This study designs a novel image encryption cryptosystem through the two-dimensional partial unwinding decomposition (2D-PUD). It consists of three stages. Firstly, a stream sequence (first part of the security key) is generated by pseudo-random number. Secondly, the plain image is decomposed into three parts by 2D-PUD: one 2D decomposition component, two 1D decomposition components, and the average intensity value of the image. Finally, the 2D decomposition component is shuffled by a generalized Arnold transform where the average intensity value is selected as second part of the security key. The diffusion scheme is subsequently applied to the scrambled image via exclusive OR operations with the randomized 1D decomposition components (third part of the security key) along rows and columns to obtain the cipher image. Due to the adaptive attribute of 2D-PUD, the generated 1D decomposition components are completely distinct for different images. In addition, we can also make them significantly different by tuning the decomposition times for the same image. Thus, the proposed algorithm is an image-content-adaptive encryption scenario that can effectively resist cryptographic attacks. Simulation results demonstrate that our proposed method has excellent encryption performance and can resist against various typical attacks, including brute force, statistical, entropy, and differential attacks.