Motor-imagery (MI) decoding based on convolutional neural networks (CNNs) is a research hotspot in intelligent rehabilitation. However, current decoding methods struggle to dynamically explore the temporal-spectral-spatio features of electroencephalogram (EEG) signals, which vary across subjects and affect decoding performance. Therefore, we propose a dynamic convolution and attention mechanism based convolutional neural network (DCAMNet). First, a filter bank divides raw EEG signals from each channel into multi-spectral bands, which are simultaneously input into the feature extraction module. Then, the dynamic convolution block calculates attention weights to obtain valuable temporal-spectral information, which is passed through the spatial convolution block and the temporal attention block to learn spatial information and temporal correlations, achieving discriminative temporal-spectral-spatio feature extraction and fusion. Finally, the classification module completes the MI decoding. Based on the public BCI Competition IV Dataset 2a, a four-class ten-fold cross-validation experiment on nine subjects achieves an average accuracy of 79.17% and an $F_1 $ score of 0.788. The results show that the DCAMNet can adaptively focus on and enhance discriminative features caused by inter-subject variability, achieving multi-domain feature extraction and fusion. Compared to current popular methods, it has advantages in decoding accuracy and generalization performance.