Pre-processing stereo transparent images: extraction of non-transparent regions by variable length pattern correspondence

One of the hallmarks of a transparent image is the superimposition of structures in the image. This gives the image its "see through" character. However, portions of a transparent image can be considered non-transparent if no superimposed structures are present. By defining a new type of pixel uniqueness, which the authors call "pattern uniqueness", the nontransparent portions of transparent images can be identified, and either removed to clarify the adjacent image features or used as templates to find and separate superimposed structures. Several adjustable criteria were defined in order to optimize the algorithm's performance: main parameters include two types of pattern similarity and minimum pattern cohesiveness. Simulated stereo image lines were processed by the algorithm with various parameter settings. An analysis of variance determined the influence of parameter alteration on algorithmic performance. It was found that the matching similarity, cohesiveness, and interaction between these variables were very important and could be adjusted to optimize noise tolerance and performance.