Enhanced Attention-Based Encoder-Decoder Framework for Text Recognition

Recognizing irregular text in natural images is a challenging task computer vision. The existing approaches still face difficulties recognizing because of its diverse shapes. In this paper, we propose simple yet powerful recognition framework based on an encoder-decoder architecture. proposed divided into four main modules. Firstly, the image transformation module, Thin Plate Spline (TPS) employed to transform readable image. Secondly, novel Spatial Attention Module (SAM) compel model concentrate regions and obtain enriched feature maps. Thirdly, deep bi-directional long short-term memory (Bi-LSTM) network used make contextual map out visual generated from Convolutional Neural Network (CNN). Finally, Dual Step Mechanism (DSAM) integrated with Connectionist Temporal Classification (CTC) - decoder re-weights features focus intra-sequence relationships generate more accurate character sequence. effectiveness our verified through extensive experiments various benchmarks datasets, such as SVT, ICDAR, CUTE80, IIIT5k. performance analyzed accuracy metric. Demonstrate that method outperforms both regular text. Additionally, robustness approach evaluated using grocery GroZi-120, WebMarket, SKU-110K, Freiburg Groceries datasets contain complex images. Still, produces superior datasets.