With the advancement of observational techniques, over one million stars have been observed to search for exoplanets. The vast amounts of light curve data generated necessitate an automated, rapid, and efficient method for screening exoplanet candidates. In recent years, deep learning methods, particularly convolutional neural networks (CNNs), have been employed to automate the screening of light curves for exoplanet candidates. However, CNNs have several drawbacks, such as slow training times due to their large network structures and a lack of ability to explain the physical meaning behind their classifications. In this paper, we propose a CNN classification model that incorporates a channel attention mechanism, which captures significant features in transit signals and enhances the model’s classification capability and interpretability. Furthermore, we have improved the fully connected neural network by incorporating the residual network structure, significantly increasing the model’s training speed. Our model is highly adaptable to various data sets, achieving 96.2% accuracy and 95.7% F1 score on the Kepler data set, and 99.9% accuracy with 99.5% F1 score on the Transiting Exoplanet Survey Satellite data set, with training times ranging from 0.3 to 0.6 hr per data set. We also analyze the feature heat maps of the channel attention mechanism to better explain the model’s classifications. Our study demonstrates that deep learning can assist in managing the current era of astronomical big data and suggests that the channel attention mechanism and residual network structure could be applied to other astronomical classification tasks.