OCR-ENHANCED KERNEL PROPOSAL NETWORK FOR ARBITRARY-SHAPED SCENE TEXT DETECTION AND RECOGNITION

Abstract

Scene text detection for arbitrary-shaped text remains challenging, especially in separating neighbouringinstances. This article proposes a Kernel Proposal Network (dubbed KPN). This novel method addresses thisissue by classifying texts into instance-independent feature maps through dynamic convolution kernels extractedfrom Gaussian center maps. To promote kernel independence, an Orthogonal Learning Loss (OLL) isintroduced. Each kernel proposal encodes self-information and positional features, enabling effective separationof text instances without heavy post-processing.In this extension, we further integrate an OCR module that leverages the KPN-detected regions to perform textrecognition. The detected arbitrary-shaped text masks are processed through a OCR model, achieving acomplete pipeline for scene text detection and recognition. Experimental results validate the effectiveness of ourunified approach on multiple challenging datasets. The code is publicly available athttps://github.com/ATV-77/OCR-Enhanced-KPN