Pellet ore is a crucial raw material for modern blast furnace ironmaking, and an appropriate particle size distribution (PSD) is one of the key factors affecting its quality. Currently, the detection of the PSD in pellet ore production mainly relies on offline manual sieving, which is inefficient, provides delayed results, and cannot meet the real-time requirements of industrial production sites. Existing image processing-based particle size detection methods have limitations in accurately detecting the edges and contours of densely packed pellets, often leading to missed detections or over-segmentation. In light of this, a dense pellet size measurement model integrating dual attention mechanisms and multi-scale cascaded networks, termed light multi-residual attention dual-UNet (LMAD-UNet), is proposed. The proposed method concatenates two UNet architectures in a parallel cascading manner to form the backbone network, thereby increasing network width and reducing feature loss during downsampling. Subsequently, a lightweight multi-scale fusion module, referred to as the lightweight multi-residual block (LMuti-Res Block), is designed to facilitate multi-scale feature extraction, minimize model parameters, and consequently enhance inference speed, while a dual attention mechanism is introduced to enhance contour feature extraction. The loss function is also improved to strengthen the learning of imbalanced contour point data. Experimental results show that the proposed method can accurately segment densely packed pellets, achieving an $F_1\text{-}{\rm score} $ of 96.85%, outperforming other comparison methods, and meeting the real-time requirements of the production process.