BPDF-SegNet: A Bidirectional Perception and Dynamic Fusion Segmentation Network to Detect Ball Screw Small Pitting Defects

It is a challenging task to detect small damages on ball screw surface by using detection using computer vision techniques due to low contrast backgrounds and significant texture interferences. Current segmentation techniques based on U-Net uses the traditional feature fusion mechanism in the decoder, which makes it challenging to reliably extract subpixel-level edge details, hence leading to edge blurring or local breaks in the segmentation mask. Due to loss of detail information, it causes the issue of missed and false detections. This research proposes BPDF-SegNet (Bidirectional Perception and Dynamic Fusion Segmentation Network) as a solution to this issue. Firstly, a bidirectional decoupled region calibration module is introduced to improve the model's capability to focus on foreground targets. Secondly, the feature maps are effectively up-sampled using deep convolution and channel shift techniques to make over for any lost features. Thirdly, the feature fusion process facilitates the effective fusing of the feature maps, enabling the network to better capture complementary information across different layers. Lastly, the focal loss function and Tversky loss function are integrated during the model training stage to reduce the impact of sample imbalance. Experiments conducted on the ball screw dataset demonstrate that the approach produces improved segmentation outcomes and offers a fresh solution for detecting industrial metal surface flaws.