MFFP-Net: Multi-directional Feature Fusion and Position-Aware Network

Vehicle re-identification faces significant challenges in feature modeling due to viewpoint variations, illumination changes, occlusions, and intra-class similarity. To this end, we propose a VeRi algorithm based on multi-directional feature fusion and position awareness, which synergistically integrates multi-dimensional features and positional information to suppress background interference and improve feature discriminability. Specifically, we design multi-directional depthwise separable convolution kernels which square kernels for local details, horizontal strip kernels for long-range dependencies, and vertical strip kernels for spatial distribution to capture comprehensive directional features; the DOConv module fuses depthwise convolution and conventional convolution to balance fine-grained texture extraction and global structural integration without additional sub-networks; and a position encoding module with horizontal-vertical tensor initialization enhances the model's perception of key components' relative positions. Experimental results on the VeRi-776 and Veri-Wild datasets demonstrate the superior performance of the proposed algorithm.