Fluid Inclusion Detection Based on Improved YOLOX

Fluid inclusions, which are minuscule cavities within minerals or rocks that contain trapped fluids, have significant implications in various scientific and industrial fields. Detecting these inclusions presents a unique challenge, especially when dealing with numerous small to medium-sized targets in densely packed photomicrographs. These conditions often lead to challenges in feature extraction and lower detection accuracy. To address this, our study introduces a refined fluid inclusion detection methodology leveraging the enhanced YOLOX algorithm. We utilize a sliding window technique to segment the photomicrograph, thereby improving the detection of smaller targets. Additionally, we have integrated the Convolutional Block Attention Module into the YOLOX algorithm’s Neck component. This integration enriches the feature representation by capturing both spatial and channel-wise attention, which significantly boosts localization accuracy. Furthermore, we have adopted the Smooth Intersection over Union Loss as our loss function, optimizing the training process for both speed and precision. Our experimental results showcase that this improved YOLOX algorithm surpasses its predecessor, demonstrating an increase of 1.8% and 1.9% in AP50 and AP50:95 metrics, respectively. These empirical findings underscore the effectiveness of the improvement method in enhancing the precision of fluid inclusion detection.