Different ocr results with almost same image using lstm only single line mode

[shishaozheng]

Hi, I used the tesseract 4.0.0 version for ios branched by chaoskyme
I try to use it for almost same images with following language and mode (official fast eng tessdata):

let tesseract:G8Tesseract = G8Tesseract(language: "eng", engineMode: G8OCREngineMode.lstmOnly)
tesseract.pageSegmentationMode = .singleLine

I try three times, but the result is quite different, just first one is correct, the other two results is so confused.