Intelligent image recognition optimization method based on deep learning
By adopting a deep learning-based intelligent image recognition optimization method in the image recognition system, using Bottleneck, Conv and Focus modules for feature extraction, and using a neural network classifier based on metaheuristic strategy, the problem of insufficient accuracy and generalization capabilities of image recognition systems in the prior art in complex scenes and multi-category images is solved, and higher recognition accuracy and more stable performance are achieved.
Patent Information
- Application Number
- CN202411623142.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-14
- Publication Date
- 2025-03-11
AI Technical Summary
When existing image recognition systems process complex scenes and multi-category images, their accuracy and generalization capabilities are limited, and due to the quality and diversity of data, resulting in insufficient generalization capabilities of models.
The intelligent image recognition optimization method based on deep learning is adopted, including data preparation, model construction, feature extraction model training, classifier training and intelligent image recognition. Feature extraction was performed using Bottleneck, Conv, and Focus modules, and image recognition was performed using a neural network classifier based on metaheuristic strategy.
Improves the accuracy and generalization ability of image recognition, can perform more stably in complex scenes and multi-category images, and significantly reduces the computing resource requirements, allowing the system to run in real time on resource-constrained devices.
Smart Images
Figure CN119600334A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of image processing and computer vision, and more specifically, to an intelligent image recognition optimization method based on deep learning. Background Art
[0002] Image recognition technology has been widely used in many fields, such as security monitoring and autonomous driving. Traditional image recognition methods are usually based on feature extraction and classifier design, such as SIFT features, HOG features, etc., and combined with classifiers such as support vector machines and decision trees for image classification and recognition. Some existing image recognition systems use a simple convolutional neural network (CNN) architecture to extract image features through multi-layer convolution and pooling operations, and then use a fully connected layer for classification. However, this system has limited accuracy and generalization ability when dealing with complex scenes and multi-category images.
[0003] Traditional image recognition systems usually rely on a large amount of labeled data for training, which is not only time-consuming and labor-intensive, but also limited by the quality and diversity of the data, which may lead to insufficient generalization of the model. In practical applications, samples of certain categories may be difficult to obtain, resulting in an unbalanced training data set, affecting the model's recognition performance on these categories. Existing feature extraction methods may not be able to fully capture the key information of the image, resulting in insufficiently rich feature representation, affecting the accuracy of subsequent classification.
[0004] Therefore, a new solution needs to be proposed to solve this problem. Summary of the invention
[0005] In view of the shortcomings of the prior art, the purpose of the present invention is to provide an intelligent image recognition optimization method based on deep learning.
[0006] The above technical objectives of the present invention are achieved through the following technical solutions: an intelligent image recognition optimization method based on deep learning, comprising the following steps: S1: data preparation;
[0007] S2: Build the model;
[0008] S3: feature extraction model training;
[0009] S4: classifier training, inputting the dimension-reduced data into the classifier for classifier training;
[0010] S5: Intelligent image recognition.
[0011] Preferably, the method for preparing data in S1 is: identifying each frame of the original image according to a computer image tool, marking each target on each frame of the original image with a target frame, and cutting out a target area from each frame of the original image according to the position of the marked target frame.
[0012] Preferably, the method for building a model in S2 is to build a sequential model using tf.keras.Sequential() and add each layer in turn.
[0013] Preferably, the feature extraction model training method in S3 comprises the following steps:
[0014] S3.1: Through the Bottleneck method: The standard Bottleneck module first reduces the number of feature channels and then expands it (by default, it is reduced to half). If a shortcut operation is required, the input and the output of the convolutional layer are directly added without changing the number of feature channels;
[0015] S3.2: Through the Conv method: The Conv module defines a standard convolutional layer, which includes a Conv2d, a BN layer, and an activation function. Yolov5 uses a Bn layer and an activation function for optimization. The convolution step size of the early backbone network of the algorithm is 2 to reduce the size of the feature map. The convolution step size of the late detect part of the network is 1, and the size of the feature map is no longer reduced.
[0016] S3.3: Through the Focus method, each output channel is sliced and concat together and then output using a standard convolution module. The Focus module reduces the length and width of the feature map and increases the depth of the feature map to retain the feature information of each position. The Focus module can replace the convolution operation with a step size of 2 to reduce the feature size, reduce the model calculation amount, and retain the original information.
