Feature augmentation and scale penalty for tiny floating detection

Rapidly increasing concerns about the impact of tiny floating objects on water health has prompted the need for more effective detection methods. The main challenge in detecting these objects is their small size, accounting for only 0.5% of the image, which significantly hampers detection efforts. Moreover, existing object detectors utilize the intersection over union (IOU) as the bounding box regression loss to enhance object localization accuracy. However, this approach penalizes larger objects more heavily than smaller ones, leading to imbalanced regression losses. To address these issues, we propose enhancements to the YOLOv4 model. Our approach incorporates the following key improvements. Firstly, we introduce a feature augmentation module (FAM) to capture multi-scale contextual features of tiny objects and low-level features. This helps overcome the challenge of limited representation of tiny objects in the deeper layers of the network. Additionally, we integrate a convolutional block attention module (CBAM) into the path aggregation network to prevent the flooding of conflicting information in the fusion of features at different levels, ensuring an accurate representation of tiny object features. Finally, we propose a scale penalty function to address the issue of imbalanced regression loss. Experimental results demonstrate that our improved model achieves impressive detection performance on the Flow-RI dataset, specifically for detecting small-scale objects. These findings highlight the efficacy of our proposed methodology in enhancing the detection of tiny floating objects and contribute to the overall goal of improving water health.