Time series similarity search under elastic distance measures (such as DTW, TWED, and MSM) is a fundamental yet computationally intensive task in many data-driven domains. In this paper, we explore a novel vision-based approach to this problem by investigating whether pretrained vision models—originally developed for natural image understanding—can be adapted to capture meaningful representations of time series through their line graph visualizations. We propose a framework, EViTS, which fine-tunes a pretrained vision model to extract discriminative features from time series plots. These features are then used to enable efficient similarity search via a two-stage process: an approximate filtering phase based on vector similarity, and an exact refinement phase using elastic distances with lower bound pruning. Extensive experiments on multiple benchmark datasets demonstrate that EViTS not only reduces the number of costly elastic distance computations but also achieves competitive or superior retrieval accuracy compared to related methods. Our findings suggest that vision-based pretrained models, when properly adapted, offer a powerful and scalable alternative for time series similarity search under elastic distances.