Recently, the Vision and Learning Forum of the 2nd World Intelligence Congress education and academic subsession was held at Tianjin University. The Vision and Learning Forum is hosted by Tianjin University and Tianjin Haihe Education Park Management Committee. Professor Zhang Jiawan, vice dean of software college of Tianjin University, attended the forum and delivered a speech. The forum specially invited Lin Zhouchen of Peking University, Ling Haibin of Temple University, Bai Xiang of Huazhong University of Science and Technology, Zuo Wangmeng of Harbin Institute of Technology, Cheng Mingming of Nankai University, Liu Risheng, Dalian University of Technology, Liu Jiaying, Peking University, Li Chunguang, Beijing University of Posts and Telecommunications, Wang Xinggang, Huazhong University of Science and Technology to present. In addition, a large number of business people, university teachers and students attended the forum, which was presided over by Professor Guo Xiaojie, Young Scholar of Peiyang.

Professor Cheng Mingming introduced the application of machine learning in image semantic recognition in detail in accordance with his research achievements. He considered pixel-level image semantic recognition as the basis for many important computer vision and computer graphics applications. Although relevant research has made rapid progress in recent years, the most advanced solutions are heavily dependent on quality and precise image interpretation of pixels. In contrast, human beings are able to learn how to achieve high-precision semantic recognition and target extraction through online search. Inspired by this phenomenon, we begin with the classification of independent semantic feature extraction technology, such as significance object detection, image segmentation and edge extraction. He also introduced how to use this kind of classification of independent image semantic features to reduce the dependence on precise annotation in the semantic learning process and then achieve an unneeded image semantic recognition technology with manual annotation. By using parallel technology and parallel tracking and mapping technology in SLAM, Professor Ling Haibin proposed a novel parallel tracking and verification framework from a new point of view, and introduced his research of real-time and high-precision tracking. Professor Bai Xiang briefly reviewed the state of text detection and mainstream recognition research technology in recent years, and introduced the latest research progress in the field of scene text detection and Photo OCR, especially the detection and recognition methods for arbitrary shape text and end-to-end recognition neural networks, and outlined the prospects of future development trends in this field. Professor Zuo Wangmeng introduced how to improve and adjust self-coding networks accordingly. Aiming at solving the problem of image coding, he proposed an importance figure subnetwork to solve the discretization problem caused by discretization and rate control, which is applied to the underlying visual problems such as image filling, face attribute editing and image coding. Optimization has always been an integral part of machine learning. However, too many relevant textbooks and research materials make it difficult for beginners to master the optimization algorithm. Therefore, Professor Lin Zhouchen systematically introduced some practical optimization algorithms by explaining process flows in relaxation optimization in his report. Professor Li Chunguang introduced the joint optimization framework of structured sparse spatial clustering to study the close relationship and data extraction of high dimensional data. The framework is based on the representation of each data point as a structured sparse linear combination with all other data points. By using partial edge information, the framework is extended to constrained subspace clustering and semi-supervised learning algorithms for partial labels. Deep learning has acquired great success in many applications, but most existing networks are usually designed using a heuristic method, therefore, they lack strict mathematical principles and derivation. Liu Risheng introduced a series of latest research methods such as integrating experimental network structures and rich task hints to build a deep learning model. Professor Liu Jiaying introduced "Visual Editing" in his report, and introduced how to use a depth learning framework to solve the problem of rain and fog in a single image by using spatial highly regular analysis to guide the synthesis of text effects. Wang Xinggang introduced some methods to solve the problem of weak supervised visual learning using end-to-end training depth network, and introduced target detection and semantic segmentation, and put forward two kinds of new network called proposal clustering learning and deep seed region growing (DSRG ) respectively.

The Vision and Learning Forum brought together top experts and scholars in the field of machine vision and machine learning at home and abroad. During the forum, they shared their own research and explored research trends in machine vision and machine learning. The forum was a great success in bringing together many scholars for the exchange of ideas.

By: Liu Min

Editors: Qin Mian and Christopher Peter Clarke