Abstract In order to improve the speed and keep high accuracy of UUVs object detection on embedded device. This paper proposes an acceleration strategy based on key frame extraction and model compression. Firstly, selecting shipwreck as target and creating a real-world underwater image dataset of the shipwreck. And this work applies both online and offline data augmentation to accelerate the model convergence and to improve the model generalization. Then, this paper builds a fusion algorithm with structural similarity, color histogram and image entropy to select the key frame of the captured video for removing the redundant information, and analyzing the reason for the complexity of the neural network and the method of model compression. Finally, this work cuts off unimportant structure of network based on channel and layer pruning to reduce model complexity and keeps the model with high accuracy by fine-tuning. The experiment shows the Billion Floating Operations of the compressed YOLOV4 model is reduced to 15.778. On the Nvidia Jetson TX2 embedded image processor, the average detection speed for video with 608 × 608 resolution can reach 15.12FPS. And the detection speed of the video composed of key frames is 2.98 times faster than the raw video.