relevant to the target are enhanced, and the channels with weaker responses and mainly from the background are suppressed, thereby obtaining the first enhanced feature that has been enhanced in the channel dimension.

[0095] Simultaneously, spatial attention processing can be applied to the fused features. By converging the features along the channel dimension, a spatial description map is obtained to characterize the importance of each spatial location. Then, a spatial attention weight map is generated based on this spatial description map. This spatial attention weight map is multiplied with the fused features position by position, so that features located near the target region are enhanced, while features located in the background region are relatively weakened, thus obtaining a second enhanced feature that has been enhanced in the spatial dimension.

[0096] Subsequently, element-wise multiplication can be performed on the first and second enhancement features to superimpose the effects of channel enhancement and spatial enhancement onto the same feature map, resulting in the final enhanced fusion feature.

[0097] When represented symbolically, the enhanced fusion feature can be expressed as:

[0098]

[0099] in, As a feature of fusion, This represents the features after applying channel attention to the fused features. The symbol represents the feature after spatial attention has been applied to the fused feature. This indicates that element-wise multiplication of the two feature maps is performed. Through the aforementioned channel-space joint attention mechanism, the response strength of the fused features to the target region can be effectively improved without increasing computational overhead, further enhancing the accuracy and robustness of subsequent target detection.

[0100] Optionally, the image recognition result includes the coordinates of the recognition box and the recognition confidence score. After the step of providing the enhanced fusion features to the preset target detection network to obtain the corresponding image recognition result, the cross-union ratio (CUP) between the recognition box and the preset warning area can be calculated based on the coordinates. When the CUP is not less than the preset CUP threshold and the recognition confidence score is not less than the preset confidence threshold, it is determined that there is an intrusion in the warning area.

[0101] In this embodiment of the invention, after inputting the enhanced fusion features into a preset target detection network and obtaining the image recognition result, it is possible to further determine whether there is an intrusion in the preset warning area based on the recognition result.

[0102] Specifically, the image recognition result output by the object detection network for each frame can include multiple candidate targets. Each candidate target contains at least the coordinate information of the recognition box and the recognition confidence score. The recognition result can also include a category label to represent the target category. To avoid the same target being repeatedly labeled by multiple close recognition boxes, non-maximum suppression processing can be performed on the multiple candidate recognition boxes output by the network. For example, the NMS algorithm can be used to filter out recognition boxes with high overlap and low confidence scores based on the degree of overlap between candidate boxes and the corresponding recognition confidence scores, and retain a set of final recognition boxes with relatively independent positions and high confidence scores.

[0103] One or more virtual alert zones can be pre-configured for the monitoring screen to represent areas requiring special protection. These virtual alert zones can be defined using a polygon mask, which determines a polygonal region on the image plane using the coordinates of multiple vertices. During implementation, the polygon mask is used to determine whether a location falls within the alert zone.

[0104] After completing nonmaximum suppression and obtaining the final bounding box, the intersection-union ratio (IUR) between the bounding box and the warning region can be calculated based on their coordinates and geometry. The IUR can be defined as the ratio of the area of ​​the intersection of the bounding box and the warning region to the area of ​​their union.

[0105] After calculating the Cross-Union Ratio (CUI), the CUI is compared with a preset CUI threshold, and the recognition confidence is compared with a preset confidence threshold. When the CUI is not less than the preset CUI threshold and the recognition confidence is not less than the preset confidence threshold, it can be determined that there is an intrusion target in the warning area, and an audible and visual alarm device is triggered accordingly, such as controlling the buzzer to sound and the warning light to flash. The alarm time, the corresponding recognition result, and related image information are recorded in the log for subsequent tracking and analysis.

