SE-YOLOv8: Small-Object Enhanced YOLOv8 for Visual-Based Reaction Yield Prediction of Catalytic Processes

Abstract

Reaction conversion, catalyst dispersion, precipitate production, and local colour changes are all included in the collection of visual data from the catalytic process. Determine tiny areas that are susceptible to yield variations in order to accurately estimate the reaction yield while maintaining a stable regression. In this study, a small-object improved YOLOv8 framework for catalytic reaction yield prediction is developed. The suggested approach is a comprehensive approach that integrates yield-oriented regression, adaptive multi-scale fusion, small-target recovery, and region-aware attention within an experimental evaluation framework. Improve yield estimate in noisy picture situations and lower the missed detection rate of weak catalytic regions. In comparison to the baseline YOLOv8-regression model, the suggested model has improved mAP by 4.8% points, decreased MAE from 6.42% to 3.91%, and decreased RMSE from 8.35% to 5.27%, as demonstrated in the comparative experiment. An applied deep visual model for intelligent catalytic process analysis is called Framework.