Low-resource languages like Gujarati create major text spotting problems in natural images because set font styles, lighting conditions, and crowded backgrounds make spotting text difficult. Our study combines U-Net architecture with ResNet50 and Squeeze-and-Excitation blocks to develop a deep learning method for Gujarati text segmentation. Our method strives to find Gujarati text in real scenes by developing solutions that solve feature detection difficulties and handle attention from every image channel. Our model reaches 91.80% segmentation precision while training and evaluating through Gujarati text image data processing on a T4 GPU inside Google Colab. Our method uses ResNet50’s pretrained layers to extract features from inputs then passes those through a decoder with up sampling layers and SE blocks that increase the visibility of important feature channels. Our findings prove that the new approach performs better than standard techniques even in challenging situations. Our hybrid deep learning design proves successful at handling Gujarati text spotting tasks while working across various language processing settings.