The Feature Extraction Of Visual Image Based On Miniature Mobile Land Robot

The general object recognition based on vision is a research hotspot in the field of computer vision. And also, it’s a challenging issue.After reading up the current mainstream project of object recognition, we decide to apply machine leaning algorithm into feature selection and extraction. Using the powerful learning ability of machine leaning, we try to carry out feature learning by means of unsupervised learning way, then get a good feature set to proceed with classification.Our thesis state the whole process of object recognition, from preprocessing, feature selection and extraction to classification. But our research focused on the self-learning of feature by means of machine learning algorithm, to get a good feature set. Deep learning algorithm is the most popular issue in the field of machine learning, its powerful learning ability in mass data let it can extract the most essential feature. Facing such a complicated problem of multi-classification, we need the ability of abstract expression ability.At present, the classical deep learning algorithm mainly include deep convolutional neural networks, deep belief networks and the relevant transformation of restricted Boltzmann machine unit, auto-encoder networks and the relevant transformation of auto-encoder unit.Our research mainly focused on the point as below:(1).In the competition of image recognition of 2012,CNNs reduced the error rate to nine percent, the convolutional feature, weights of shared, pooling and so on included in CNNs, playing a great role on the recognition.in order to utilize the inborn character of CNNs in the field of computer vision, we designed a targeted unit--convolutional auto-encoder.(2).For the deep learning architecture made up of different units have its own merit in the issue of classification, in order to make full use of the merits of different deep learning architecture, we designed a targeted mixed learning architecture, it contains CAE unit and RBM unit, promoting the recognition rate effectively.(3). After completing the feature learning, we need to checkout that whether our feature map is complete. So we designed a concurrent mixed deep learning architecture to accomplish it. We put the feature map of second line as noise, blending it with the feature set of the mixed deep learning architecture.at last, we use the feature set to make classification. If we change the second line to the same good line of feature extraction, then blending them to make classification, thereby utilize the way of mixed feature and feature correlation to check out the classification.