[0106] In one possible embodiment, intrusion detection can also be based on the relationship between the center point of the detection box and the virtual warning area. For example, the position of the center point of the detection box can be calculated based on its coordinates. When the center point of the detection box falls within the warning area defined by the polygon mask, and simultaneously meets the conditions that the intersection-union ratio (IU) is greater than a preset threshold and the recognition confidence level is greater than a preset threshold, an audible and visual alarm is triggered. Through this method, based on suppressing redundant detection boxes, the recognition results output by the target detection network can be combined with the geometric relationship of the preset virtual warning area to achieve automatic identification and real-time alarm of intrusion behavior in key areas of the industrial site.

[0107] Optionally, the method can also acquire multi-frame image recognition results within a preset time window, the multi-frame image recognition results including the coordinates of the recognition box in each frame and the corresponding recognition confidence; based on the changes in the coordinates of the recognition box and the changes in the recognition confidence in the multi-frame image recognition results, calculate a stability index to characterize the stability of the recognition results; dynamically adjust the cross-union ratio threshold and the recognition confidence threshold based on the stability index, wherein when the stability index meets a first preset condition, the cross-union ratio threshold and / or the recognition confidence threshold are reduced, and when the stability index meets a second preset condition, the cross-union ratio threshold and / or the recognition confidence threshold are increased.

[0108] In this embodiment of the invention, the cross-union ratio threshold and the recognition confidence threshold used for alarm judgment can be dynamically adjusted by combining the recognition results of multiple frames of images within a preset time window, so as to adapt to environmental changes and fluctuations in detection results in industrial sites.

[0109] Specifically, during system operation, multiple frames of image recognition results within a preset time window can be continuously acquired, such as the coordinates of the recognition boxes and the corresponding recognition confidence scores of the most recent frames, and the multi-frame recognition results can be analyzed as time series data. In each frame, a recognition box associated with the warning area can be selected, and its coordinate information (such as the position, width, and height of the center point of the recognition box) and the recognition confidence score corresponding to the recognition box can be recorded, and the above data can be accumulated and stored in the time dimension.

[0110] After obtaining multiple frames of recognition results within a preset time window, the changes in the bounding box coordinates and the recognition confidence can be statistically analyzed to construct a stability index characterizing the stability of the recognition results. For example, for bounding box coordinates, the offset of the center point of the bounding box between consecutive frames, the magnitude of changes in the width and height of the bounding box, or the variance of coordinate changes can be calculated to measure the stability of the target position over time. For recognition confidence, the mean, variance, or fluctuation range of the recognition confidence in each frame can be statistically analyzed to reflect the stability of the network's recognition results for that target. If the change in the bounding box position is small and the fluctuation in recognition confidence is small within the preset time window, the recognition result can be considered relatively stable; conversely, if the bounding box position jumps significantly between multiple frames or the fluctuation in recognition confidence is large, the recognition result can be considered unstable. Based on the above changes in position and confidence, a comprehensive stability index can be constructed, for example, by combining the coordinate changes and confidence fluctuations with certain weights to quantitatively evaluate the stability of the recognition results.

[0111] After calculating the stability index, the cross-union ratio (CUNR) threshold and identification confidence threshold used for intrusion detection can be dynamically adjusted based on the value of the stability index. When the stability index meets the first preset condition, such as the center point offset of the identification box in multiple consecutive frames being less than a preset offset threshold and the fluctuation range of the identification confidence being lower than a preset fluctuation threshold, the current identification result can be determined to be in a stable state. At this time, in order to improve the sensitivity of the alarm response, the CUNR threshold and / or identification confidence threshold can be appropriately reduced, so that the system can trigger the alarm earlier when the target continues to approach or stay near the warning area, thereby improving the response speed to potential risks. When the stability index meets the second preset condition, such as the identification box position fluctuating greatly between multiple frames or the identification confidence changing drastically over time, the identification result can be determined to have a large degree of uncertainty. At this time, in order to reduce the false alarm rate, the CUNR threshold and / or identification confidence threshold can be appropriately increased, so that the system only triggers the alarm when the overlap is high enough and the identification confidence is stable enough, thereby suppressing false alarms caused by noise, reflection interference, or instantaneous false detection in complex environments.

