Multilayer feature fusion and attention-based network for crops and weeds segmentation

Distinguishing weeds from crops is a critical challenge in agriculture, with the existing agriculture semantic segmentation networks simply combining low-level with high-level features at the encoder and decoder stages to improve performance. However, a simple low-level and high-level feature fusion may not be effective due to the semantic and spatial resolution gap. Hence, this paper proposes a novel dual attention network (DA-Net), based on branch attention blocks in the encoding stage and spatial attention blocks in the decoding stage, to bridge the gap between low-level and high-level features. Our method first adds a branch selection module at the residual connection between the encoder and decoder, enabling low-level futures to select higher-level features for fusion adaptively. Then, a cascaded convolution block utilizing asymmetric convolution is constructed, supporting the receptive field’s expansion without increasing the computational burden or the parameter cardinality. We design a spatial attention block in the fusion stage to capture rich contextual dependencies. Finally, we construct a novel block named densely channel fusion, which utilizes a sub-pixel layer to encode most channel information into spatial information. The experimental results demonstrate that DA-Net is superior to ExFuse, Ddeeplabv3 + , and PSPNet on three public datasets, with each added component significantly affecting the overall performance.