Reserach On Image Target Detection And Segmentation Algorithm Based On Deep Learning

In recent years, with the improvement upon technology, artificial intelligence technology represented by deep learning has been increasingly used on all aspects of scientific research and engineering, such as automatic driving, man-machine conversation, content-based image retrieval and face identification for which the artificial intelligence technology are broadly used. AlphaGo which had ever been so famous for a while is another example of actual application of deep learning technology. The artificial intelligence technology sources from bionics while most information of the human beings from vision. Therefore, image recognition technology is a crucial field of the artificial intelligence. It is a key question of application of artificial intelligence system to know how to understand scenes and recognize catalogues of the objects.For the traditional image recognition calculation, two steps including feature extraction and classification of the images are conducted separately, which requires human beings to have the characteristics of artificial construction and selection of targets. Not only does it increase human's workload, but also makes human beings unable to design the characteristics excellent enough when facing complicated unfamiliar problems. The key point is that the characteristics designed by human beings are often superficial and the superficial ones concerning only one aspect such as gradient, color or texture. This brings many limits to the traditional image recognition calculation, so there is a limited space for improvement of performance. However, in recent years, with the improvement of hardware level, the deep neural network has begun to become pragmatic. The most outstanding characteristic is an integration of feature extraction and classification into an individual neural network. With deepening network structure, it can extract the deep high-level characteristics of the object. This has brought great-leap-forward development to the artificial intelligence technology in recent years. For scene recognition of robots and automated vehicles, this essay mainly researches the calculation based on deep learning including the following:(1) Realize a fast detection model for scenes of pedestrians and vehicles based on convolutional neural networks. In this paper, we refer to the Distilling Knowledge, Fully Convolutional Networks and other algorithms, improve the real-time performance of pedestrian vehicle detection model based on convolutional neural network. In the field of automatic driving, it is very important to deal with the traffic information in real time. So the work of this chapter also has a strong practical significance.(2) Propose a method of transform class segmentation into instance-aware segmentation. Semantic segmentation algorithm based on the deep neural network can effectively get the recognition of the scene pixel level labels. But when dealing with the same class objects, the semantic segmentation algorithm cannot get a single object. This greatly limits the application of semantic segmentation algorithm. This paper presents an object segmentation algorithm based on the combination of deep neural network and local features, which can effectively deal with the problem of the same kind of conglutination objects, and then get a single object.(3) We propose a method for text position detection based on deep neural network. Through the use of adhesion text images for training neural network can avoid undetected adhesion text problem. In dealing with the text in the natural scene, it is very important to overcome the problem of text adhesion.(4) Realize the image recognition system based on deep learning. In order to reach a balance between thesis and engineering application, there aren't simulation software in this essay. We use the current popular source program libraries together with our own codes. Because all codes are visible and there is no copyright problem, it gives our work strong engineering properties and pragmatic values.