Atrial fibrillation (AF) is the most common arrhythmia that increases the risk of cardiac diseases such as stroke; hence, its early detection is important. Although many studies have been conducted on automatic AF detection, the accuracy of identification needs to be improved for large amounts of data such as Holter electrocardiograms (ECGs). In rare clinical cases, AF shows regular R‒R intervals (RRIs). These cases are difficult to diagnose even by experienced cardiologists or to detect automatically. To detect AF accurately with minimum number of oversights, identification of AF including regular RRIs is necessary. This paper presents a new method for improving the accuracy of AF detection. The detection is realized in two stages with different AF identification models, after eliminating noise in the ECG waveforms using a finite impulse response bandpass filter. In the first stage, a hybrid convolutional neural network (CNN)‒long short-term memory model trained with segmented ECG waveforms and their RRIs is used to identify AF with irregular RRIs. In the second stage, non-AF events are input. AF cases with regular RRIs are identified using an independent CNN model trained with waveforms that include P- or fibrillatory waves. In this study, 24-hour Holter ECG data obtained from 60 subjects were used to train both models. For evaluation, Holter ECG data from ten subjects not used for training were used, which yielded a detection accuracy of 94.7% and a sensitivity of 98.5% for AF. In addition, evaluation using another untrained dataset from two subjects whose AF had regular RRIs showed an accuracy of 96.2% and a sensitivity of 98.8% for AF. These results demonstrate the feasibility of the proposed method for AF detection.