<p>Premature ventricular contraction (PVC) is a common cardiac arrhythmia that requires accurate and timely detection for effective clinical management. Although convolutional neural networks (CNNs) have demonstrated significant potential in the automated analysis of electrocardiogram (ECG) signals, developing a PVC detection model with high generalization to unseen patients remains a major challenge due to the inherent complexity and variability of ECG signal morphology. In this study, we propose a novel dual-path fusion network (DPFNet) by integrating attention mechanism for robust PVC detection. DPFNet employs a dual-path CNN architecture that simultaneously extracts features from two complementary data representations: one-dimensional (1D) time-domain heartbeat sequences and two-dimensional (2D) wavelet scalograms. The first path utilizes multi-scale convolution and depthwise separable convolution to process 1D signals, effectively capturing temporal dynamics. The second path analyzes 2D scalograms by combining multi-scale convolutions, strided convolutions as an alternative to traditional pooling layers, and integrates our proposed 2D residual convolutional block attention module (2D-Res-CBAM). The features extracted from both paths are adaptively fused through an attention mechanism and subsequently classified using a multilayer perceptron (MLP). The proposed DPFNet was rigorously evaluated on the MIT-BIH Arrhythmia Database using a strict inter-patient paradigm. Experimental results demonstrate that DPFNet achieves excellent performance, attaining an F1 score of 92.38%, an accuracy of 98.90%, and a specificity of 98.97% on the inter-patient test set. The overall performance surpasses that of various state-of-the-art methods evaluated under identical experimental conditions.</p>