11:00 - 12:00 | Mon 28 Oct | The Great Room IV | MoC-T2
Deep learning-based object detection has recently received significant attention among scholars and practitioners. However, the acquired images often suffer from visual quality degradation under severe weather conditions, which could lead to negative effects on object detection in practical applications. Most previous studies proposed to implement object detection based on the assumption that image restoration techniques (e.g., image dehazing and low-light image enhancement, etc.) could improve visual quality while boosting detection accuracy. In contrast, we assumed that the image restoration techniques (i.e., image preprocessing) may also degrade the fine image details resulting in failing to promote object detection performance. In this work, according to the physical imaging process under severe weather conditions, we directly proposed to synthetically generate the degraded images with training labels to enlarge the original training datasets, which commonly contain only clear natural images under normal weather conditions. The advanced YOLOv3 model was then trained and tested on the enlarged dataset which contain both synthetic and realistic ship images generated under different weather conditions. Experiments have been conducted to compare the proposed method with other competing methods which implement training model only with clear images and testing model with (or without) image preprocessing. Results illustrated that our model could achieve superior detection performance under different conditions.
No information added