A convolutional neural network optimization method based on autonomous learning

Abstract

The application relates to the technical field of computer vision and artificial intelligence, in particular to a convolutional neural network optimization method based on autonomous learning, which comprises the following steps: acquiring original image data; inputting the original image data into a convolutional neural network for forward propagation; recording feature maps extracted from adjacent layers during the forward propagation; inputting second feature maps into a feedback prediction channel to output predicted feature maps; calculating the difference between the first feature maps and the predicted feature maps to generate error signals; inputting the error signals into a modulator to output modulated feature maps; replacing the original first feature maps with the modulated feature maps and re-performing forward propagation calculation from the first layer; and using the recognition result output by the final network for decision control of an intelligent driving system. The application effectively offsets the influence of input disturbance or interlayer information deviation on feature expression, so that the features output by the network are more stable.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer vision and artificial intelligence, and in particular to a convolutional neural network optimization method based on autonomous learning. BACKGROUND

[0002] Convolutional Neural Network (CNN) is one of the core models in the field of deep learning, and is widely used due to its excellent performance in computer vision tasks such as image recognition and object detection. Classical CNN models such as AlexNet, VGGNet, GoogLeNet and ResNet usually adopt a hierarchical feedforward structure. In this structure, the input data (such as images) pass through convolutional layers, activation function layers (such as ReLU) and pooling layers in turn, and the information flows from bottom to top in a one-way manner. With the increase of the number of layers, the extracted features are gradually transformed from low-level details such as edges and textures into high-level abstract semantic information.

[0003] The advantage of this feedforward structure is that it can efficiently learn feature representations with translation invariance from images through the mechanisms of "local receptive field" and "parameter sharing". However, this structure also has inherent limitations. First, the one-way nature of information flow means that high-level semantic features are difficult to effectively feedback and guide the optimization of low-level features. When the feature map undergoes pooling and other downsampling operations, some spatial detail information will inevitably be lost, and the subsequent layers cannot recover these lost details. Second, in complex real-world application scenarios, such as the visual perception system of intelligent driving, the model often needs to deal with image quality degradation problems caused by rain, night low light, target occlusion or motion blur. When facing these input disturbances, the features extracted by the existing feedforward CNN may already have deviations, which will be amplified layer by layer during forward propagation, eventually leading to errors in high-level semantic judgment, such as missing detection or misidentification of pedestrians and vehicles. SUMMARY

[0004] The purpose of the present application is to effectively offset the influence of input disturbances (such as noise and illumination changes) or interlayer information deviation on feature expression by allowing CNN specific layers to receive high-level feedback information and compare it with their own features, adjust errors, make the network output features more stable, reduce the risk of feature distortion caused by external interference, significantly enhance the overall robustness, and improve the stability and anti-interference ability of feature extraction.

[0005] In a first aspect, an embodiment of the present application provides a convolutional neural network optimization method based on autonomous learning, which comprises:

[0006] Obtaining original image data containing pedestrians, vehicles or traffic signs collected by a vehicle-mounted camera;

[0007] The original image data is input into a convolutional neural network for forward propagation. The network contains multiple sequentially connected layers, wherein the spatial resolution of the feature map decreases with each layer as the depth increases.

[0008] During the forward propagation process, feature maps extracted from adjacent layers are recorded, wherein the feature maps extracted from adjacent layers include those from the previous layer with higher spatial resolution. The first feature map, and from the immediately following hierarchy Then, the next level with lower spatial resolution +1 second feature map;

[0009] The second feature map is input into a feedback prediction channel, which includes an upsampling layer and a learnable network module for detail enhancement. Through the transformation of the feedback prediction channel, the output is similar in spatial size and number of channels to the first feature map. Figure One A consistent predictive feature map that is enhanced in spatial detail;

[0010] Calculate the difference between the first feature map and the predicted feature map to generate an error signal;

[0011] The error signal is input to a modulator, which generates a modulated signal based on the error signal, adaptively adjusts the first feature map, and outputs a modulated feature map.

[0012] The original first feature map is replaced with the modulated feature map, and the process is repeated from the first feature map. Layered forward propagation calculation;

[0013] The final network output recognition results are used for decision-making and control in the intelligent driving system.

[0014] Optionally, the learnable network module for detail enhancement is a feature fusion module;

[0015] The feedback prediction channel generates the prediction feature map through the following steps:

[0016] The second feature map is upsampled to initially align its spatial size and number of channels with the first feature map.

[0017] The upsampled features are compared with those from the convolutional neural network. The detailed features from the shallower hierarchical jumps are then fused together;

[0018] The fused features are processed through at least one convolutional layer to output the predicted feature map.

[0019] Optionally, the learnable network module for detail enhancement is an attention mechanism module.

[0020] The feedback prediction channel generates the predicted feature map by the following steps:

[0021] The second feature map is upsampled to preliminarily align its spatial size and channel number with the first feature map;

[0022] The upsampled feature is input into the attention mechanism module to generate a feature weight for enhancing key details;

[0023] The upsampled feature is weighted by the feature weight, and the predicted feature map is output.

[0024] Optionally, the modulator generates a modulation signal according to the error signal and adaptively adjusts the first feature map, and the specific implementation is as follows:

[0025] Channel-level modulation, including: generating a set of channel scaling coefficients and channel bias coefficients according to the error signal, and scaling and biasing each channel of the first feature map;

[0026] Or,

[0027] Feature map-level modulation, including: generating a feature correction matrix with the same shape as the first feature map according to the error signal, and adding the matrix to the first feature map;

[0028] Or,

[0029] Convolution kernel-level modulation, including: generating a convolution kernel adjustment amount according to the error signal, adjusting the weight parameters of the first layer convolution kernel, and recalculating the first feature map using the adjusted convolution kernel.

[0030] Optionally, based on the computing power of the deployment device, the implementation of the modulator is dynamically selected, including:

[0031] When the computing efficiency is prioritized, the channel-level modulation with the lightest computing load is selected;

[0032] When the spatial details of the feature map need to be finely corrected, the feature map-level modulation with a higher computing load than the channel-level modulation but lower than the convolution kernel-level modulation is selected;

[0033] When the network core parameters need to be optimized for a long time and the computing resources allow, the convolution kernel-level modulation with the heaviest computing load is selected.

