Detecting arrhythmias, a significant category of cardiovascular disorders, plays a pivotal role in early diagnosis and treatment of heart diseases. In this context, our goal in this paper is to propose an innovative approach using deep learning, specifically a Bilinear Convolutional Neural Network (BCNN), to detect and to classify arrhythmias using electrocardiogram (ECG) signals transformed into grayscale images, after removing noise and balancing the ECG data using the AC-GAN model. Our BCNN model leverages a unique bilinear pooling technique, which facilitates end-to-end learning of feature extractors, combining two pre-trained and fine-tuned CNNs (VGG16 and ResNet50), effectively establishes connections among local features in a translation-invariant manner. The results of the proposed BCNN model achieved an impressive accuracy of 99.24%, surpassing related studies and prior research. This breakthrough holds promise for enhancing the automated, early detection and classification of arrhythmias, thereby contributing significantly to the diagnosis and treatment of various cardiovascular diseases.