[0017] Preferably, the method for training the classifier in S4 is: using a neural network classifier based on a meta-heuristic strategy to perform image recognition classification, and using the argparse.ArgumentParser() method to find one of the optimal or approximately optimal solutions in the search space.
[0018] Preferably, the method of intelligent image recognition in S5 is: after the model training is completed, the trained model is used to process new data, and the newly collected image is processed by the result_to_xywh_landmark_conf method, and the features are processed by the trained feature extraction model and feature dimensionality reduction model, and the obtained features are input into the classifier for classification.
[0019] Preferably, S3.2 uses the SiLU activation function.
[0020] In summary, the present invention has the following beneficial effects: This solution has the following advantages:
[0021] It has good fault tolerance, parallel processing and self-learning capabilities, allows samples to have large defects and distortions, runs fast, has good adaptive performance and has high resolution.
[0022] There is no need for complex preprocessing of samples. A two-dimensional image can be directly input into the model according to the extracted XML, and the classification result will be given at the output. There is no need for complex preprocessing, and feature extraction and pattern classification are completely put into a black box.
[0023] It has a multi-layer hierarchical network structure, which reduces the number of trainable parameters in the network by mining the spatial correlation of data, thereby improving the efficiency of the back-propagation algorithm of the forward propagation network.
[0024] The use of a more advanced feature extraction architecture improves the ability to express features of complex images, thereby greatly improving recognition accuracy.
[0025] Through innovative data augmentation and regularization techniques, the generalization ability of the model is enhanced, making it more stable in new scenarios and data sets.
[0026] Optimized model compression and quantization algorithms significantly reduce computing resource requirements, enabling the system to run in real time on resource-constrained devices. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION
[0028] The present invention is described in detail below in conjunction with the accompanying drawings and embodiments.
[0029] Example:
[0030] 1. An intelligent image recognition optimization method based on deep learning, such as Figure 1 As shown, the following steps are included:
[0031] S1: Data preparation; The method of data preparation in S1 is: identify each frame of the original image according to the computer image tool, mark each target on each frame of the original image with a target frame, and cut out the target area from each frame of the original image according to the position of the marked target frame;
[0032] S2: Build a model; The method of building a model in S2 is: use tf.keras.Sequential() to build a sequential model and add each layer in sequence;
[0033] S3: feature extraction model training;
[0034] S4: classifier training, inputting the dimension-reduced data into the classifier for classifier training;
[0035] S5: Intelligent image recognition.
[0036] The feature extraction model training method in S3 includes the following steps:
[0037] S3.1: Through the Bottleneck method: The standard Bottleneck module first reduces the number of feature channels and then expands it (by default, it is reduced to half). If a shortcut operation is required, the input and the output of the convolutional layer are directly added without changing the number of feature channels;
[0038] S3.2: Through the Conv method: The Conv module defines a standard convolution layer, which includes a Conv2d, a BN layer, and an activation function. The Conv module is the most basic component of the network. Due to the large depth of the target detection network at this stage, various measures have been taken to prevent the occurrence of gradient explosion or diffusion. Among them, Yolov5 uses a Bn layer and an activation function for optimization. The early backbone network of the algorithm extracts features, and the convolution step size is 2, which reduces the size of the feature map, continuously abstracts meaningful features, and reduces the calculation amount of the subsequent network. Different from this, the convolution step size of the late detect part of the network is 1, and the size of the feature map is no longer reduced; S3.2 uses the SiLU activation function;
[0039] In deep neural networks, the output of SiLU automatically adapts to the size of the input, which helps prevent the gradient from disappearing or exploding, thereby improving the training efficiency of the network. Unlike monotonic activation functions such as ReLU, SiLU can convert between positive and negative values, providing richer information processing capabilities. SiLU can reduce the number of parameters required for the model, making the model more lightweight. SiLU can improve the performance of certain deep learning models, especially in image classification and natural language processing tasks. The operation content of the Conv method can increase the generalization ability of the model and increase the types of samples obtained.
[0040] S3.3: Through the Focus method, each output channel is sliced and concat together and then output using a standard convolution module. The Focus module reduces the length and width of the feature map and increases the depth of the feature map to retain the feature information of each position. The Focus module can replace the convolution operation with a step size of 2 to reduce the feature size, reduce the model calculation amount, and retain the original information.
