Dual Stream Cross Domain Feature Fusion for Land-Oceanic Object Detection

Unmanned platforms are equipped with dual-light sensor, but the detectors are usually trained on visible-light images. For land-oceanic objects, it is challenging to achieve the requirements of multiperiod and all-weather detection. To solve this problem, we combined with Siamese neural network and detector, the land-oceanic object detection based on dual stream cross domain feature fusion method was proposed. Firstly, the two streams model was constructed based on backbone of yolov4. The model uses two streams to extract the feature of visible-light image and infrared image respectively; then, the adaptive weight redistribution module under different lighting environments was proposed; finally, we compared three feature fusion strategies of early fusion, middle fusion and late fusion. Extensive experiments are carried out on various benchmark datasets with different methods. The experimental results show that early fusion, middle fusion and late fusion have achieved better results than other methods, among them, middle fusion has achieved best results.