[0034] Optionally, the convolutional neural network is configured to include multiple feedback self-learning structures. ​

[0035] wherein each of the feedback self-learning structures is established based on a hierarchical pair including a preceding hierarchical layer with a higher spatial resolution and a succeeding hierarchical layer +1 with a lower spatial resolution following the preceding hierarchical layer +1;

[0036] The plurality of feedback self-learning structures collectively form a multi-layer feedback self-learning system.

[0037] Optionally, the preceding hierarchical layer with the higher spatial resolution is a layer in the convolutional neural network in which the spatial resolution of the first output feature map is lower than the input image resolution.

[0038] In a second aspect, an embodiment of the present application provides a convolutional neural network optimization device based on autonomous learning, the device comprising:

[0039] an image data acquisition module configured to acquire original image data containing pedestrians, vehicles or traffic signs collected by a vehicle-mounted camera;

[0040] an image data input module configured to input the original image data to a convolutional neural network for forward propagation, the network containing a plurality of sequentially connected hierarchical layers, wherein the spatial resolution of a feature map decreases layer by layer as the hierarchical layers deepen;

[0041] a feature map recording module configured to record feature maps extracted from adjacent hierarchical layers during the forward propagation, wherein the feature maps extracted from adjacent hierarchical layers include a first feature map from a preceding hierarchical layer with a higher spatial resolution and a second feature map from a succeeding hierarchical layer +1 with a lower spatial resolution

[0042] a predicted feature map generation module configured to input the second feature map to a feedback prediction channel, the feedback prediction channel containing an up-sampling layer and a learnable network module for detail enhancement, and output a predicted feature map enhanced in spatial details and having the same spatial size and channel number as the first feature map through transformation of the feedback prediction channel; Figure One

[0043] an error signal calculation module configured to calculate the difference between the first feature map and the predicted feature map to generate an error signal;

[0044] ​The modulated feature map generation module is configured to input the error signal into a modulator, and the modulator generates a modulation signal according to the error signal and adaptively adjusts the first feature map, and outputs a modulated feature map.

[0045] The feature map adjustment module is configured to replace the original first feature map with the modulated feature map, and re-perform forward propagation calculation from the first layer.

[0046] The recognition result output module is configured to use the recognition result output by the final network as a decision control of the intelligent driving system.

[0047] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0048] at least one processor;

[0049] a memory for storing instructions executable by the at least one processor;

[0050] The at least one processor is configured to execute the instructions to implement the method of the first aspect.

[0051] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium having a computer program stored thereon, and the program is executed by a processor to implement the method of the first aspect.

[0052] The embodiment of the present application introduces an inter-layer feedback self-learning mechanism composed of a feedback prediction channel and a modulator in the network, and constructs a multi-level, closed-loop feedback correction system. The mechanism can use the abstract semantic information of the high layer as a guide signal, reconstruct the bottom layer features in detail through the learnable feedback prediction channel, and dynamically correct the deviation in the bottom layer feature expression with the help of the flexible modulator. This innovative design effectively overcomes the problems of feature distortion, detail loss and semantic deviation of traditional feedforward convolutional neural networks in complex scenes, significantly improves the feature robustness and detail preservation ability of the model in multiple interference environments such as sudden light changes, motion blurring, and rainy and snowy weather. This method not only effectively reduces the missed detection and false detection of key targets such as pedestrians and vehicles, but also improves the stability of reasoning while maintaining high accuracy, ultimately provides more reliable and accurate recognition results for the perception module of the intelligent driving system, lays a solid technical foundation for vehicle decision and control, and greatly improves the driving safety level. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The overall architecture schematic diagram provided by the embodiment of the present application;

[0054] Figure 2 The schematic diagram of three modulation modes provided by the embodiment of the present application;​

[0055] Figure 3 A flow chart of a convolutional neural network optimization method based on autonomous learning is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0056] The present application will be described in detail below through embodiments.

[0057] Convolutional Neural Networks (CNN) as an important model of deep learning, plays a key role in image recognition, visual analysis and other fields. It stands out with two core designs: one is through the local receptive field mechanism, let the network focus on the local area of the image features, without paying attention to the global redundant information; the second is to adopt the parameter sharing strategy, so that the neurons of the same convolution layer share a set of parameters, greatly simplifying the network size. These two points together create the efficiency and strong feature extraction ability of CNN in processing high-dimensional data such as images and speech.

[0058] The reason why the convolutional neural network can become the core tool in the field of computer vision is that it precisely meets the needs of visual tasks: through the local receptive field, it efficiently captures edge, texture and other detailed features; with parameter sharing, it greatly reduces the computational burden; with translational invariance, it stably identifies targets in different positions; with multi-layer structure, it realizes deep feature mining from pixels to semantics. These capabilities enable it to automatically complete feature engineering without human intervention, easily handle large-scale data, and have outstanding generalization effect. In addition, the architecture is flexible and expandable, and can adapt to various tasks such as image classification, target detection and semantic segmentation, so the application is very wide.

[0059] Traditional CNN is a feedforward structure: information passes through convolution, nonlinear activation, pooling and other levels from the input layer to higher layers. With each layer, the features become more abstract, but also may lose the fine information of the bottom layer (such as spatial details). The deeper the number of layers, the more likely the gradient transmission will be attenuated or exploded (especially for very deep networks), which will make training difficult.

[0060] In core intelligent driving scenarios such as vehicle camera recognition of pedestrians, vehicles, and road signs, this invention improves the stability and robustness of information through optimized network design. Specifically, after receiving "feedback prediction / semantic information" from higher layers, a layer of the CNN compares its current features with its own and calculates errors, adaptively adjusting the layer (activation or parameters) accordingly. This improves representational robustness and detail preservation. This design effectively addresses complex interferences in intelligent driving, such as blurred vision in rainy weather, insufficient lighting at night, and mixed traffic of multiple vehicles and pedestrians. It reduces missed pedestrian detection, misidentification of vehicles, and misinterpretation of signs, while accurately preserving key details, providing more reliable perceptual basis for decision-making and further enhancing driving safety.

[0061] To ensure the plan is clearly described, the following will combine... Figure 1 The overall technical solution of this invention is described in detail. This embodiment takes pedestrian recognition in intelligent driving as an example, but this method is also applicable to the recognition of other targets such as vehicles and traffic signs. Specifically, it may include the following steps:

[0062] 1. Network initialization and data preparation.

[0063] First, a multi-layered convolutional neural network (backbone) is constructed, such as using a ResNet or VGG architecture. This network consists of sequentially connected convolutional layers, activation function layers (such as ReLU), and pooling layers. Then, raw road images captured by an onboard camera are used as input data.

