Aquaculture is a vital sector in global food production, with shrimp farming being a key branch within it. Accurate individual counting of shrimp fry is crucial for effective aquaculture management. However, traditional manual counting methods are time-consuming, labor-intensive, and prone to errors. Additionally, challenges such as mutual occlusion and varying lighting conditions in densely populated areas of shrimp fry can lead to inaccurate counting. In this paper, we propose the use of a high-resolution network, HRNET, to extract important information from images at multiple scales, mitigating information loss caused by external factors like lighting and overlapping due to mutual occlusion. By incorporating attention mechanisms, we enhance the feature representation of the backbone network for small targets, thereby improving recognition and counting accuracy. Furthermore, to address differences between dense and sparse regions, we employ dilated convolutions to expand the receptive field and replace the mean squared error (MSE) loss, which solely focuses on pixel-level average errors, with the DMS-SSIM loss function that considers similarity in subtle image structures and quality. Experimental results demonstrate a significant improvement in both mean absolute error (MAE) and MSE.

报告人简介：

龙伟，博士，硕导，湖州师范学院云顶集团优惠活动大厅教师，同济大学电子与云顶集团优惠活动大厅访问学者。主要研究智能信息处理、计算机视觉与感知智能、人机共融与智能应用等方向，目前在智慧水产养殖信息化方面有着较深的研究。先后主持及参与了国家自然科学基金、省级重点研发项目、省级自然科学基金及市级重点研发项目等课题，在Molecules、Frontiers in Psychology、Multimedia Tools and Applications、International Conference on Networking Systems of AI、《江苏农业学报》等国内外期刊和会议上发表高水平SCI、EI论文20多篇，发明专利、软件著作权共计20余项。

打印: