Instance Segmentation Method Based On Several Attention Mechanisms

When the traditional axis-aligned bounding box detects objects with obvious rotation directions in natural images,especially in dense scenes,they often cannot fit the edges of the objects well.This problem may cause great interference to subsequent detection and segmentation,and affects the detection and segmentation performance of the instance segmentation model.This problem is more obvious in the object detection process of remote sensing images.For this reason,some remote sensing image object detection methods introduce the idea of oriented object detection.By redefining the rotated or inclined directional bounding box,the interference to the detection process is alleviated.In this paper,the idea of oriented object detection is extended to general natural images.Based on the original instance segmentation model Center Mask,an oriented instance segmentation model Ori-Center Mask is designed based on three attention mechanisms,including spatial attention,channel attention,and self-attention.Ori-Center Mask,the oriented instance segmentation model proposed in this thesis,mainly makes the following improvements: Firstly,the center-ness branch and the bounding box regression branch are optimized,and the original center-ness function is generalized in combination with direction perception,and a directional center-ness formula that can be applied to any direction is proposed.Secondly,through comparative experiments,a center-to-corner oriented bounding box prediction strategy with better localization performance is determined.Finally,an interactive branch based on the selfattention mechanism is added in parallel,which enhances the prediction performance of the model for oriented bounding boxes and segmentation masks of objects by outputting self-attention guided feature maps.Through multiple sets of ablation experiments,this paper fully demonstrates and analyzes the role and contribution of the main innovative work of this method in the basic instance segmentation model Center Mask.It confirms the performance improvement of Ori-Center Mask compared with the original instance segmentation model and the effectiveness of related work.