[0112] In one possible embodiment, the stability index can be designed in a hierarchical or interval format. For example, the stability index can be divided into three intervals: "high stability," "medium stability," and "low stability." When the stability index falls into the "high stability" interval, a relatively low set of cross-union ratio (CUI) and recognition confidence thresholds is selected; when the stability index falls into the "low stability" interval, a relatively high set of thresholds is selected; and when the stability index is in the intermediate interval, the default thresholds remain unchanged. Through an adaptive threshold adjustment mechanism based on multi-frame recognition results, this invention improves the overall robustness of the system while balancing the timeliness and accuracy of alarms under different operating conditions, reducing false alarms or missed alarms caused by single-frame anomalies, making it more suitable for long-term continuous monitoring scenarios deployed in industrial sites.

[0113] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0114] In one embodiment, an image recognition device is provided, which corresponds one-to-one with the image recognition methods described in the above embodiments. For example... Figure 2 As shown, the image recognition device includes a first acquisition module 201, a feature extraction module 202, a feature fusion module 203, a feature enhancement module 204, and a target detection module 205. Detailed descriptions of each functional module are as follows:

[0115] The first acquisition module 201 is used to acquire the amplitude map to be identified and the corresponding depth map;

[0116] The feature extraction module 202 is used to perform feature extraction processing based on the amplitude map and the depth map to obtain the amplitude features of the amplitude map and the depth features of the depth map;

[0117] The feature fusion module 203 is used to perform feature fusion processing on the amplitude feature and the depth feature based on the channel saliency of the amplitude feature and the depth feature to obtain fused features;

[0118] Feature enhancement module 204 is used to perform attention enhancement processing based on the fused features to obtain enhanced fused features;

[0119] The target detection module 205 is used to provide the enhanced fusion features to a preset target detection network to obtain the corresponding image recognition result.

[0120] Optionally, the first acquisition module 201 is further configured to:

[0121] Obtain the initial amplitude map and initial depth map acquired by the target depth camera;

[0122] The pixel values ​​of each pixel in the initial amplitude map and the initial depth map are mapped to a preset numerical range to obtain the amplitude map and the depth map to be aligned.

[0123] Alignment processing is performed on the amplitude map to be aligned and the depth map to be aligned to obtain the amplitude map to be identified and the corresponding depth map.

[0124] Optionally, the feature fusion module 203 is further configured to:

[0125] Based on the channel saliency, a first fusion weight for the amplitude feature and a second fusion weight for the depth feature are generated;

[0126] The amplitude feature and the depth feature are weighted and fused according to the first fusion weight and the second fusion weight to obtain the fused feature.

[0127] Optionally, the feature fusion module 203 is further configured to:

[0128] Based on the channel saliency, a first weight score is calculated to characterize the importance of the amplitude feature and a second weight score is calculated to characterize the importance of the depth feature;

[0129] The first fusion weight is calculated based on the first weight score and the second weight score;

[0130] The second fusion weight is calculated based on the first fusion weight.

[0131] Optionally, the feature enhancement module 204 is further configured to:

[0132] Channel attention processing is performed based on the fusion features to obtain the first enhanced feature;

[0133] Spatial attention processing is performed based on the fusion features to obtain the second enhanced feature;

[0134] The first enhancement feature and the second enhancement feature are multiplied element-wise to obtain the enhanced fusion feature.

[0135] Optionally, the image recognition result includes the coordinates of the recognition box and the recognition confidence score, and the device further includes:

[0136] The first calculation module is used to calculate the intersection-over-union ratio (IoU) between the recognition box and the preset warning area based on the coordinates.

[0137] The first determining module is used to determine that an intrusion exists in the warning area when the cross-union ratio is not less than a preset cross-union ratio threshold and the identification confidence level is not less than a preset confidence threshold.

[0138] Optionally, the device further includes:

[0139] The second acquisition module is used to acquire the recognition results of multiple frames within a preset time window. The recognition results of multiple frames include the coordinates of the recognition box of each frame and the corresponding recognition confidence.

