Discover the SciOpen Platform and Achieve Your Research Goals with Ease.
Search articles, authors, keywords, DOI, etc.
To address challenges of complex backgrounds and small targets in real-world waste classification, this study proposes an enhanced object detection framework integrating synthetic data generation with model architecture optimization. The generated trash classification dataset with 12 Categories (GTC12) was constructed by combining semantic descriptions from large language models and images synthesized via Stable Diffusion, employing a filtering mechanism to ensure high quality. Compared with existing public benchmarks, the synthetic dataset exhibits closer alignment with real scenes in color complexity and object-to-image area ratio. For model design, a dynamic receptive-field convolution attention module (RFAC) and a nonlinear mapping module were introduced based on Vision Transformer Adapter (ViT-Adapter) to strengthen feature extraction and robustness against complex backgrounds. On the hybrid "real+synthetic" dataset GTC12-Mix, the enhanced model’s mean average precision (mAP50:95) increased from 0.77 to 0.80; accuracy increased from 0.85 to 0.86 in fine-grained binary classification experiments on the class-balanced dataset GTC12-Bal. Results demonstrate high diversity and realism of the synthetic dataset, along with superior detection capability of the proposed model for complex backgrounds and small targets.
This is an open access article under the CC BY-NC-ND 4.0 license (https://creativecommons.org/licenses/by-nc-nd/4.0/).
Comments on this article