Recently, there has been a surge in learned image compression techniques that outperform standard image compression algorithms, leading to a growing inter-est in the application of neural networks for effective image compression. Most of these new techniques prioritize improving image quality, without adequately considering processing power, processing time and system resource constraints. In this paper, we propose a promised autoencoder-based lossy image compres-sion method. Our aim is to develop a compression technique that achieves good compression performance while minimizing computational complexity. The method we propose involves partitioning the input image into two distinct groups according to the content of image blocks. For each group, we employ separate autoencoders to achieve a reduced bit rate while enhancing image quality. Subsequently, a quantization process is applied before entropy encoding to get the final outputs. The experimental results demonstrate that our proposed method achieve better PSNR than JPEG compression in low bit rates.