Bag of Strategies Set New State-of-the-art for Algae Object Detectors

Deep learning-based detection of marine microalgae in natural waters can meet the need for rapid monitoring, facilitating researchers in marine and environmental sciences, while also paving the way for downstream cellular analysis tasks. We use a new training scheme for marine microalgae detection that consists of two phases: a teacher benchmark model phase and a student model learning phase. Using teacher model supervision to get better student model training results. Through a simple and fast image fusion method, we can obtain more realistic algae-generated images to extend the training set and eventually improve the convergence speed and performance of the model. Based on the algorithms of YOLOv5 and YOLOv6, we use the DHLC backbone network fusion method to fuse features from different levels of C3 modules and BepC3 modules together as the input of the PANet middle layer. We also use the module in BoTNet network to obtain stronger feature extraction capability by introducing self-attention mechanism in the yolo model. Since there are many small targets in marine microalgae images, we also extend the YOLOv6l model to the more powerful YOLOv6l-P6 model, which can get better detection results in the input image size of 1280. In addition, we also use time-test augmentation (TTA), weighted boxes fusion (WBF) and Single-class wighted boxes fusion (SinWBF) techniques to optimize the performance of each class. These strategies greatly improve the model detection performance and robustness under the conditions of small amount of marine microalgae microscopic image data. Finally our solution won the first place on the “Vision Meets Algae” Object Detection Challenge, and got 58.25 MAP.