[0064] 2. Select the feedback level pair and perform forward propagation.

[0065] As attached Figure 1 As shown, in this embodiment, we choose to implement a feedback self-learning mechanism on two adjacent layer pairs of the network, i.e., layer pairs ( , +1) and hierarchy pairs ( +1, +2). This multi-layered feedback structure can form a more refined closed-loop correction system.

[0066] Forward propagation to hierarchy : Input image after the first The network processing before and after the layer (usually including convolution → activation → pooling) yields the first feature map output by that layer. This feature map has a high spatial resolution, preserving rich spatial details such as edges and textures.

[0067] Forward propagation to hierarchy +1: First feature map Continue to spread forward, after the first... +1 layer (convolution→activation→pooling) to get the second feature map At this time, the second feature map has a lower spatial resolution but contains more abstract semantic information.

[0068] Forward propagation to level +2: the second feature map continues to be propagated forward and is processed by the +2 layer to get the feature map . It has a lower spatial resolution and the most abstract semantic information.

[0069] 3, feedback prediction process (corresponding to the Figure 1 module in FIG. 2) G

[0070] The core of this step is to use high-level semantic information to predict and enhance low-level features.

[0071] For level pair , +1: input the high-level feature into the feedback prediction channel . The predictor first uses an upsampling layer (such as transpose convolution or bilinear interpolation) to make the spatial size and the number of channels of consistent with . Then, through a learnable network module for detail enhancement (for example, an “upsampling + feature fusion” collaborative structure, an “upsampling + attention mechanism”, etc.), the upsampled feature is fused with the detail feature from the shallower layer of the network, and then a convolution layer is used for fine correction. Finally, the feedback prediction channel outputs a predicted feature map that is completely aligned with in spatial size and channel number, and is more detailed and accurate in details.

[0072] For level pair +1, +2: similarly, input the feature into the feedback prediction channel , and output the predicted feature that is aligned with through the internal upsampling and detail enhancement module.

[0073] 4, error calculation and modulation process (corresponding to the Figure 1 module in FIG. 2) H

[0074] This step realizes adaptive optimization of the layer by calculating the error and modulating the feature.​​​

[0075] At the level Calculate the first feature map With predicted feature map The difference between them generates an error signal. = - This error signal reflects the deviation between high-level semantic predictions and the actual features at the lower level.

[0076] Error signal Input to modulator ( There are typically three modulation methods, which will be described in detail in the following examples. In this example, the second modulation method (feature map-level modulation) will be used as an example. This modulator is a lightweight convolutional subnetwork that will... As input, after nonlinear transformation, the output is a... Feature correction matrix of the same shape .

[0077] Next, feature modulation is performed: the correction matrix is ​​added to the original feature map to obtain the modulated feature map. = + This operation is equivalent to... Targeted enhancements and repairs were performed on areas where there were discrepancies or lost details.

[0078] At the level +1: The process is exactly the same. Calculation = - and through the modulator ( Generate modulated features .

[0079] in, Figure 1 In It refers to three modulation methods. The meaning is a broad concept and can be understood as a modulator, without referring to any specific modulator.

[0080] 5. Re-propagate and output.

[0081] Using the modulated feature map Replace the original feature map Then, from the first The layer begins to re-execute the forward propagation computation, i.e., using the optimized version. To calculate the first The outputs of layer +1 and subsequent layers are used. The recalculated features are propagated and eventually pass through a fully connected layer and a classifier to output the recognition results of targets such as pedestrians and vehicles in the image. This recognition result is then sent to the decision control module of the intelligent driving system to implement subsequent functions such as vehicle warning, planning, or control.

[0082] 6. Training Process

[0083] During end-to-end training of the entire network, the feedback prediction channel G and modulator H The parameters will be optimized along with the main network using the gradient descent algorithm. The loss function typically employs a combination of cross-entropy loss and feature reconstruction loss to simultaneously ensure classification accuracy and feature quality.

[0084] Through the above combination Figure 1 As can be seen from the detailed description, the present invention introduces a feedback self-learning loop consisting of a feedback prediction channel and a modulator at multiple levels, enabling the network to dynamically correct and enhance the expression of low-level features using high-level semantic information, thereby significantly improving the robustness and accuracy of visual perception in complex driving environments.

[0085] In the above Figure 1 Based on the overall architecture shown, one of the core innovations of this invention lies in the modulator. H Flexible design and implementation. For example... Figure 2 As shown, modulator ( For error signals The following three different mechanisms can be used to handle this, in order to adapt to different computing resource constraints and performance requirements.

[0086] After the aforementioned steps, the first [item] has been obtained. Original feature map of the layer And the error signal obtained through feedback prediction and error calculation. The modulator's goal is to utilize... Generate a modulated signal, for Optimize.

[0087] Modulation Mechanism 1: Channel-level Modulation

[0088] This mechanism achieves fast, lightweight adjustments by globally scaling and biasing each channel of the feature map.

[0089] (1) Error information compression: compression of error signals The global average pooling operation is performed along the spatial dimensions to compress the feature map into a C-dimensional vector. This vector serves as a global error descriptor at the channel level, summarizing the average error level of each channel.

[0090] (2) Generating modulation parameters: The C-dimensional error descriptor is input into a small parameter generation network consisting of two fully connected layers. The network outputs two C-dimensional vectors: channel scaling coefficients and channel bias coefficients .

[0091] (3) Applying modulation: The generated and are applied to the original feature map to correct . The calculation process is .

[0092] where represents channel-wise multiplication. This operation is equivalent to adaptively enhancing or suppressing the feature responses of different channels based on the error.

[0093] This method has minimal computational overhead and is very suitable for edge devices with limited computing power. It can quickly correct the intensity bias of the entire feature channel and achieve coarse-grained and efficient adjustment of the feature distribution.

[0094] Modulation mechanism two: feature map level modulation

[0095] This mechanism aims to make more detailed, pixel-level corrections to the spatial details of the feature map.

[0096] (1) Error feature extraction: The error signal is input into a lightweight convolutional subnetwork. The subnetwork is usually composed of 1-2 convolutional layers and is combined with activation functions such as ReLU to introduce nonlinearity.

[0097] (2) Generating correction matrix: The convolutional subnetwork finally outputs a feature correction matrix that is exactly the same in spatial size and number of channels as the original feature map . Each element of the matrix represents the amount of fine adjustment needed at the corresponding spatial position and channel.