[0041] The method of classifier training in S4 is: using a neural network classifier based on a meta-heuristic strategy for image recognition classification, and using the argparse.ArgumentParser() method to find one of the optimal and approximately optimal solutions in a wide search space to improve the performance of the model.
[0042] The method of intelligent image recognition in S5 is: after the model training is completed, the trained model is used to process new data. The result_to_xywh_landmark_conf method is used to process the newly collected images through the trained feature extraction model and feature dimensionality reduction model, and then the obtained features are input into the classifier for classification.
[0043] This scheme has good fault tolerance, parallel processing and self-learning capabilities, allows samples to have large defects and distortions, runs fast, has good adaptive performance, and has high resolution; it does not require complex preprocessing of samples, and can directly input a two-dimensional image into the model according to the extracted XML, and give the classification results at the output end, without complex preprocessing, and completely puts feature extraction and pattern classification into a black box; it has a multi-layer hierarchical network structure, and the convolutional neural network is considered to be the first truly successful, multi-layer hierarchical network with robust deep learning method. It reduces the number of trainable parameters in the network by mining the spatial correlation of data, thereby improving the efficiency of the back propagation algorithm of the forward propagation network. In CNN, small areas in the image are treated as the bottom input data in the hierarchy, and information passes through each layer in the network through forward propagation. In each layer, filters are formed to obtain some significant features of the observed data. Because the local perception area can obtain some basic features, such as the boundaries and corners in the image, this method can provide a certain degree of relative invariance to displacement, stretching and rotation. The close connections and spatial information between layers in CNN make it particularly suitable for image processing and understanding, and it can automatically extract rich relevant features from images.
[0044] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.
Claims
1. An intelligent image recognition optimization method based on deep learning, characterized in that: The following steps are involved: S1: Data preparation; S2: Build the model; S3: feature extraction model training; S4: classifier training, inputting the dimension-reduced data into the classifier for classifier training; S5: Intelligent image recognition.
2. The deep learning-based intelligent image recognition optimization method according to claim 1, characterized in that: The data preparation method in S1 is: identify each frame of the original image according to the computer image tool, mark each target on each frame of the original image with a target frame, and cut out the target area from each frame of the original image according to the position of the marked target frame.
3. The deep learning-based intelligent image recognition optimization method according to claim 2, characterized in that: The method of building a model in S2 is to build a sequential model using tf.keras.Sequential() and add each layer in turn.
4. The method for intelligent image recognition optimization based on deep learning according to claim 3, characterized in that: The feature extraction model training method in S3 includes the following steps: S3.1: Through the Bottleneck method: The standard Bottleneck module first reduces the number of feature channels and then expands it (by default, it is reduced to half). If a shortcut operation is required, the input and the output of the convolutional layer are directly added without changing the number of feature channels; S3.2: Through the Conv method: The Conv module defines a standard convolutional layer, which includes a Conv2d, a BN layer, and an activation function. Yolov5 uses a Bn layer and an activation function for optimization. The convolution step size of the early backbone network of the algorithm is 2 to reduce the size of the feature map. The convolution step size of the late detect part of the network is 1, and the size of the feature map is no longer reduced. S3.3: Through the Focus method, each output channel is sliced and concat together and then output using a standard convolution module. The Focus module reduces the length and width of the feature map and increases the depth of the feature map to retain the feature information of each position. The Focus module can replace the convolution operation with a step size of 2 to reduce the feature size, reduce the model calculation amount, and retain the original information.
5. The method for intelligent image recognition optimization based on deep learning according to claim 4, characterized in that: The method of classifier training in S4 is: using a neural network classifier based on a meta-heuristic strategy for image recognition classification, and using the argparse.ArgumentParser() method to find one of the optimal or approximately optimal solutions in the search space.
6. The method for intelligent image recognition optimization based on deep learning according to claim 5, characterized in that: The method of intelligent image recognition in S5 is: after the model training is completed, the trained model is used to process new data. The result_to_xywh_landmark_conf method is used to process the newly collected images through the trained feature extraction model and feature dimensionality reduction model, and then the obtained features are input into the classifier for classification.
7. The method for intelligent image recognition optimization based on deep learning according to claim 4, characterized in that: S3.2 uses the SiLU activation function.