A study was conducted on image segmentation obtained by different types of sensors on Unmanned Surface Vessel(USV),combining the image segmentation large model with low-cost fine-tuning techniques to enhance the practicality of large models in unmanned systems.This article first applies the Segment Anything Model(SAM)to the segmentation tasks of visible light,infrared,and sonar images of USV.The comparison results show that SAM has good segmentation performances in visible light and infrared images,but there is a serious decline in sonar image segmentation performance.Subsequently,this article used Low-rank Adaptation(LoRA)algorithm to fine tune the SAM encoder to adapt to sonar image segmentation tasks.The segmentation performance of the fine tuned SAM on sonar images has been significantly improved,with Dice improving by 52.88 compared to the initial SAM,proving the effectiveness of the method.Compared to the scheme of retraining all parameters of the large model,LoRA uses low-rank approximation to reduce the dimensionality of the weight matrix,thereby significantly reducing the number of trainable parameters in the model.It performs better in terms of training efficiency and cost reduction.
维普
万方数据