In this work, we investigate the use of DM-time images as input for convolutional neural networks (CNNs) to classify pulsar and transient radio signals. Our previous work highlighted significant limitations with spectrogram-based models, particularly low sensitivity in detecting faint pulses amidst noise. The decision was made to use DM-time images, which capture detailed dispersion characteristics, offering enhanced detection capabilities for weak signals. We developed minimalist CNN architectures, ranging from one to three convolutional layers, optimized for real-time processing with reduced computational demands. The models were trained and tested using datasets derived from Crab Pulsar observations, with promising results demonstrating robust pulse detection even under challenging signal-to-noise conditions. The sensitivity of the models was evaluated against both real and synthetic data, showing high accuracy for pulses with SNR greater than 8.5, and maintaining sensitivity for weaker pulses under certain configurations. Furthermore, performance tests on a high-performance cluster revealed the feasibility of using these models in real-time applications, with scalable improvements in execution time as CPU resources were increased. This work provides a foundation for efficient and scalable real-time pulse classification in radio astronomy.
