Vision-based underwater target real-time detection for autonomous underwater vehicle subsea exploration

Autonomous underwater vehicles (AUVs) equipped with online visual inspection systems can detect targets during operations, which is of great significance to subsea exploration. However, the undersea scene has some instinctive challenging problems, such as poor lighting conditions, sediment burial, and marine biofouling mimicry, makes it difficult for traditional target detection algorithms achieve online, reliable, accurate targets. To solve above issues, this paper proposes a real-time object algorithm based on lightweight convolutional neural network model. improve imaging quality images, contrast limited adaptive histogram equalization fused multicolor space (FCLAHE) model designed enhance image Afterwards, spindle-shaped backbone designed. The inverted residual block group convolutions are used extract depth features ensure accuracy one hand reduce parameter volume other under complex scenarios. Through extensive experiments, precision, recall, mAP proposed reached 91.2%, 90.1%, 88.3%, respectively. It also noticeable that method been integrated into embedded GPU platform deployed in AUV system practical average computational time 0.053s, satisfies requirements detection.