Reduced-order modeling stands as a pivotal method in curbing the computational expenses linked with expansive fluid dynamics quandaries by employing proxy numerical simulations. Within this realm, downscaling and reconstruction methods serve as fundamental constituents of reduced-order modeling. The traditional intrinsic orthogonal decomposition relies on linear mapping, often relinquishing a substantial amount of nonlinear flow information within the flow field. Meanwhile, autoencoders equipped with fully-connected structures, maybe encounter a parameter explosion when handling larger-scale flow field meshes, impeding effective training. Convolutional autoencoders necessitate uniform interpolation across the flow field to attain a uniform flow field snapshot, yet this process frequently introduces interpolation errors and unwarranted temporal overheads. This paper introduces an innovative solution: a non-interpolated convolutional autoencoder, designed to extract nonlinear features from the flow field while curbing parameter count, evading interpolation errors, and mitigating additional computational burdens. Illustratively, in a two-dimensional cylindrical winding flow scenario, both the reduced dimensional reconstruction display root mean square errors of approximately 1×10-3. Notably, the velocity cloud and absolute error cloud vividly exhibit the non-interpolated convolutional autoencoder's remarkable prowess in reconstruction.