236 / 2025-06-15 15:39:37

Dual-Branch Pyramid Feature Fusion Algorithm for Small Object Segmentation

semantic segmentation,multi-scale,feature fusion,dilated convolution,image processing

      Current mainstream semantic segmentation algorithms still face issues such as small object loss, inaccurate edge segmentation, low segmentation accuracy, and inefficient fusion of deep and shallow feature information. Based on the PSPNet framework, a Dual-Branch Pyramid Feature Fusion Network is proposed. The standard convolution in the residual module of the backbone network ResNet50 is replaced by dilated convolution to enlarge the receptive field, aiming to enhance the attention to small objects. Additionally, a dual-branch interactive connection structure is employed to strengthen the interaction of local semantic information between branches. An improved progressive pyramid pooling module (M-PPM) is introduced in the main branch to enhance multi-scale contextual information, and an efficient channel attention mechanism module (ECA) is added to improve the model's feature representation capability. In the auxiliary branch, a small dilated pyramid module (H-SPN) that fuses high and low attention mechanisms is designed to enhance the extraction of low-level features. The upsampling method is modified to use a cubic interpolation algorithm to improve the smoothness of edge segmentation. The proposed method is experimentally evaluated on the mainstream PASCAL VOC2012 dataset, achieving an average intersection-over-union (IoU) of 84.46%, yielding state-of-the-art results. Compared with other methods, the proposed approach provides more accurate segmentation of small objects and addresses the issue of missing segmentations.