A target detection method and system based on deep learning

Abstract

The present invention discloses a target detection method based on deep learning, comprising the following steps: S1: classification, deep generative model, establishing the probability distribution of model data, by sharing parameters at multiple spatial positions, CNN can extract common features and generate new data samples, which is particularly suitable for processing structured data, such as images and videos; S2: positioning, building a deep neural network DNN architecture, extracting spatial features, identifying key points, and recursive neural network RNN ​​is used to process time series data, capture motion patterns, and simulate the dependencies between data; S3: detection, using the data model of positioning to obtain a detection model of deep learning, combined with region nomination and convolutional neural network, it can automatically learn and extract image features and obtain an accuracy data model. The present invention maintains a high accuracy, and it effectively captures the detailed information of the object by predicting bounding boxes and classification confidence on feature maps at different levels.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a target detection method and system based on deep learning. Background Art

[0002] As a basic problem in the field of computer vision, object detection has always attracted much attention and has been widely used in the fields of unmanned driving, video surveillance, video analysis, etc. Before the emergence of deep learning, traditional object detection methods are generally divided into three steps: first, region selection, then feature extraction of the target region, and finally classification and regression of the prediction results. Representative methods include background subtraction and optical flow.

[0003] At present, deep learning-based object detection methods are mainly divided into two categories: region proposal-based and bounding box regression-based models. Region proposal-based object detection is mainly divided into two steps: first, possible regions are generated and features are extracted through convolutional neural networks (CNNs), and then they are put into classifiers for classification and position correction.

[0004] The idea of ​​the object detection algorithm based on bounding box regression is to directly regress the bounding box of the predicted target object. The specific method is to first transform the image to a fixed size and divide it into an S*S grid, then extract the features of each grid through a convolutional neural network, and finally directly output the prediction result on the classifier. The object detection algorithm based on region proposal is not suitable for real-time detection because it is time-consuming in the process of extracting the region of interest. However, it also has problems such as over-reliance on pre-trained models and insufficient detection accuracy, which need to be improved. Summary of the invention

[0005] Based on the technical problems existing in the background technology, the present invention proposes a target detection method and system based on deep learning.

[0006] The present invention proposes a target detection method based on deep learning, comprising the following steps: S1: Classification, deep generative model, establishes the probability distribution of model data, by sharing parameters at multiple spatial locations, CNN can extract common features and generate new data samples, which is particularly suitable for processing structured data such as images and videos; S2: Positioning, building a deep neural network (DNN) architecture, extracting spatial features, identifying key points, and recurrent neural network (RNN) is used to process time series data, capture motion patterns, and simulate dependencies between data; S3: Detection, using the positioning data model to obtain a deep learning detection model, combined with region nomination and convolutional neural network, it can automatically learn and extract image features to obtain an accuracy data model; The S3 comprises the following steps: Step 1: Candidate region generation: First, the candidate regions in the image are generated by the Selective Search or Edge Boxes method, and the image model is updated in real time using sensor data to improve positioning accuracy; Step 2: Feature extraction: Extract features for each candidate region using a convolutional neural network (CNN). The advantage of this method is that it has higher accuracy because they use a more complex model to process each candidate region and process each candidate region independently.

[0007] Step 3: Classification and regression: Classify the extracted features to determine whether each region contains the target object, and regress the bounding box to accurately locate the position of the target object.

[0008] Preferably, the deep neural network (DNN) architecture extracts rich features through multimodal data including images or sensors, and learns similarity metrics of image features.

[0009] Preferably, the deep learning model can fuse data from different sensors such as Wi-Fi, Bluetooth, and IMU, enhance positioning accuracy, and associate the correlation between different source data.

[0010] Preferably, the CNN model is used to extract key points in an image, and the model is trained using unlabeled images using self-supervised learning techniques.

[0011] A system involved in a target detection method based on deep learning, comprising: An input module, used for acquiring an image to be processed; A preprocessing module, used for preprocessing the image to be processed to obtain a first image; A feature extraction module, used for inputting the first image into a convolutional neural network for feature extraction to obtain a first feature map, and using the first feature map as a first set of input data; The first to nth convolution modules are used to perform a convolution operation on the i-th feature map to obtain the i+1-th feature map; The first to nth downsampling modules are used to perform a pooling convolution operation on the i-th feature map, and then combine the result of the pooling convolution operation on the i-th feature map with the i+1-th feature map to obtain the i+1-th group of input data; the pooling convolution operation includes a 2*2 pooling with a step size of 2 and a 1*1 convolution; the combination is an addition and averaging operation; The detection module is used to detect the types of objects contained in the image to be processed and the position of each object according to the first to n+1th groups of input data to obtain a detection result.

[0012] The beneficial effects of the present invention are: 1. The input of each convolution operation is a combination of the outputs of all previous convolution operations. Therefore, the present invention fuses feature maps of different scales and reduces the number of parameters. At this time, feature maps of multiple sizes are used for detection, thereby improving the detection capability of objects of different sizes. The speed is also very fast while maintaining a high accuracy. It effectively captures the detailed information of the object by predicting the bounding box and classification confidence on feature maps at different levels.

