Algae detection is a specialized object detection that identifies multi-scale algae in microscopic aquatic environments. Due to the high density, similar appearance, and background interference of raw water samples, existing methods often detect algae from clean images and may face limitations in (a) directly handling raw water images with impurity interference (e.g. suspended particles, bubbles, and other microorganisms), and (b) accurately capturing multi-scale algae features from dense and overlapping background. To address these challenges, we introduce an end-to-end Lite Real-Time Detection Transformer (LRTDT), which incorporates three components: (a) a Multi-Channel Downsample Block (MCDB) designed to enhance hierarchical features by channel-wise feature separation and subsequent fusion module to mitigate interference of impurities; (b) Hierarchical Fusion Encoder (HFE) that simplifies the fusion of multi-scale feature through a lightweight architecture to improve the multi-scale feature extraction; and (c) a Generalized Algae Intersection over Union (GAIoU) which incorporates a feature overlap term into GIoU loss to enhance focus area of feature in complex background. Extensive experiments on public benchmark datasets and a self-collected High Quality Algae dataset (HQAD) demonstrate the performance and generalization of LRTDT. Source code, dataset and pre-trained models are available at https:/github.com/he13689/LRTDETR.
He et al. (Fri,) studied this question.