Word Segmentation via Convolutional Neural Network Probability Maps

The patent transforms the image processing approach by changing from traditional binary thresholding to color space transformation (RGB to HSV), and further to neural network-based probability classification. This parameter change enables the system to handle varied lighting conditions and colors by working in the HSV color space where hue, saturation, and value can be independently analyzed, and ultimately uses probability values from neural networks to robustly distinguish text from background under varying conditions.

The patent replaces manual threshold setting and rule-based text detection with an automated neural network system. The convolutional neural network automatically learns optimal detection parameters from training data, eliminating the need for manual intervention and artificial rules. This substitution of mechanical/threshold-based methods with intelligent automated systems resolves the reliability issue under varying conditions.

The patent performs preliminary actions by pre-processing images through color space transformation (RGB to HSV) and normalizing the V channel before neural network processing. This preliminary preparation of data in an optimal format accelerates the subsequent neural network processing while maintaining high accuracy. The pre-computed probability maps from the neural network also serve as preliminary results that directly enable fast word segmentation without requiring additional threshold tuning.

The patent substitutes the mechanical binarization process with a neural network-based probability classification system. Instead of applying fixed or manual thresholds to binary images, the system uses a trained neural network to directly output probability values indicating text likelihood, which are then used for segmentation. This substitution simultaneously improves speed by eliminating iterative threshold adjustment and accuracy by using learned patterns from diverse training data.

The patent segments the complex neural network processing into distinct modular stages: (1) color space transformation from RGB to HSV, (2) normalization of the V channel, (3) convolutional neural network processing to generate probability maps, and (4) threshold-based segmentation using the probability maps. This segmentation of the processing pipeline into independent, well-defined stages reduces overall system complexity by making each stage manageable and interchangeable.

The patent introduces intermediary representations that simplify the interface between different processing stages. The HSV color space serves as an intermediary between raw RGB images and neural network input, providing a more suitable representation for text detection. The probability map generated by the neural network acts as an intermediary between complex feature learning and simple threshold-based segmentation, decoupling these operations and reducing overall system complexity.