Attention-Enhanced Dual-Path Convolutional Neural Network with Ensemble Gradient Boosting for Content-Adaptive Bitrate Ladder Prediction in H.266/DASH Streaming

Adaptive bitrate streaming over HTTP (DASH) requires content-aware bitrate ladders to balance bandwidth utilisation and quality of experience, particularly for computationally demanding codecs such as H.266/Versatile Video Coding (VVC). This paper introduces MHCA-EGB (Multi-Head Cross-Attention with Ensemble Gradient Boosting), an integrated framework that systematically combines established attention-based fusion, multi-scale pooling, and ensemble classification techniques in a purpose-designed architecture that predicts Pareto-optimal bitrate resolution pairs for H.266-encoded video delivered over DASH by jointly modelling video content complexity and compression-domain artefacts. The proposed architecture extends the dual-path 2D–3D Convolutional Neural Network (CNN) paradigm with three key contributions: (i) a Multi-Head Cross-Attention fusion module that replaces naïve channel concatenation, enabling learned bidirectional feature alignment and content-adaptive emphasis of quality-discriminative representations; (ii) a Temporal Pyramid Pooling layer that captures multi-scale temporal dynamics from short-burst motion to long-range scene transitions at 2-frame, 4-frame, and 8-frame granularities; and (iii) a stacked ensemble classifier combining XGBoost, LightGBM, and CatBoost with a logistic regression meta-learner for robust bitrate cluster assignment. Comprehensive evaluation on 101 diverse 4K Ultra-High-Definition (UHD) sequences from four benchmark datasets, encoded with VTM 22.0 at seven resolutions and eleven quantisation parameter (QP) values (7777 total encodes), demonstrates that MHCA-EGB achieves an average Bjøntegaard Delta Rate (BD-Rate) of −5.47% relative to the exhaustive convex hull (a value attributable to the BD-Rate polynomial fitting methodology when the predicted and reference ladders use overlapping operating-point subsets; all predicted points are physically encoded and verified) while reducing encoding time by 67.2% with only 0.12 Video Multi-Method Assessment Fusion (VMAF) regret (fifty times below the just-noticeable difference threshold). Ablation analysis confirms that cross-attention fusion contributes the largest novel gain (+1.42% BD-Rate), followed by temporal pyramid pooling (+0.79%) and ensemble stacking (+0.55%), with a Pearson correlation of ρ=0.87 between content complexity and BD-Rate magnitude confirming that the framework delivers the greatest value on high-complexity premium content.