[0098] (3) Applying modulation: The generated correction matrix is directly added element-wise to the original feature map to obtain the modulated feature map:

[0099] = +

[0100] The method can preserve spatial information, realize enhancement or repair of details of a specific region in a feature map, and for example, strengthen blurred pedestrian edges. The calculation amount is moderate, and excellent performance can be provided in a scene requiring fine correction.

[0101] Modulation mechanism three: kernel-level modulation

[0102] This mechanism is the most fundamental optimization method, which directly adjusts the core parameters of the network, the convolution kernel weight, to realize long-term and root cause improvement of feature extraction capability.

[0103] (1) Error information compression and coding: first, the error signal is adaptively pooled to reduce its dimension and flatten it into a one-dimensional vector. Then, the vector is encoded through several fully connected layers to capture the complex mapping relationship between the error and the convolution kernel weight adjustment.

[0104] (2) Generation of convolution kernel adjustment amount: the encoded vector passes through a final fully connected layer to generate the error signal , that is, the "convolution kernel adjustment amount", which has the same shape as the original convolution kernel.

[0105] (3) Update parameters and recalculate: add the adjustment amount to the original convolution kernel (or act on the original convolution kernel in other ways) to obtain a new and better convolution kernel.

[0106] This operation is the key difference: it is not to generate a new feature map, but to update the network parameters. Then, the updated convolution kernel must be used to perform convolution operation on the original input of the first layer, so as to calculate a new and modulated feature map .

[0107] This method directly optimizes the feature extractor itself, and the effect is the most persistent and fundamental, which can significantly improve the essential ability of the network. However, it has the largest calculation and memory overhead, and is usually used in a training environment with sufficient computing power or for offline optimization of key network layers.

[0108] The modulator of the present application provides a variety of solutions from light to heavy, from temporary correction to permanent optimization. In practical applications, according to the real-time requirements and precision needs of different modules (such as forward collision warning and traffic sign recognition) of intelligent driving systems, the most suitable modulation mechanism can be flexibly configured for different levels in the network, so as to achieve the best balance between performance and efficiency.

[0109] After a detailed description of the overall technical solution and three modulation methods of the embodiments of the present invention, the following will provide a detailed description of a convolutional neural network optimization method based on autonomous learning provided by the embodiments of the present invention.

[0110] like Figure 3 As shown in the figure, an embodiment of the present invention provides a convolutional neural network optimization method based on autonomous learning, which may include the following steps:

[0111] S310 acquires raw image data containing pedestrians, vehicles, or traffic signs captured by an onboard camera.

[0112] In this step, the system captures real-time images of the road environment using image sensors deployed on the vehicle (such as front-facing cameras and surround-view cameras). This raw image data forms the basic input to the intelligent driving perception system, and its content typically includes complex scene information, such as roads, pedestrians, other vehicles, traffic lights and signs, etc.

[0113] S320 inputs the original image data into the convolutional neural network for forward propagation. The network contains multiple sequentially connected layers, in which the spatial resolution of the feature maps decreases layer by layer as the layer deepens.

[0114] The raw image obtained in step S310 is input into a predefined base convolutional neural network (such as VGG, ResNet, or other backbone networks). This network consists of multiple hierarchical modules connected sequentially, each typically containing convolution, non-linear activation (such as ReLU), and pooling operations. As the data propagates forward through the network layer by layer, the spatial dimensions (height and width) of the feature maps are gradually reduced through pooling or stride convolution, while the number of feature channels increases, and the semantic abstraction level continuously improves.

[0115] S330, during the forward propagation process, records feature maps extracted from adjacent layers. These feature maps include those from the previous layer with higher spatial resolution. The first feature map, and from the immediately following hierarchy Then, the next level with lower spatial resolution +1 is the second feature map.

[0116] During the forward propagation of the network, the system selects one or more adjacent hierarchical pairs ( , +1), and cache the feature maps from the outputs of these two levels. Specifically:

[0117] First feature map From the hierarchy Output. This level is usually located in the middle or lower part of the network, and its output feature maps have high spatial resolution, preserving rich spatial details such as pedestrian contours, vehicle edges, textures, etc.

[0118] Second feature map : the immediately preceding, deeper level +1 output. Its spatial resolution is lower than that of the preceding level , due to further pooling or convolutional down-sampling, but contains higher-level, more abstract semantic information (e.g., concepts such as "part of a vehicle", "human form", etc.).

[0119] As an implementation of an embodiment of the present application, the preceding level , whose output feature maps have a spatial resolution lower than that of the input image, is the first level in the convolutional neural network whose output feature maps have a spatial resolution lower than that of the input image. This definition provides an objective, clear and efficient starting point for the core feedback mechanism. The determination method is described in detail below in conjunction with the network structure.

[0120] 1. Network structure analysis.

[0121] Consider a typical convolutional neural network structure, whose initial part is usually as follows:

[0122] Input image → Convolutional layer 1 (Conv1) → Activation function (ReLU) → Pooling layer (Pool1) → Convolutional layer 2 (Conv2)

[0123] Assume that the spatial resolution of the input image is [H, W] (e.g., [224, 224]). Convolutional layer 1 (Conv1) usually uses a small stride (e.g., stride = 1) and padding (padding = 1), and the size of its output feature maps usually remains unchanged or changes very little, and the spatial resolution is still considered to be at the [H, W] level. Pooling layer 1 (Pool1) is a typical spatial down-sampling operation. Taking a max pooling with a stride of 2 as an example, it halves the size of the input feature maps in both the width and height dimensions. Therefore, the spatial resolution of its output feature maps is explicitly reduced to [H / 2, W / 2].

[0124] 2. Determination of the level .

[0125] According to the above definition, the level whose output feature maps have a spatial resolution lower than that of the input image is the level immediately following the first spatial down-sampling operation. In this example, pooling layer 1 (Pool1) is the first operation that performs spatial down-sampling, which reduces the feature map resolution from [H, W] to [H / 2, W / 2]. Therefore, the convolutional layer 2 (Conv2) that follows it is the preceding level And the level after the convolution layer 2 (for example, the immediately subsequent pooling layer 2 or convolution layer 3) naturally becomes the next level +1.

