A Semantic Segmentation and Distortion Correction of Fragment Perforation Small Target

Abstract

In order to solve the superimposed image distortion caused by explosion shock wave and shooting tilt in warhead explosion test, and the problem of small target of the fragment perforation missed detection, a method combining image correction and small target semantic segmentation is proposed. Based on mesh planarization and homomorphic inverse projection transformation, the distortion of fragment perforation image is corrected. For the reliability and accuracy of small fragment perforation segmentation, based on a encoder-decoder structure, a DSA-EF(Double Self-Attention and Embedded Fusion) semantic segmentation net is proposed. In the encoder part, swin transformer block is as the backbone network to extract the multi-scale features, especially the small fragment perforation features, and the receptive field expansion block is used to fuse the multi-scale features in order to enhance features representation. In the decoder part, the channel attention mechanism of ECA block is introduced to consider the correlation between different channels of multi-scale target feature, and remove interference from redundant information, while retain shallow small fragment perforations. Through simulation experiments, this method can effectively correct the superimposed image distortion of the target plate. The ablation experiments verify the specific impact of each component on model performance, and proved the effectiveness of each modules for the entire method. Compared with five semantic segmentation methods of FCN and Seg Former et al., it shows that this method has better segmentation performance, effectively solves the problem of small fragment perforation target recognition, and improves the segmentation reliability and accuracy of small target.

Keywords

• Image correction
• small target detection
• attention mechanism
• image segmentation
• area calculation