What kind of neural network is usually used for image processing
Google AI Writing Project: Neural Network Pseudo-Original
What are the improvements in convolutional neural networks
Recent advances in the study of convolutional neural networks have sparked enthusiasm for perfecting stereo matching reconstruction . From a conceptual point of view, learning-based algorithms can capture global semantic information, such as prior conditions based on highlights and reflections, to facilitate more robust matching.
At present, some two-view stereo matching has been explored, and neural networks are used to replace hand-designed similarity measures or regularization methods. These methods show better results and gradually outperform traditional methods in the field of stereo matching.
In fact, the stereo matching task is perfectly suitable for using CNN, because the image pairs have been rectified, so the stereo matching problem is transformed into pixel-by-pixel disparity estimation in the horizontal direction.
Unlike binocular stereo matching, the input of MVS is an arbitrary number of views, which is a thorny problem to be solved by deep learning methods.
And only a few works are aware of this problem. For example, SurfaceNet reconstructs the color voxel cube in advance, and forms a 3D cost body with the color information of all pixels and camera parameters, and the formed 3D cost body is the input of the network.
However, limited by the huge memory consumption of the 3D cost body, it is difficult to increase the size of the SurfaceNet network: SurfaceNet uses a heuristic "divide and conquer" strategy, which takes a long time for large-scale reconstruction of scenes.
How can a fully convolutional neural network improve image segmentation accuracy?
The fully convolutional neural network can improve the accuracy of image segmentation through the neural network model. There are also many fully convolutional neural network solutions.
Baidu/Google search for overfitting, and individuals will first try to reduce the network size, such as the number of layers, the number of convolution filters, and the number of units in the fully connected layer.
Others such as Dropout, data enhancement/expansion, regularization, earlystop, and batchnorm can also be tried.
Fully convolutional neural network hidden layer: The hidden layer of the fully convolutional neural network convolutional neural network includes three common structures: convolutional layer, pooling layer, and fully connected layer. Inception may be included in some more modern algorithms. Complex constructions such as modules and residual blocks.
Among common architectures, convolutional and pooling layers are unique to convolutional neural networks. The convolution kernel in the convolution layer contains weight coefficients. Fully convolutional neural networks and pooling layers do not contain weight coefficients, so pooling layers may not be considered as independent layers in the literature.
Taking LeNet-5 as an example, the order of the three common structures in the hidden layer is usually: input-convolution layer-pooling layer-full connection layer-output.
What to do with "graph" structured data using convolutional neural networks
The convolutional neural network has the following applications for research: 1. Shape recognition based on the convolutional network The shape of the object is the basis for the human visual system to analyze and identify objects. The geometric shape is the expression of the essential characteristics of the object and has translation , scaling and rotation invariance, so in the field of pattern recognition, it is of great significance for the analysis and recognition of shapes, and two-dimensional images are a special case and part of three-dimensional images, so the recognition of two-dimensional images is a three-dimensional image recognition Foundation.
2. Face detection based on convolutional network Convolutional neural network is different from traditional face detection methods. It directly acts on input samples, uses samples to train the network and finally realizes the detection task.
It is a non-parametric face detection method, which can save a series of complex processes such as modeling, parameter estimation, parameter testing, and model reconstruction in traditional methods. This paper targets faces of any size, position, pose, orientation, skin color, facial expression, and lighting condition in an image.
3. Character recognition system In classic pattern recognition, features are generally extracted in advance. After extracting many features, it is necessary to perform correlation analysis on these features to find the features that best represent characters, and remove features that are irrelevant to classification and autocorrelation.
However, the extraction of these features is too dependent on human experience and subjective consciousness. The difference in the extracted features has a great impact on the classification performance, and even the order of the extracted features will also affect the final classification performance. At the same time, the quality of image preprocessing will also affect the extracted features.
What is CNN (Convolutional Neural Network)?
In digital image processing, we use convolution to filter because the convolution template we use is indeed a high-pass, low-pass, band-pass and other physical filters in the frequency domain.
However, in the neural network, the parameters of the template are trained. I think it is a purely mathematical thing. It is difficult to understand that it has any meaning in the frequency domain, so I don’t think the convolution in the neural network has a filtering effect. . Then talk about personal understanding.
First of all, regardless of whether it is a convolutional neural network or not, as long as it is a neural network, it is essentially using layers of simple functions (whether it is sigmoid or Relu) to fit an extremely complex function, and the fitting process is through time and time again. Back propagation is used to adjust parameters to minimize the cost function.
What to do with "graph" structured data using convolutional neural networks
The convolutional neural network has the following applications for research: 1. Shape recognition based on the convolutional network The shape of the object is the basis for the human visual system to analyze and identify objects. The geometric shape is the expression of the essential characteristics of the object and has translation , scaling and rotation invariance, so in the field of pattern recognition, it is of great significance for the analysis and recognition of shapes, and two-dimensional images are a special case and part of three-dimensional images, so the recognition of two-dimensional images is a three-dimensional image recognition Foundation.
2. Face detection based on convolutional network Convolutional neural network is different from traditional face detection methods. It directly acts on input samples, uses samples to train the network and finally realizes the detection task.
It is a non-parametric face detection method, which can save a series of complex processes such as modeling, parameter estimation, parameter testing, and model reconstruction in traditional methods. This paper targets faces of any size, position, pose, orientation, skin color, facial expression, and lighting condition in an image.
3. Character recognition system In classic pattern recognition, features are generally extracted in advance. After extracting many features, it is necessary to perform correlation analysis on these features to find the features that best represent characters, and remove features that are irrelevant to classification and autocorrelation.
However, the extraction of these features is too dependent on human experience and subjective consciousness. The difference in the extracted features has a great impact on the classification performance, and even the order of the extracted features will also affect the final classification performance. At the same time, the quality of image preprocessing will also affect the extracted features.