[0126] Selecting this level as the feedback starting point has the following significant benefits:

[0127] 1. The best balance between details and semantics is achieved. The feature map of this level (such as Conv2) has just undergone the first downsampling and still retains very rich spatial detail information, much higher than the features in deeper layers of the network. At the same time, it has been processed by a layer of convolution and nonlinear activation and has higher-level semantic information than the original pixels. The introduction of feedback enables the high-level information to most effectively act on these "semi-processed" features with remaining details, achieving accurate correction.

[0128] 2. An objective and unambiguous execution standard is provided. This definition avoids ambiguous expressions such as "low-middle layer". For any given CNN architecture, a person skilled in the art can uniquely and accurately determine the position of the level by simply analyzing the first change in the size of the feature map during forward propagation, ensuring the repeatability of the scheme and the stability of the patent.

[0129] 3. The efficiency of the feedback mechanism is ensured. Starting from the front of the network, the feedback can correct feature deviations as early as possible to prevent errors from being propagated and amplified in the deep network. At the same time, since the resolution of this layer has not been reduced for many times, the cost of calculating the feedback prediction channel G and the modulator H is relatively controllable, achieving an optimal balance between performance and overhead.

[0130] This embodiment objectively defines the level as the level whose resolution first decreases in the network, providing a starting point for the entire feedback self-learning system that is significantly effective and easy to implement, which is one of the key designs for the efficient operation of the present application.

[0131] S340, input the second feature map into a feedback prediction channel, the feedback prediction channel comprising an upsampling layer and a learnable network module for detail enhancement, and output a predicted feature map that is enhanced in spatial details and has the same spatial size and channel number as the first feature Figure One map through the transformation of the feedback prediction channel.

[0132] This step is the core of introducing feedback information. The second feature map from the high level is input into a specially designed feedback prediction channel . The predictor is not a simple upsampler, and its structure includes:

[0133] 1. Upsampling layer. Used to increase the resolution of the second feature map to the same size as the first feature map. 2. Learnable network module for detail enhancement. spatial dimension is enlarged, so that it is preliminarily aligned with the size of the first feature map and has the same number of channels as .

[0134] 2. A learnable network module for detail enhancement. This module refines the rough features after upsampling. Its implementation can be a "upsampling + feature fusion" structure (i.e., fusing the details from the skip connections of the shallower layers of the network), or a "upsampling + attention mechanism" structure (i.e., adaptively enhancing the details of the key regions). Finally, the predictor outputs a predicted feature map which not only has the same size and channel as , but also reconstructs and enhances the spatial details under the guidance of semantics.

[0135] S350, calculate the difference between the first feature map and the predicted feature map to generate an error signal.

[0136] Compare the ideal predicted feature generated by the feedback prediction channel with the first feature actually extracted by the network. By calculating the difference between the two (e.g. = - or = - ), an error signal is generated. This signal quantifies the deviation between the current network bottom feature extraction result and the high-level semantic expectation, directly reflecting the loss of details or distortion of features caused by image quality degradation or limitations of the model itself.

[0137] S360, input the error signal to a modulator, which generates a modulation signal according to the error signal and adaptively adjusts the first feature map to output a modulated feature map.

[0138] The error signal is sent to a modulator ( ). The role of the modulator is to convert the error signal into specific feature correction instructions. Depending on the different requirements for computational efficiency and correction granularity in different application scenarios, one of the three mechanisms can be used.

[0139] Mechanism one, channel-level modulation, includes: generating a set of channel scaling coefficients and channel bias coefficients according to the error signal, and scaling and biasing each channel of the first feature map.

[0140] Specifically, a set of channel-level scaling and biasing parameters is generated according to to scale and bias Perform overall calibration.

[0141] Mechanism two, feature map level modulation, includes: generating a feature correction matrix with the same shape as the first feature map according to the error signal, and adding the matrix to the first feature map.

[0142] Specifically, according to generate a fine correction matrix , and point by point to repair spatial details.

[0143] Mechanism three, kernel level modulation, includes: generating a kernel adjustment amount according to the error signal, adjusting the weight parameters of the first layer kernel, and recalculating the first feature map using the adjusted kernel.

[0144] Specifically, according to dynamically adjust the kernel weight of the first layer, and recalculate the feature from the root. By performing the above adaptive adjustment on the first feature map , the modulator outputs an optimized and modulated feature map .

[0145] In an implementation, based on the computing power of the deployment device, dynamically selecting the implementation of the modulator can include the following steps:

[0146] When the computing efficiency is prioritized, the channel level modulation with the lightest computing load is selected;

[0147] When the spatial details of the feature map need to be finely corrected, the feature map level modulation with a higher computing load than the channel level modulation but lower than the kernel level modulation is selected;

[0148] When the network core parameters need to be optimized for a long time and the computing resources allow, the kernel level modulation with the heaviest computing load is selected.

[0149] In this implementation, first, according to the real-time requirements, precision requirements of different functional modules of the intelligent driving system, and the computing resources of the deployment hardware, a modulation strategy is formulated.

[0150] For embedded platforms with extremely limited computing resources (such as microprocessors used for basic alarm functions), the computing efficiency is prioritized, and the system defaults to the channel level modulation with the lightest computing load. This way, modulation parameters are generated through global pooling and a lightweight fully connected layer to realize fast calibration of feature channels with minimal computing overhead, ensuring stable operation of the system under resource constraints.

[0151] ​For a vehicle-mounted intelligent computing unit with medium computing power (such as a domain controller for panoramic perception), the task requires fine correction of the spatial details of the feature map to improve the accuracy of positioning and contour. At this time, the system selects feature map level modulation with higher computational load than channel level modulation but significantly lower than convolution kernel level modulation. This method generates a spatial detail correction matrix through a lightweight convolution subnetwork, which can enhance specific regions of the feature map at the pixel level, effectively handling partially occluded or blurred targets.

[0152] For a cloud server with sufficient computing power or when performing offline training of the model, the goal is not only to process the current data, but also to optimize the core parameters of the network for a long time to improve the intrinsic ability of the model. In this case, when the computing resources allow, the system selects the convolution kernel level modulation with the heaviest computational load. This method optimizes the feature extraction ability from the root by dynamically adjusting the convolution kernel weights, thereby realizing the continuous evolution of the model performance, especially suitable for scene-specific optimization of pre-trained models.

[0153] In actual operation, the system is built-in a resource scheduler. The scheduler evaluates the available computing resources and the current task requirements in real time or according to the preset configuration.

[0154] When the system starts a real-time perception task with low latency and high throughput, the resource scheduler will identify the need for "priority to computing efficiency" and assign channel level modulation to the relevant network layers.