[0140] The second calculation module is used to calculate a stability index to characterize the stability of the recognition results based on the changes in the coordinates of the recognition box and the changes in the recognition confidence in the multi-frame image recognition results.

[0141] An adjustment module is used to dynamically adjust the cross-union ratio threshold and the identification confidence threshold based on the stability index. When the stability index meets a first preset condition, the cross-union ratio threshold and / or the identification confidence threshold are decreased. When the stability index meets a second preset condition, the cross-union ratio threshold and / or the identification confidence threshold are increased.

[0142] Each module in the aforementioned image recognition device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0143] In one embodiment, a computer device is provided, which may be a terminal device, and its internal structure diagram may be as follows: Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a readable storage medium storing computer-readable instructions. The network interface communicates with external terminals via a network connection. When executed by the processor, the computer-readable instructions implement an image recognition method. The readable storage medium provided in this embodiment includes both non-volatile and volatile readable storage media.

[0144] In this application embodiment, a computer device is provided, including a memory, a processor, and computer-readable instructions stored in the memory and executable on the processor. When the processor executes the computer-readable instructions, it implements the steps of the image recognition method described above.

[0145] In one embodiment of the application, a readable storage medium is provided, which stores computer-readable instructions. When the computer-readable instructions are executed by a processor, they implement the steps of the image recognition method described above.

[0146] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware with computer-readable instructions. These computer-readable instructions can be stored in a non-volatile readable storage medium or a volatile readable storage medium. When executed, these computer-readable instructions can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0147] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0148] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An image recognition method characterized by, The method comprises: obtaining an amplitude map to be identified and a corresponding depth map; based on the amplitude map and the depth map, feature extraction processing is performed to obtain amplitude features of the amplitude map and depth features of the depth map; based on the channel saliency of the amplitude features and the depth features, the amplitude features and the depth features are subjected to feature fusion processing to obtain fusion features; based on the fusion features, attention enhancement processing is performed to obtain enhanced fusion features; the enhanced fusion features are provided to a preset target detection network to obtain a corresponding image recognition result; the method comprises: obtaining an initial amplitude map and an initial depth map collected by a target depth camera; mapping the pixel values of each pixel in the initial amplitude map and the initial depth map to a preset numerical interval to obtain an amplitude map to be aligned and a depth map to be aligned, the preset numerical interval being the [0, 1] interval; based on the amplitude map to be aligned and the depth map to be aligned, alignment processing is performed to obtain the amplitude map to be identified and the corresponding depth map; based on the channel saliency of the amplitude features and the depth features, the amplitude features and the depth features are subjected to feature fusion processing to obtain fusion features, comprising: According to the channel saliency, a first fusion weight of the amplitude feature and a second fusion weight of the depth feature are generated, and the amplitude feature and the depth feature are input to a weight generation module and the depth feature The values of α and β are dynamically obtained by converging and nonlinearly transforming the feature channels. The weight generation module adopts a lightweight neural network structure, including one or more layers of fully connected layers, per-channel convolution layers or gate units, to learn the dynamic generation law of the fusion weight from the input features. The amplitude feature and the depth feature are weighted and fused according to the first fusion weight and the second fusion weight to obtain the fusion feature, that is wherein, is a magnitude feature, is a depth feature, and a and β are the first fusion weight and the second fusion weight described above; the first fusion weight of the amplitude features and the second fusion weight of the depth features are generated according to the channel saliency, comprising: based on the channel saliency, a first weight score representing the importance of the amplitude features and a second weight score representing the importance of the depth features are calculated; based on the first weight score and the second weight score, the first fusion weight is calculated; based on the first fusion weight, the second fusion weight is calculated.

