Optimization of Japanese text automatic translation system based on computer vision

Abstract

In response to the rapidly growing demand for Japanese visual text translation in scenarios such as cross-border e-commerce, games, and online education, this paper addresses the issues of unstable recognition, semantic misalignment, and excessive inference costs that often occur in "OCR + machine translation" under complex backgrounds, vertical-horizontal mixed layouts, and low-resolution conditions. It has constructed a set of automated translation optimization framework with clear engineering mechanisms. Starting from the complexity index of the structure, it introduces a character recognition loss that combines structural perception and uncertainty weighting, and explicitly models the cascading distortion of visual uncertainty on translation quality in multiple scenarios. At the visual-language decoding end, it designs character/line/block three-level alignment and consistency regularization, and adds an ambiguity-driven gating mechanism for context. At the system level, it achieves multi-configured inference under efficiency constraints through "difficulty perception + budget-driven" dynamic routing. Experiments show that the complete system improves BLEU from 32.5 to 36.8, and the semantic consistency score increases from 78.0 to 85.6. The lightweight configuration reduces FLOPs from 45.3 G to 18.7 G while compressing the end-to-end latency from 185 ms to 92 ms, demonstrating a better synergy of accuracy and efficiency. This provides a feasible path for the engineering deployment of Japanese visual text in cross-terminal and multi-scenario conditions.