EDCST-Rain: Enhanced Density-Aware Cross-Scale Transformer for Robust Object Classification Under Diverse Rainfall Conditions

Rain degradation significantly impairs object classification systems, causing accuracy drops of 40-60% under severe conditions and limiting autonomous vehicle deployment. While preprocessing approaches attempt deraining before classification, they suffer from error propagation and computational overhead. This paper introduces EDCST-Rain, an Enhanced Density-Aware Cross-Scale Transformer specifically designed for robust classification under diverse rain conditions. The architecture consists of five integrated components: a Rain Density Encoding Module that captures rain streak density, accumulation, and orientation; a Swin-Tiny Backbone for hierarchical feature extraction; and three rain-specific mechanisms: directional attention modules adapting to rain streak orientation, accumulation-aware processing handling lens droplet distortions, and adaptive cross-scale fusion integrating multi-resolution information. We develop a comprehensive physics-based rain simulation framework covering four rain types (drizzle, moderate, heavy, storm) and implement a curriculum learning strategy that progressively introduces rain complexity during training. Extensive experiments on CIFAR-10 demonstrate that EDCST-Rain achieves 83.1% clean accuracy while maintaining 71.8% under severe rain (86.4% retention), representing a 10-percentage-point improvement over state-of-the-art methods. With 15.8 million parameters and a 14.3 ms GPU inference time, enabling real-time operation, EDCST-Rain provides a practical, weather-robust perception framework suitable for autonomous systems operating under adverse weather conditions.