[0155] When the system processes critical scenes (such as passing through complex intersections) and needs to generate high-precision target segmentation maps, the scheduler will identify the need for "fine correction of spatial details", even if the current computing power is not the most abundant, it will enable feature map level modulation for related layers to ensure the output quality.

[0156] When the vehicle is parked or the network connection is good, the system can upload part of the data to the cloud. The cloud server uses its abundant computing resources to perform incremental training on the received difficult example samples using convolution kernel level modulation, realizes the "long-term optimization" of the core parameters of the network, and downloads the optimized model parameters to the vehicle end, completing the ability iteration of the entire system.

[0157] The above dynamic selection mechanism enables the present application to flexibly adapt to various hardware platforms from low-end to high-end, achieving "on-demand allocation" under different resource constraints. It not only guarantees the basic functions and real-time performance in resource-constrained environments, but also can exert the best performance of the model under appropriate conditions, realizing the efficient unification of computing efficiency and recognition accuracy, greatly improving the practical value and deployment flexibility of the recognition method in industrial practice.

[0158] S370, using the modulated feature map to replace the original first feature map, and re-performing self Layer-wise forward propagation calculation.

[0159] The higher-quality features obtained by step S360 Replace the original Then, the network re-executes the forward propagation from the first layer. This means that the subsequent layers (Layer +1, +2,...) will perform secondary calculations based on this corrected and enhanced feature, thereby obtaining more accurate high-level semantic features.

[0160] S380, the recognition result output by the final network is used for decision control of the intelligent driving system.

[0161] After the above processing including the feedback self-learning process, the accuracy and robustness of the classification or detection result (such as "there is a pedestrian 10 meters ahead", "there is a vehicle on the left side") output by the network are significantly improved. The result is transmitted to the decision and control center of the intelligent driving system as the control basis for key decisions such as vehicle path planning, collision warning, adaptive cruise, etc., thereby effectively improving driving safety in complex real road conditions.

[0162] The embodiment of the application introduces an inter-layer feedback self-learning mechanism composed of a feedback prediction channel and a modulator into the network, and constructs a multi-level, closed-loop feedback correction system. The mechanism can use high-level abstract semantic information as a guide signal, reconstruct the bottom layer features in detail through a learnable feedback prediction channel, and dynamically correct the bias in the bottom layer feature expression with the help of a flexible modulator. This innovative design effectively overcomes the problems of feature distortion, detail loss and semantic bias faced by traditional feedforward convolutional neural networks in complex scenes, significantly improving the feature robustness and detail preservation ability of the model in multiple interference environments such as sudden changes in light, motion blur, and rainy and snowy weather. This method not only effectively reduces the missed detection and false detection of key targets such as pedestrians and vehicles, but also improves the stability of reasoning while maintaining high accuracy, ultimately providing a more reliable and accurate recognition result for the perception module of the intelligent driving system, laying a solid technical foundation for vehicle decision and control, and greatly improving driving safety.

[0163] Based on the embodiment shown in Figure 3 As an implementation manner of the embodiment of the application, the learnable network module for detail enhancement is a feature fusion module;

[0164] The feedback prediction channel generates the predicted feature map through the following steps a1 to a3:

[0165] Step a1, up-sample the second feature map to preliminarily align its spatial size and channel number with the first feature map.

[0166] Will come from high level +1 Second Feature Map Input upsampling layer. Because... Having undergone preceding pooling or strided convolution operations, its spatial size is smaller than that of the first feature map. This step restores its spatial dimensions to a similar level using upsampling operations such as bilinear interpolation or transposed convolution. Same scale, and make its number of channels the same as With the same number of channels, preliminary alignment of spatial dimensions is achieved, laying the foundation for subsequent feature fusion.

[0167] Step a2, compare the upsampled features with those from the convolutional neural network. The detailed features that are connected by shallower layers are then fused together.

[0168] The upsampled features are compared with those obtained from the network through the skip connection mechanism. The original detailed features from shallower layers are fused. These shallower features (e.g., those from the first layer) -1 or The -2 layer has higher spatial resolution and retains rich details such as edges and textures, but its semantic level is lower. This step uses channel concatenation or element-wise addition to complementarily fuse the upsampling results of high-level features (rich in semantic information) with shallow features (rich in detail information), effectively solving the problem of detail blurring caused by simple upsampling.

[0169] Step a3: Process the fused features through at least one convolutional layer to output a predicted feature map.

[0170] The fused features are input into a refinement module consisting of at least one convolutional layer. The functions of this convolutional layer are: ① to reduce the dimensionality and integrate the information of the concatenated features; ② to enhance the model's expressive power through a non-linear activation function; and ③ to adaptively learn how to optimally combine the fused semantic and detailed information, correcting potential semantic biases and structural irregularities. Finally, this module outputs a feature with good spatial size, number of channels, and... The predicted feature maps are completely identical and have been enhanced in both spatial detail and semantic accuracy. .

[0171] This implementation effectively utilizes the diverse features within the network by constructing a feedback prediction path of "upsampling-feature fusion-convolutional refinement," achieving accurate feedback and detailed reconstruction of high-level semantic information to low-level features, and significantly improving the quality of feature representation.

[0172] exist Figure 3Based on the embodiment shown, as an implementation manner of the embodiment of the application, the learnable network module for detail enhancement is an attention mechanism module.

[0173] The feedback prediction channel generates the predicted feature map through the following steps b1 to b3:

[0174] Step b1, up-sampling the second feature map to preliminarily align its spatial size and channel number with the first feature map.

[0175] The high-level +1 second feature map is input into an up-sampling layer, and through operations such as bilinear interpolation or transposed convolution, its spatial size is restored to the same scale as the first feature map , completing the basic alignment of the spatial dimension and making its channel number the same as that of the first feature map .

[0176] Step b2, inputting the up-sampled feature into an attention mechanism module to generate a feature weight for enhancing key details.

[0177] The up-sampled feature is input into an attention mechanism module. The attention mechanism module analyzes the spatial and channel information of the feature map, and adaptively learns and generates a feature weight map. Specifically, for spatial attention, the attention mechanism module evaluates the importance of each spatial position in the feature map, generates a spatial weight matrix, and highlights the key regions related to target recognition (such as the edges of pedestrians and the outlines of vehicles), while suppressing background interference. For channel attention, the attention mechanism module evaluates the importance of different feature channels, generates a channel weight vector, and strengthens the feature channels that contribute more to the current recognition task.

