Lightweight context-awareness hybrid-attention network for waste segmentation in cluttered scenes

With the acceleration of urbanization, municipal solid waste is increasing at an alarming rate, posing a significant obstacle to achieving sustainable development. On the one hand, improper disposal of hazardous waste causes environmental pollution. On the other hand, inefficient sorting of recyclable waste results in resource waste. Therefore, automatic waste sorting systems based on computer vision have received more attention. To achieve waste segmentation in a cluttered industrial environment, this paper proposes a lightweight context-awareness hybrid-attention network, which is suitable for industrial terminal devices with limited resources. Specifically, we introduce an efficient spatial cascade module based on the multi-branch architecture, which can extract richer spatial features under different receptive fields. In addition, we use a plug-and-play feature enhancement module based on the Transformer architecture, which can effectively model long-range dependencies and enhance important information. At the same time, we use the channel shuffle operation to achieve information exchange between different groups. To fuse detailed information and semantic features, we design a novel semantic fusion module. It not only uses a spatial awareness module to extract multi-scale features, but also uses a channel awareness module to enhance critical features. Experimental results show that our model outperforms other methods. It not only achieves satisfactory segmentation results, but also has fewer model parameters.