TC-YOLOv5: rapid detection of floating debris on raspberry Pi 4B

Floating debris is a prominent indicator in measuring water quality. However, traditional object detection algorithms cannot meet the requirement of high accuracy due to the complexity of the environment. It is difficult for some deep learning-based object detection algorithms to achieve fast detection due to the limited performance of embedded devices. To address the above issues, this paper proposes TC-YOLOv5, which improves the detection accuracy by integrating the convolutional block attention module and vision transformer. To ensure the efficiency and low resource consumption of the algorithm, we selectively remove some convolutional layers and reduce some redundant calculations. We evaluated the performance of TC-YOLOv5 on a dataset with multiple species of floating debris, which can process an image in an average of 1.18 s on a Raspberry Pi 4B and achieve the mean average precision (mAP@0.5) of 84.2%. The detection accuracy, speed, and floating-point operations (FLOPs) of TC-YOLOv5 are better than some algorithms of the YOLOv5 series, such as YOLOv5n, YOLOv5s, and YOLOv5m. The above data show that TC-YOLOv5 realizes high-precision, low resource consumption, and rapid detection.