This study proposes the DLF Multi-Scale Network (DLF) for recognizing Big Five personality traits from multimodal (visual/audio) video features, addressing limitations in feature stability and cross-modal correlation inherent in traditional methods. The model integrates a Discrete Wavelet Transform-CNN (DWTC)module that decomposes video frames into multi-scale components(e.g., facial expressions, body movements) to extract robust visual representations, alongside an MLP for processing audio spectral data; a Learnable Multi-head Self-Attention (LMSA) mechanism then aligns cross-modal features to resolve spatiotemporal inconsistencies before decision-level fusion. Evaluated on the First Impressions V1 dataset, DLF-MSN achieves 0.9177 average accuracy, outperforming baselines. This work advances behavior-personality mapping with applications in mental health assessment and personalized recommendation systems.