In human–robot hybrid intelligent warehouses, pallets often come in various shapes and sizes, posing challenges for AGVs to automate pallet picking. This, in turn, reduces the overall operational efficiency of the warehouse. To address this issue, this paper proposes a lightweight component segmentation network using a dual-attention mechanism to achieve precise segmentation of the pallet’s stringer board and accurate localization of the pallet slots. To overcome the challenge of redundant computations in existing semantic segmentation models, which are unable to balance spatial details and high-level semantic information, this network utilizes a dual-branch attention mechanism within an encoder–decoder architecture to effectively capture spatial details. On this basis, a residual structure is introduced to reduce redundant network parameters, addressing issues like vanishing and exploding gradients during training. Due to the lack of a public pallet image segmentation dataset, the network was tested using a custom-made dataset. The results show that by extracting intermediate-, low-, and high-level features from dual-branch input images and integrating them to construct multi-scale images, precise segmentation of various types of pallets can be achieved with limited annotated images. Furthermore, to comprehensively evaluate the model’s robustness, additional pallet localization experiments were conducted under varying illumination conditions and background noise levels. The results demonstrate that the proposed method can effectively identify and locate multi-category pallet targets while maintaining high segmentation accuracy under different lighting conditions and background interferences, verifying the model’s robustness in complex warehousing environments. Compared to the traditional model, the proposed model in this paper achieves a 10.41% improvement in accuracy and a 32.8% increase in image processing speed. The segmentation network we proposed is used for pallet-positioning experiments and has achieved good positioning results in pallet images taken from different distances and angles.