[0178] Step b3, weighting the up-sampled feature using the feature weight to output the predicted feature map.

[0179] The spatial weight matrix and the channel weight vector generated in step b2 are applied to the up-sampled feature. First, the different feature channels are scaled by the channel weight, and then the different spatial positions are emphasized by the spatial weight, achieving targeted enhancement of key details. This dual attention weighting mechanism enables the network to automatically focus on the most relevant feature components for the recognition task, and finally outputs a predicted feature map that has significantly enhanced details while maintaining semantic information.

[0180] This implementation introduces an attention mechanism, enabling the feedback prediction channel to adaptively focus on key details, effectively solving the problems of irregular structures and blurred details in complex scenes, and providing a higher-quality basis for subsequent error calculation and feature modulation.

[0181] In Figure 3 the illustrated embodiment, as an implementation manner of the embodiment of the present application, the convolutional neural network is configured to contain multiple feedback self-learning structures;

[0182] Each feedback self-learning structure is established based on a level pair , +1). The level pair includes a previous level with a higher spatial resolution and a subsequent level +1 with a lower spatial resolution.

[0183] The multiple feedback self-learning structures collectively constitute a multi-level feedback self-learning system.

[0184] Specifically, first, a deep convolutional neural network is constructed as a basic architecture. In the network, multiple non-continuous level pairs are selected to establish feedback self-learning structures. For example, a first feedback structure based on the level pair , and a second feedback structure based on the level pair , may be established at the same time, where to ensure that the two feedback structures do not overlap in levels.

[0185] The execution of the multi-level feedback self-learning system includes the following steps:

[0186] Step c1, forward propagation and feature recording.

[0187] The input image data is normally forward propagated in the network. When propagated to the level of the first feedback structure, the system records the first feature map output by the level; when propagated to the level , the second feature map is recorded. Similarly, when propagated to the levels and of the second feedback structure, the corresponding feature maps and are recorded, respectively.

[0188] Step c2, parallel feedback prediction

[0189] The feedback prediction channels in each feedback structure work in parallel:

[0190] The first feedback prediction channel G ( → ​​) receiving feature map , outputting a predicted feature map aligned with . .

[0191] second feedback prediction path G ( → ) receiving feature map , outputting a predicted feature map aligned with . .

[0192] Step c3, parallel error calculation and modulation.

[0193] Each feedback structure independently calculates error and performs feature modulation:

[0194] calculating a first error signal = - , generating a modulated feature map through a first modulator .

[0195] calculating a second error signal = - , generating a modulated feature map through a second modulator .

[0196] Step c4, sequential re-computation mechanism.

[0197] Since the feedback structures are established at different depths, the system adopts a sequential re-computation strategy:

[0198] First, replace the original feature map with the modulated feature map , and start re-forward computation from level ; when the re-computation proceeds to level , replace the original feature map with the feature map obtained at this time , and continue to complete the computation of the remaining levels. The present implementation builds a multi-layer feedback self-learning system, which achieves the following beneficial effects:

[0199] 1. Progressive feature optimization: the bottom feedback structure mainly corrects spatial details and local features, and the middle feedback structure calibrates at a higher semantic level, forming a progressive optimization process from details to semantics.

[0200] 1. Progressive feature optimization: the bottom feedback structure mainly corrects spatial details and local features, and the middle feedback structure calibrates at a higher semantic level, forming a progressive optimization process from details to semantics.

[0201] ​2. Error propagation suppression: By setting feedback correction mechanisms at multiple key points in the network, the inter-layer propagation of feature bias is effectively suppressed, significantly improving the feature expression quality of the entire network.

[0202] 3. Training stability improvement: The multi-layer feedback structure provides more short-circuit paths for gradient propagation, alleviating the gradient vanishing problem in deep networks, making the training process more stable and efficient.

[0203] 4. Anti-interference ability enhancement: Different levels of feedback structure have different robustness to different types of input disturbance, and multi-layer cooperation ensures stable performance of the system in complex environment.

[0204] The design of this multi-layer feedback self-learning system enables the network to fine-grainedly calibrate feature expression from multiple abstraction levels, demonstrating excellent performance and robustness in complex visual perception tasks of intelligent driving.

[0205] In a second aspect, an embodiment of the present application provides a convolutional neural network optimization device based on autonomous learning, which comprises:

[0206] An image data acquisition module for acquiring original image data containing pedestrians, vehicles or traffic signs collected by a vehicle-mounted camera;

[0207] An image data input module for inputting the original image data to a convolutional neural network for forward propagation, the network comprising multiple sequentially connected levels, wherein the spatial resolution of the feature map decreases layer by layer as the level deepens;

[0208] A feature map recording module for recording the feature map extracted from the adjacent level during the forward propagation process, wherein the feature map extracted from the adjacent level includes: a first feature map from the previous level with higher spatial resolution, and a second feature map from the next level +1 with lower spatial resolution after the level ;

[0209] A predicted feature map generation module for inputting the second feature map to a feedback prediction channel, the feedback prediction channel comprising an up-sampling layer and a learnable network module for detail enhancement, and through the transformation of the feedback prediction channel, outputting a predicted feature map which is enhanced in spatial details and has the same spatial size and channel number as the first feature Figure One ;

[0210] An error signal calculation module for calculating the difference between the first feature map and the predicted feature map to generate an error signal;

[0211] The modulated feature map generation module is configured to input the error signal into a modulator, which generates a modulation signal according to the error signal and adaptively adjusts the first feature map, and output a modulated feature map.

[0212] The feature map adjustment module is configured to replace the original first feature map with the modulated feature map and perform forward propagation calculation again from the first layer.

[0213] The recognition result output module is configured to use the recognition result output by the final network as a decision-making control of the intelligent driving system.

[0214] In a third aspect, an embodiment of the present application provides an electronic device, comprising:

[0215] At least one processor;

[0216] Memory for storing instructions executable by the at least one processor;

[0217] The at least one processor is configured to execute the instructions to implement the method of the first aspect.

[0218] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium having stored thereon a computer program, the program being executed by a processor to implement the method of the first aspect.

