Addressing the insufficient interaction and integration of multi-band EEG information, the low multi-class recognition accuracy caused by single feature extraction methods, and the difficulty in simultaneously capturing global features and local details of signals, this paper proposed a dual-stream complementary feature emotion recognition network based on heterogeneous information density visualization (HIDV-DSCF-ENet) . First, the study developed a heterogeneous information density visualization method that converted preprocessed multi-band EEG signals into two types of visualization images with complementary information densities: high-density information encoded cross-band coupling information and spectral interaction features using RGBA four-channel encoding, while low-density information encoded intrinsic structural features within frequency bands using grayscale single-channel encoding. Then, the paper designed an asymmetric dual-stream feature extraction structure to process interactive semantic features of high-density information and discriminative structural features of low-density information separately. Finally, the study designed a heterogeneous feature interaction function based on dual learning mechanism to learn the complementary characteristics between the two information densities and achieve dual learning heterogeneous feature fusion. Experiments on the DEAP dataset demonstrate that the proposed method achieves accuracies of 98.17% and 97.88% for 4-class emotion recognition based on valence-arousal labels and 8-class emotion recognition based on valence-arousal-dominance labels, respectively, outperforming existing methods and providing new insights for EEG-based emotion recognition research.