Ship Image Segmentation Based on Multi-dimensional Information and Feature Fusion

Ship images in inland waterway scenarios have the characteristics of fixed background and continuous image. Therefore, motion detection is a feasible preprocessing method to mine the dataset information. For the feature extraction problem of RGB image and motion information, an asymmetric two-branch structure is adopted to perform convolution operation, respectively. For the problem of the scale gap between ship targets, we design a multi-scale convolutional attention module, which is used to capture the features of targets of different scales by setting up different receptive fields. To fuse features of the two branches, an adaptive feature fusion module is designed and inserted into both the encoder and decoder, which can effectively fuse the features of two branches. Experimental results show that the segmentation method proposed in this work can reach 94.88% of mIoU in the ship image dataset, compared to 92.88% of U-Net.