[0219] The embodiment of the present application introduces an inter-layer feedback self-learning mechanism composed of a feedback prediction channel and a modulator in the network, and constructs a multi-level, closed-loop feedback correction system. The mechanism can use the abstract semantic information of the higher layer as a guide signal, reconstruct the bottom layer features in detail through the learnable feedback prediction channel, and dynamically correct the deviation in the bottom layer feature expression with the help of the flexible modulator. This innovative design effectively overcomes the problems of feature distortion, detail loss and semantic deviation faced by traditional feedforward convolutional neural networks in complex scenes, significantly improves the feature robustness and detail preservation ability of the model in multiple interference environments such as sudden light changes, motion blurring, and rainy and snowy weather. This method not only effectively reduces the missed detection and false detection of key targets such as pedestrians and vehicles, but also improves the stability of reasoning while maintaining high accuracy, ultimately providing more reliable and accurate recognition results for the perception module of the intelligent driving system, laying a solid technical foundation for vehicle decision-making and control, and greatly improving the driving safety level.​

Claims

1. A convolutional neural network optimization method based on autonomous learning, characterized in that, The method includes: Acquire raw image data containing pedestrians, vehicles, or traffic signs captured by vehicle-mounted cameras; The original image data is input into a convolutional neural network for forward propagation. The network contains multiple sequentially connected layers, wherein the spatial resolution of the feature map decreases with each layer as the depth increases. During the forward propagation process, feature maps extracted from adjacent layers are recorded, wherein the feature maps extracted from adjacent layers include those from the previous layer with higher spatial resolution. The first feature map, and from the immediately following hierarchy Then, the next level with lower spatial resolution +1 second feature map; the previous level with higher spatial resolution , is the layer in the convolutional neural network whose spatial resolution of the first output feature map is lower than that of the input image; The second feature map is input into the feedback prediction channel, which includes an upsampling layer and a learnable network module for detail enhancement. Through the transformation of the feedback prediction channel, a predicted feature map that is consistent with the first feature map in terms of spatial size and number of channels and is enhanced in terms of spatial details is output. Calculate the difference between the first feature map and the predicted feature map to generate an error signal; The error signal is input to a modulator, which generates a modulated signal based on the error signal, adaptively adjusts the first feature map, and outputs a modulated feature map. The original first feature map is replaced with the modulated feature map, and the process is repeated from the first feature map. Layered forward propagation calculation; The final network output recognition results are used for decision-making and control in intelligent driving systems. The convolutional neural network is configured to include multiple feedback self-learning structures; each feedback self-learning structure is based on a hierarchical pair ( , +1) Establish; multiple feedback self-learning structures together constitute a multi-layer feedback self-learning system.

2. The method according to claim 1, characterized in that, The learnable network module used for detail enhancement is a feature fusion module; The feedback prediction channel generates the prediction feature map through the following steps: The second feature map is upsampled to initially align its spatial size and number of channels with the first feature map. The upsampled features are compared with those from the convolutional neural network. The detailed features from the shallower hierarchical jumps are then fused together; The fused features are processed through at least one convolutional layer to output the predicted feature map.

3. The method according to claim 1, characterized in that, The learnable network module used for detail enhancement is an attention mechanism module; The feedback prediction channel generates the prediction feature map through the following steps: The second feature map is upsampled to initially align its spatial size and number of channels with the first feature map. The upsampled features are input into the attention mechanism module to generate feature weights for enhancing key details; The upsampled features are weighted using the feature weights to output the predicted feature map.

4. The method according to claim 1, characterized in that, The modulator generates a modulation signal based on the error signal and adaptively adjusts the first feature map. Specifically, this is achieved as follows: Channel-level modulation includes: generating a set of channel scaling coefficients and channel bias coefficients based on the error signal, and scaling and biasing each channel of the first feature map; or, Feature map-level modulation includes: generating a feature correction matrix with the same shape as the first feature map based on the error signal, and adding the matrix to the first feature map; or, Convolutional kernel-level modulation includes: generating a convolutional kernel adjustment amount based on the error signal, used to adjust the kernel level... The weight parameters of the convolution kernel are adjusted, and the first feature map is recalculated using the adjusted convolution kernel.

5. The method according to claim 4, characterized in that, Based on the computing power of the deployed equipment, the implementation method of the modulator is dynamically selected, including: When computational efficiency is a priority, the channel-level modulation with the lightest computational load is selected; When fine-tuning of the spatial details of the feature map is required, the feature map-level modulation with a computational load higher than channel-level modulation but lower than convolution kernel-level modulation is selected. When long-term optimization of core network parameters is required and computational resources allow, the convolutional kernel-level modulation with the heaviest computational load should be selected.

6. A convolutional neural network optimization device based on autonomous learning, characterized in that, The device includes: The image data acquisition module is used to acquire raw image data containing pedestrians, vehicles, or traffic signs captured by the vehicle-mounted camera. An image data input module is used to input the original image data into a convolutional neural network for forward propagation. The network contains multiple sequentially connected layers, wherein the spatial resolution of the feature map decreases layer by layer as the layer deepens. The feature map recording module is used to record feature maps extracted from adjacent layers during the forward propagation process, wherein the feature maps extracted from adjacent layers include those from the previous layer with higher spatial resolution. The first feature map, and from the immediately following hierarchy Then, the next level with lower spatial resolution +1 second feature map; the previous level with higher spatial resolution , is the layer in the convolutional neural network whose spatial resolution of the first output feature map is lower than that of the input image; A predictive feature map generation module is used to input the second feature map into a feedback prediction channel. The feedback prediction channel includes an upsampling layer and a learnable network module for detail enhancement. Through the transformation of the feedback prediction channel, a predictive feature map that is consistent with the first feature map in terms of spatial size and number of channels and is enhanced in terms of spatial details is output. An error signal calculation module is used to calculate the difference between the first feature map and the predicted feature map, and generate an error signal; The modulated feature map generation module is used to input the error signal to a modulator, which generates a modulation signal according to the error signal, adaptively adjusts the first feature map, and outputs the modulated feature map. The feature map adjustment module is used to replace the original first feature map with the modulated feature map and re-perform the self-adjustment. Layered forward propagation calculation; The recognition result output module is used to apply the final recognition results output by the network to the decision control of the intelligent driving system. The convolutional neural network is configured to include multiple feedback self-learning structures; each feedback self-learning structure is based on a hierarchical pair ( , +1) Establish; multiple feedback self-learning structures together constitute a multi-layer feedback self-learning system.

7. An electronic device, characterized in that, include: At least one processor; Memory for storing the at least one processor-executable instruction; The at least one processor is configured to execute the instructions to implement the method as described in any one of claims 1-5.

8. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method as described in any one of claims 1-5.

Images