[0013] 2. Learn the similarity measurement of image features and associate the correlation between different source data. It can automatically learn and extract image features, improving the accuracy and efficiency of detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A flowchart of a target detection method based on deep learning proposed by the present invention; Figure 2 This is a module block diagram of a deep learning-based target detection system proposed in the present invention. DETAILED DESCRIPTION

[0015] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0016] Reference Figure 1-2 , a target detection method based on deep learning, comprising the following steps: S1: Classification, deep generative model, establishes the probability distribution of model data, by sharing parameters at multiple spatial locations, CNN can extract common features and generate new data samples, which is particularly suitable for processing structured data such as images and videos; S2: Positioning, building a deep neural network (DNN) architecture, extracting spatial features, identifying key points, and recurrent neural network (RNN) is used to process time series data, capture motion patterns, and simulate dependencies between data; S3: Detection, using the positioning data model to obtain a deep learning detection model, combined with region nomination and convolutional neural network, it can automatically learn and extract image features to obtain an accuracy data model; The S3 comprises the following steps: Step 1: Candidate region generation: First, the candidate regions in the image are generated by the Selective Search or Edge Boxes method, and the image model is updated in real time using sensor data to improve positioning accuracy; Step 2: Feature extraction: Extract features for each candidate region using a convolutional neural network (CNN). The advantage of this method is that it has higher accuracy because they use a more complex model to process each candidate region and process each candidate region independently.

[0017] Step 3: Classification and regression: Classify the extracted features to determine whether each region contains the target object, and regress the bounding box to accurately locate the position of the target object.

[0018] In the present invention, the deep neural network DNN architecture extracts rich features through multimodal data including images or sensors, and learns similarity metrics of image features.

[0019] In the present invention, the deep learning model can fuse data from different sensors such as Wi-Fi, Bluetooth, and IMU, enhance positioning accuracy, and associate the correlation between different source data.

[0020] In the present invention, the CNN model is used to extract key points in the image, and the model is trained using unlabeled images using self-supervised learning technology.

[0021] A system involved in a target detection method based on deep learning, comprising: An input module, used for acquiring an image to be processed; A preprocessing module, used for preprocessing the image to be processed to obtain a first image; A feature extraction module, used for inputting the first image into a convolutional neural network for feature extraction to obtain a first feature map, and using the first feature map as a first set of input data; The first to nth convolution modules are used to perform a convolution operation on the i-th feature map to obtain the i+1-th feature map; The first to nth downsampling modules are used to perform a pooling convolution operation on the i-th feature map, and then combine the result of the pooling convolution operation on the i-th feature map with the i+1-th feature map to obtain the i+1-th group of input data; the pooling convolution operation includes a 2*2 pooling with a step size of 2 and a 1*1 convolution; the combination is an addition and averaging operation; The detection module is used to detect the types of objects contained in the image to be processed and the position of each object according to the first to n+1th groups of input data to obtain a detection result.

[0022] In the present invention, the input of each convolution operation is a combination of the outputs of all previous convolution operations. Therefore, the present invention fuses feature maps of different scales and reduces the amount of parameters. At this time, feature maps of multiple sizes are used for detection, thereby improving the detection capability of objects of different sizes. The speed is also very fast while maintaining a high accuracy. It effectively captures the detailed information of the object by predicting the bounding box and classification confidence on feature maps at different levels.

[0023] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A target detection method based on deep learning, characterized in that: The following steps are involved: S1: Classification, deep generative model, establishes the probability distribution of model data, and CNN can extract common features by sharing parameters at multiple spatial locations; S2: Positioning, building a deep neural network (DNN) architecture, extracting spatial features, identifying key points, and recurrent neural network (RNN) is used to process time series data and capture motion patterns; S3: Detection, using the positioning data model to obtain a deep learning detection model, combined with region nomination and convolutional neural network to obtain an accuracy data model; The S3 comprises the following steps: Step 1: Candidate region generation: First, generate candidate regions in the image through Selective Search or Edge Boxes method; Step 2: Feature extraction: Extract features for each candidate region using a convolutional neural network (CNN). Step 3: Classification and regression: Classify the extracted features to determine whether each region contains the target object, and regress the bounding box to accurately locate the position of the target object.

2. The target detection method based on deep learning according to claim 1, characterized in that: The deep neural network (DNN) architecture extracts rich features through multimodal data including images or sensors, and learns similarity metrics of image features.

3. The target detection method based on deep learning according to claim 1, characterized in that: The deep learning model can fuse data from different sensors such as Wi-Fi, Bluetooth, and IMU, enhance positioning accuracy, and associate the correlation between different source data.

4. The target detection method based on deep learning according to claim 1, characterized in that: The CNN model is used to extract key points in an image, and the model is trained using unlabeled images using self-supervised learning technology.

5. The system according to claim 1, wherein: include: An input module, used for acquiring an image to be processed; A preprocessing module, used for preprocessing the image to be processed to obtain a first image; A feature extraction module, used for inputting the first image into a convolutional neural network for feature extraction to obtain a first feature map, and using the first feature map as a first set of input data; The first to nth convolution modules are used to perform a convolution operation on the i-th feature map to obtain the i+1-th feature map; The first to nth downsampling modules are used to perform a pooling convolution operation on the i-th feature map, and then combine the result of the pooling convolution operation on the i-th feature map with the i+1-th feature map to obtain the i+1-th group of input data; the pooling convolution operation includes a 2*2 pooling with a step size of 2 and a 1*1 convolution; the combination is an addition and averaging operation; The detection module is used to detect the types of objects contained in the image to be processed and the position of each object according to the first to n+1th groups of input data to obtain a detection result.

Images