报告题目：Advances in Underwater Optical and Sonar Image Enhancement and Quality Assessment
报告时间：2024年10月19日 上午 10：00
报告地址：广东技术师范大学东校区工业中心605
报 告 人：任金昌 教授
报告人简介：
任金昌，罗伯特戈登大学计算机科学系教授、英国国家海底中心（NSC）高光谱成像实验室主任、国际研究生院主任、Transparent Ocean 计划负责人、IEEE高级会员、Marquis Who's Who终身成就奖获得者。他毕业于中国西北工业大学，先后获得计算机软件学士学位、图像处理硕士学位和计算机视觉博士学位，后在英国布拉德福德大学获得电子成像与媒体传播博士学位。 任金昌教授的研究领域包括高光谱成像、图像处理、计算机视觉、大数据分析和机器学习。多年来，他主持了多项国家级和国际合作重大科研项目，研究总经费超过350万英镑。任教授已经发表了350多篇国际高水平期刊/会议论文，并担任多个国际期刊的副主编，包括IEEE TGRS和J. of the Franklin Institute等。此外，他还主持和共同主持了多项国际会议和研讨会。他指导的学生在多个会议上屡屡获奖，包括IET图像与视觉的最佳博士论文奖和其他会议/研讨会。
报告摘要：
In underwater environments, imaging devices with optic sensors face numerous challenges including water turbidity, light attenuation, scattering, and the presence of particles, which collectively degrade image quality, reduce contrast, and distort colours. To mitigate these issues, SOund and NAvigation Ranging (SONAR) sensors are often employed to capture information using sound pulses reflected from the scene. This imaging technique serves as a valuable complement in scenarios with poor lighting conditions. However, sonar images also exhibit limitations such as lower resolution, susceptibility to environmental factors like salinity and temperature, vulnerability to underwater currents and noises from marine life, and the presence of shadow zones, complicating the differentiation between original objects and their associated shadows.
Despite these limitations, the fusion of these two modalities can yield highly informative results. Consequently, researchers have concentrated on developing quality assessment techniques to ensure the acquisition of high-quality data, supplemented by enhancement methods to further refine data quality. Enhancements encompass haze removal, contrast and resolution enhancement, and enhancement of colour distribution in optical images. Additionally, for sonar images, emphasis is placed on noise reduction and shadow removal to improve overall clarity and interpretability.
In this talk, we will discuss the devised quality assessment techniques for both image types utilizing machine learning algorithms trained to correlate images, mapped within a predetermined feature space, with corresponding quality scores. Our experimental findings have demonstrated the effectiveness of these methods, exhibiting strong correlation with human evaluations. Furthermore, to facilitate enhancement, we have introduced innovative deep neural network architectures enriched with attention-driven inception modules and autoencoders. Through experimentation, we have showcased the efficacy of our developed networks in enhancing image quality, mitigating noise, and demonstrating robust generalization capabilities across publicly available datasets. Finally, applications of the developed techniques to tackle real challenges in offshore energy, subsea operations, military and environmental sectors are demonstrated.
