To enhance the performance of future elastic optical networks, a modulation format recognition method based on amplitude density features is proposed. This method uses amplitude density features as input to an improved MobileNetV2 model to identify modulation formats and introduces a normalized attention mechanism (NAM) for accurate identification of transmission signal modulation formats. The feasibility of this approach is verified in 28 GBaud quadrature phase shift keying(QPSK), 8-level quadrature amplitude modulation (8QAM), 16QAM, 32QAM, 64QAM, and 128QAM transmission systems. The experimental results show that the minimum optical signal-to-noise ratio (OSNR) required for achieving 100% recognition accuracy for each modulation format is below its corresponding 20% forward error correction (FEC) threshold. Furthermore, within a broad OSNR range, a recognition accuracy of 99.62% is achieved. In optical networks with residual dispersion, this approach still maintains high recognition performance.