Transformer architectures have transformed computer vision since their introduction, replacing or augmenting convolutional neural networks with self-attention mechanisms that let a model weigh relationships between distant parts of an image. For road garbage recognition, this global context is valuable: a crumpled plastic bag half-hidden at the edge of a curb can be distinguished from a patch of fallen leaves only by considering the surrounding scene. The problem is computational cost. Self-attention scales poorly with the number of image tokens, and full-size vision transformers demand memory and processing power far beyond what an embedded board mounted on a cleaning vehicle can supply. Earlier approaches to garbage detection relied on convolutional networks or heavyweight transformers that either sacrificed contextual understanding or required cloud-level hardware, neither of which suits a camera bolted to a truck sweeping a street at dawn.

The new model, developed by Suheng Peng, Jiacai Liao, Libo Cao, Cong Duan and Minghai Zhang at Hunan University and Changsha University of Science and Technology, builds on SegFormer, a transformer architecture already designed with efficiency in mind. SegFormer uses a hierarchical encoder that processes the image at multiple scales and a simple, all-MLP decoder, avoiding the most expensive parts of classic transformer designs. But even SegFormer’s attention blocks can strain an embedded processor. The team’s central contribution is a module they call Residual Multi-axis Hadamard Attention, which reorganizes how the network computes attention so that it remains expressive while dramatically cutting the arithmetic burden.

The key idea is to split the feature representation into groups along different axes of the image and then fuse information across those groups using element-wise multiplication, the operation mathematicians call the Hadamard product. Instead of letting every image token attend to every other token, which grows quadratically with image size, the module promotes what the authors describe as complementary feature interactions across multiple axes through residual cross-axis feature-weighted fusion. In practical terms, features computed along the horizontal axis of an image are weighted and combined with features computed along the vertical axis, and a residual connection preserves the original information through the operation. This enhances the sharing of information between groups without reconstructing the full attention matrix, keeping computational complexity low while still allowing the model to build an efficient representation of the scene’s context.

Attention redesign alone, however, does not address a second, subtler problem in garbage segmentation: class imbalance and category confusion. Roads are mostly clean surfaces, and the objects that litter them come in wildly varying shapes, sizes and frequencies. Small, thin items such as bottles or paper scraps occupy few pixels, while large debris dominates. A segmentation network can easily learn to favor the frequent background classes and blur the rare or visually similar garbage categories together. To counter this, the researchers propose a Feature Memory module, a component that stores and continuously updates a memory of category distribution information as the network processes images. By consulting this memory, the model can reinforce the characteristic features of each garbage category and improve its feature representations for classes that appear infrequently or look alike, effectively giving the network a running summary of what it has learned about each type of trash.

The two modules were integrated into a lightweight SegFormer backbone and evaluated on the RK3568, a low-power embedded development board of the sort that could realistically be installed on municipal cleaning vehicles. The results were striking on both fronts. Inference speed increased by 12.48 frames per second compared with the baseline, a meaningful margin for a real-time system that must keep pace with a moving vehicle, and the mean intersection-over-union score, the standard accuracy metric for semantic segmentation, improved by 4.54 percent. The model outperformed other semantic segmentation approaches tested under the same conditions, indicating that the gains were not simply a matter of trimming the network until it was fast but inaccurate. The combination of higher speed and higher accuracy on constrained hardware is precisely the trade-off that most lightweight architectures struggle to achieve.

Beyond segmentation itself, the team also revisited how street cleanliness should be measured. Cleanliness assessment has historically been a subjective exercise; earlier studies have compared municipal scorecards with citizen surveys and proposed quantitative indices for specific cities, but these approaches are labor-intensive and inconsistent. The researchers reformulated the road cleanliness assessment metric used in their pipeline to improve the accuracy and stability of the overall assessment system. With a segmentation model that reliably delineates garbage regions frame by frame, a well-defined metric can convert those pixel-level predictions into a stable, reproducible cleanliness score, turning a fleet of camera-equipped vehicles into a continuous environmental monitoring network rather than an occasional inspection tool.

The practical implications extend across urban management. Cities spend substantial budgets on street sweeping, and decisions about where to deploy cleaning resources are often guided by sparse manual inspections. A real-time, edge-deployed system that segments garbage, classifies it by category and produces a quantitative cleanliness index could allow routing decisions to be made on current data, direct vehicles to the dirtiest streets first, and provide objective evidence of service quality. The authors position the work as bridging the gap between high-accuracy transformer-based models and practical edge deployment, offering what they describe as a cost-effective solution for real-time environmental monitoring. Because the model runs on an inexpensive development board rather than a server, the marginal cost of adding intelligence to each vehicle is low.

The study also sits within a broader research movement applying deep learning to waste management, from convolutional detectors for recycling facilities to vision transformers for waste classification and multimodal systems for plastic identification. What distinguishes this contribution is its explicit focus on the deployment constraint: rather than chasing leaderboard accuracy on powerful GPUs, the authors optimized for the hardware that municipal systems can actually afford to install. The datasets generated during the study are not publicly available because they form part of an ongoing investigation, though the authors state they are available from the corresponding author on reasonable request. If the approach proves robust in field trials, the sight of a street-sweeping truck that not only cleans the road but also maps, counts and reports every piece of litter it passes may soon become a routine feature of the smart city.

Subject of Research: Lightweight transformer-based semantic segmentation for real-time road garbage detection and cleanliness assessment on edge devices

Article Title: Transformer-based lightweight model and category enhancement for road garbage segmentation

Article References: Peng, S., Liao, J., Cao, L., Duan, C., & Zhang, M. (2026). Transformer-based lightweight model and category enhancement for road garbage segmentation. Neural Computing and Applications, 38(19), Article 792. https://doi.org/10.1007/s00521-026-12554-6

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12554-6

Keywords: transformer, semantic segmentation, road garbage detection, edge computing, lightweight model, SegFormer, attention mechanism, cleanliness assessment, computer vision, waste management, embedded deployment, deep learning