Abstract: Autonomous localization is a key technology in robotics. Visual localization has become a widely adopted localization approach in recent years, as it can effectively alleviate the limitations of GPS, Bluetooth, and Wi-Fi in indoor environments, where these signals are susceptible to interference and instability. To address the insufficient positional representation of feature descriptors and the limited coordinate accuracy of keypoints in conventional visual localization methods, a feature enhancement module for indoor visual lo-calization is proposed. First, learnable Fourier features are employed to encode keypoint coordinates, and the resulting positional encodings are deeply fused with the original feature descriptors to enhance their spatial representation capability; second, random projective transformations are applied to a single input image, and geometrically corresponding feature matches are constructed between the original and trans-formed images; Finally, local neighborhood information around the matched keypoints is extracted, and a convolutional neural network is used to predict coordinate offsets, thereby achieving subpixel refinement of keypoint coordinates from a single image. Experimental results on the HPatches, InLoc, and Wanda da-tasets demonstrate that the proposed module improves the accuracy and matching performance of feature extraction methods without significantly increasing computational cost, thereby substantially improving the localization performance of visual localization methods.