WiMi Hologram Cloud a leading global Hologram Augmented Reality (“AR”) Technology provider announced that a new image classification method is developed to automatically extract features from images using a hierarchical structure inspired by the animal visual system. The method combines bionic pattern recognition (BPR) with CNN, which can fully utilize the geometric structure of the high-dimensional feature space to achieve better classification performance and therefore overcome some of the drawbacks of traditional pattern recognition. The method has been validated in several experiments and, in most cases, achieves higher classification performance than traditional methods.
Top AI ML Insights:Â Why Architecture Matters with Generative AI and Cloud Security
Convolutional Neural Network (CNN) is a deep learning model specialized for processing images. It can automatically extract features from an image through convolution and pooling operations and perform classification using fully connected layers. Convolutional operations involve applying a convolutional kernel (also known as a filter) to each position on the image and outputting the result as a feature map. Pooling operation means down-sampling the feature map to reduce the amount of computation and risk of overfitting.
In traditional CNN image recognition classification models, the softmax function is used for classification. softmax function converts a set of scores into a probability distribution, where each score represents the confidence score that the image belongs to a certain category. Traditional pattern recognition methods usually use hyperplanes in the feature space to segment categories. However, this approach has some downsides, such as the need to manually select features and the difficulty in handling nonlinear data. On the contrary, BPR can overcome these problems by performing class recognition through geometric cover sets that are concatenated in a high-dimensional feature space.
BPR is a bionic-based pattern recognition method, the basic idea of which is to simulate the processing of sensory information using biological systems, and to view the pattern recognition process as taking place in a high-dimensional feature space. In this high-dimensional space, each sample point is regarded as an object rather than a point. Therefore, different classes of samples are distributed in different regions of the high-dimensional feature space, and these regions are called geometric coverage sets. Each geometric covering set consists of a set of geometric objects, which are called geometric primitives, e.g., spheres, cones, polyhedra, etc. By appropriate combinations of geometric primitives, coverage sets with high classification performance can be constructed to enable the recognition of categories.
WiMi combines BPR with CNN to achieve better image classification results. Specifically, CNN image classification based on BPR can map CNN features into a high-dimensional feature space and construct a geometric coverage set in that space, and then display new samples in that space and determine the class they belong to.
WiMi BPR-based CNN image classification uses a mapping function to display CNN features in a high-dimensional feature space. This function can be a simple nonlinear transform such as a polynomial transform or a radial basis function (RBF) transform. It is also possible to learn this mapping function using some more complex functions such as a neural network or a support vector machine (SVM) to transform the CNN features into a form that is easier to classify in the high-dimensional feature space.
This image classification technique has been shown to have high classification performance in high-dimensional feature spaces with geometric primitives, such as spheres, cones or polyhedra, to construct geometric coverage sets. Optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms, can be used to search for the optimal combination of geometric primitives to construct the best geometric coverage set. Finally, a classifier, such as a K-nearest neighbor algorithm or an SVM, is used to identify the class to which the new sample belongs.
The specific way to realize the image classification that combines BPR with CNN is as follows:
Preparation of training and test dataset: a dataset containing images of many different categories needs to be collected. This dataset should contain two parts: the training dataset and the test dataset. The training dataset is used to train the CNN model and the test dataset is used to test the performance of the classifier.
Training CNN model and extracting image features: a CNN model is trained using the training dataset and the features of each image are extracted using the model. These features will be used to construct a geometric coverage set in a high-dimensional feature space.
Mapping CNN features into high-dimensional feature space: a mapping function needs to be used to map the CNN features into the high-dimensional feature space. This mapping function can be learned using some nonlinear transforms such as polynomial transforms or RBF transforms, or using more complex functions such as neural networks or SVMs.
Data Management News: Planet Completes Acquisition of Sinergise; Set to Expand Planet’s Earth Data Platform
Constructing geometric coverage sets: geometric coverage sets are constructed using some geometric primitives that have been shown to have high classification performance in high-dimensional feature spaces, such as spheres, cones, or polyhedra. Then, we can use some optimization algorithms, such as genetic algorithms or particle swarm optimization algorithms, to search for the optimal combination of geometric primitives to construct the best geometric coverage set.
Classifying new samples: a classifier, such as a K-nearest neighbor algorithm or SVM, is used to identify the category to which the new sample belongs. We can map the features of the new sample into a high-dimensional feature space, then find the nearest geometric cover set in that space, and finally classify the new sample into the category represented by the cover set.
This image classification technique is characterized by combining CNN and BPR to classify images by constructing geometric cover sets in a high-dimensional feature space. Compared to the current traditional CNN model using the softmax function for classification, the softmax function has limited capacity and cannot well handle complex classification problems, such as image classification. In addition, the CNN model cannot fully utilize the geometric structure of the high-dimensional feature space, and thus cannot achieve optimal classification performance. As well, traditional pattern recognition methods usually require manual selection of features and classifiers, which requires a lot of labor and time costs. By combining BPR and CNN, this technique can overcome some of the shortcomings of traditional pattern recognition, improve the performance of image classification, and can handle complex image classification problems. This method in image classification can overcome some of the current shortcomings of traditional pattern recognition as well as in most cases, higher classification performance than traditional methods. And it can deal with complex image classification problems, such as image recognition, target detection and image segmentation.
At present, the image classification technology based on CNN has been widely used in many fields, and the method of WiMi combined with BPR can overcome the limitations of traditional pattern recognition methods and improve the accuracy and reliability of image classification. It is believed that with the continuous development and progress of technology, this technology will have wider applications and more outstanding performance in the future.
Recommended CIO Influence News:Â Liquidware Launches Liquidware Ready Program
[To share your insights with us, please write to sghosh@martechseries.com]
