Text recognition of an arbitrary shape has always been a challenging task, and although recent research has improved, it has always performed poorly for ambiguous, severely curved text recognition. We believe that most algorithms directly compress visual features into one-dimensional sequences with loss of spatial information, followed by poor complementarity of global and local information and heavy reliance on the depth of the backbone network. In this work, an optional 2D feature scene text recognition network is proposed in this paper to robustly recognize arbitrarily shaped text and fuzzy text. Specifically, a transformer-based encoding and decoding framework is adopted to introduce deformable convolutional kernels in the feature extraction part and reconstruct the feature extraction network, followed by proposing a deep multi-head transposition attention network and a deep feed-forward network applied to the transformer encoding layer to both aggregate local and non-local pixel interactions and alleviate computational bottlenecks. With extensive experiments, the algorithm proposed in this paper proves to be better able to solve arbitrary shapes and fuzzy text recognition, more robust and accurate than previous methods, and achieves the best performance on several benchmark datasets.