| Aero-engine lockwire segmentation is the key step to realize the detection of its winding and twist braid density. Due to the complex background, uneven illumination and small percentage of the target region, the identification precision of aero-engine lockwire is low. This paper propose an improved Mask R-CNN model for lockwire instance segmentation. Firstly, the gamma correction of R, G and B channels is carried out to transform the lockwire image into pseudo-color image and enhance the contrast. Then, the dynamic snake-shaped convolution was incorporated into Resnet, the backbone network of Mask R-CNN, to make the network to adaptively focus on the slender and curved local structure during feature extraction. Finally, the CBAM attention mechanism was introduced in the feature fusion phase, in order to improve the perception of small target and preserve the shallow features of small target. The experimental results show that, the mask and?bounding box AP50 of the improved module was 82.54% and 93.13% . Compared to basic mode, the AP50 was improved by 5.83% and 3.47%, witch proves the proposed method can improve the segmentation accuracy and reduce the missed and false detection. |
