双向注意力特征调制的路面坑槽分割方法OA

Potholes are one of the major pavement diseases,which raise driving safety hazards.It is of great importance to detect potholes in a timely and accurate manner during vehicle driving.Aiming at the segmentation challenges caused by the low contrast between pavement potholes and their surrounding areas and the insignificant features of small-sized potholes,this paper proposes a lightweight bi-directional attention feature modulation network(BAFMNet).This network embeds a hierarchical feature transformation neck(HFTN)between the backbone and the decoding head.By taking advantage of the complementary strengths of features in adjacent stages,HFTN achieves efficient feature refinement for each encoded feature.To provide comprehensive object cues for the subsequent decoding head,a cross-stage information complementation(CSIC)module is introduced between the features of two adjacent stages in HFTN.As the core of the CSIC module,a novel bi-directional attention feature fusion(BAFF)module drives affine transformation through the cross-attention mechanism,enabling in-depth interaction of semantic features in adjacent stages.Experimental results show that compared with other methods,the proposed method achieves the best segmentation accuracy with the smallest model parameters and fast inference speed.