Article: Real-time image/surveillance waste sorting via MatMul-free-based encoder-decoder learning structure Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.68 pp.91 - 107 Abstract: The amount of garbage continues to rise, making intelligent garbage classification increasingly important for future resource recovery. Current methods still rely largely on static images, which perform poorly in dynamic real-world settings. Moreover, practical applications such as surveillance cameras or mobile inspections face additional challenges including computational efficiency, scene diversity, and long-term robustness, which traditional approaches cannot adequately address. This paper presents a real-time garbage classification framework suitable for both image and video surveillance. We design an encoder-decoder structure that eliminates matrix multiplication, significantly reducing computational cost. Additionally, we introduce a dynamic tanh (DyT) layer to enhance normalisation, replace the traditional feedforward module with a Kolmogorov-Arnold network (KAN) for better interpretability of features, and employ dense layers without matrix multiplication to further boost efficiency. Experiments demonstrate that our method achieves an effective balance of accuracy, computational cost, and robustness, making it well-suited for complex, dynamic garbage detection scenarios. Inderscience Publishers - linking academia, business and industry through research

Title: Real-time image/surveillance waste sorting via MatMul-free-based encoder-decoder learning structure

Authors: Ruoxi Cui

Addresses: Ulster College, Shaanxi University of Science and Technology, Xi'an, 710021, China

Abstract: The amount of garbage continues to rise, making intelligent garbage classification increasingly important for future resource recovery. Current methods still rely largely on static images, which perform poorly in dynamic real-world settings. Moreover, practical applications such as surveillance cameras or mobile inspections face additional challenges including computational efficiency, scene diversity, and long-term robustness, which traditional approaches cannot adequately address. This paper presents a real-time garbage classification framework suitable for both image and video surveillance. We design an encoder-decoder structure that eliminates matrix multiplication, significantly reducing computational cost. Additionally, we introduce a dynamic tanh (DyT) layer to enhance normalisation, replace the traditional feedforward module with a Kolmogorov-Arnold network (KAN) for better interpretability of features, and employ dense layers without matrix multiplication to further boost efficiency. Experiments demonstrate that our method achieves an effective balance of accuracy, computational cost, and robustness, making it well-suited for complex, dynamic garbage detection scenarios.

Keywords: waste sorting; image; MatMul-free; transformer; dynamic tanh; DyT; Kolmogorov-Arnold network; KAN.

DOI: 10.1504/IJICT.2026.154347

International Journal of Information and Communication Technology, 2026 Vol.27 No.68, pp.91 - 107

Received: 06 Jan 2026
Accepted: 17 Mar 2026
Published online: 23 Jun 2026 *

Title: Real-time image/surveillance waste sorting via MatMul-free-based encoder-decoder learning structure

Keep up-to-date