2. The image recognition method of claim 1, wherein, based on the fusion features, attention enhancement processing is performed to obtain enhanced fusion features, comprising: based on the fusion features, channel attention processing is performed to obtain a first enhanced feature; based on the fusion features, spatial attention processing is performed to obtain a second enhanced feature; the first enhanced feature and the second enhanced feature are subjected to element-by-element multiplication to obtain the enhanced fusion features.

3. The image recognition method according to any one of claims 1 to 2, characterized in that, The image recognition result comprises the coordinates of the recognition box and the recognition confidence. After the enhanced fusion features are provided to the preset target detection network to obtain the corresponding image recognition result, the method further comprises: based on the coordinates, the intersection over union of the recognition box and a preset warning area is calculated; when the intersection over union is not less than a preset intersection over union threshold and the recognition confidence is not less than a preset confidence threshold, it is determined that the warning area is invaded.

4. The image recognition method of claim 3, wherein, The method further comprises: obtaining a plurality of image recognition results in a preset time window, the plurality of image recognition results comprising the coordinates of the recognition box and the corresponding recognition confidence of each frame; based on the changes in the recognition box coordinates and the recognition confidence in the plurality of image recognition results, a stability index representing the stability of the recognition result is calculated; The intersection over union threshold and the recognition confidence threshold are dynamically adjusted based on the stability index, wherein the intersection over union threshold and / or the recognition confidence threshold are reduced when the stability index meets a first preset condition, and the intersection over union threshold and / or the recognition confidence threshold are increased when the stability index meets a second preset condition.

5. An image recognition apparatus characterized by comprising: The device comprises: The first acquisition module is configured to acquire an amplitude map to be recognized and a corresponding depth map; The feature extraction module is configured to perform feature extraction processing based on the amplitude map and the depth map to obtain amplitude features of the amplitude map and depth features of the depth map; The feature fusion module is configured to perform feature fusion processing on the amplitude features and the depth features based on channel saliency of the amplitude features and the depth features to obtain fusion features; The feature enhancement module is configured to perform attention enhancement processing based on the fusion features to obtain enhanced fusion features; The target detection module is configured to provide the enhanced fusion features to a preset target detection network to obtain a corresponding image recognition result; The first acquisition module 201 is further configured to: acquire an initial amplitude map and an initial depth map collected by a target depth camera; map pixel values of each pixel in the initial amplitude map and the initial depth map to a preset numerical interval to obtain an amplitude map to be aligned and a depth map to be aligned, the preset numerical interval being an interval [0, 1]; perform alignment processing based on the amplitude map to be aligned and the depth map to be aligned to obtain the amplitude map to be recognized and the corresponding depth map; The feature fusion module 203 is further configured to: According to the channel saliency, a first fusion weight of the amplitude feature and a second fusion weight of the depth feature are generated, and the amplitude feature and the depth feature are input to a weight generation module The weight generation module is configured to dynamically obtain the values of the first fusion weight and the second fusion weight by converging the amplitude feature and the depth feature and performing nonlinear transformation on the amplitude feature and the depth feature The weight generation module adopts a lightweight neural network structure including one or more full connection layers, per-channel convolution layers or gate units to learn a dynamic generation rule of the fusion weight from the input features. The amplitude feature and the depth feature are weighted and fused according to the first fusion weight and the second fusion weight to obtain the fusion feature, that is wherein, is a magnitude feature, is a depth feature, and a and β are the first fusion weight and the second fusion weight described above; The feature fusion module 203 is further configured to: based on the channel saliency, calculate a first weight score for representing an importance degree of the amplitude features and a second weight score for representing an importance degree of the depth features; based on the first weight score and the second weight score, calculate the first fusion weight; based on the first fusion weight, calculate the second fusion weight. 6.A computer device, comprising a memory, a processor, and computer readable instructions stored on the memory and running on the processor, characterized in that, The processor executes the computer readable instructions to implement the image recognition method of any one of claims 1 to 4.

7. A readable storage medium, having stored thereon computer readable instructions, characterized in that, The computer readable instructions are executed by the processor to implement the image recognition method of any one of claims 1 to 4.

Images