68 lines
1.4 KiB
Go
68 lines
1.4 KiB
Go
|
package gocv
|
||
|
|
|
||
|
|
import (
|
||
|
|
"image"
|
||
|
|
)
|
||
|
|
|
||
|
|
// FP16BlobFromImage is an extended helper function to convert an Image to a half-float blob, as used by
|
||
|
|
// the Movidius Neural Compute Stick.
|
||
|
|
func FP16BlobFromImage(img Mat, scaleFactor float32, size image.Point, mean float32,
|
||
|
|
swapRB bool, crop bool) []byte {
|
||
|
|
|
||
|
|
// resizes image so it maintains aspect ratio
|
||
|
|
width := float32(img.Cols())
|
||
|
|
height := float32(img.Rows())
|
||
|
|
|
||
|
|
square := NewMatWithSize(size.Y, size.X, img.Type())
|
||
|
|
defer square.Close()
|
||
|
|
|
||
|
|
maxDim := height
|
||
|
|
var scale float32 = 1.0
|
||
|
|
if width > height {
|
||
|
|
maxDim = width
|
||
|
|
scale = float32(size.X) / float32(maxDim)
|
||
|
|
}
|
||
|
|
if width < height {
|
||
|
|
scale = float32(size.Y) / float32(maxDim)
|
||
|
|
}
|
||
|
|
|
||
|
|
var roi image.Rectangle
|
||
|
|
if width >= height {
|
||
|
|
roi.Min.X = 0
|
||
|
|
roi.Min.Y = int(float32(size.Y)-height*scale) / 2
|
||
|
|
roi.Max.X = size.X
|
||
|
|
roi.Max.Y = int(height * scale)
|
||
|
|
} else {
|
||
|
|
roi.Min.X = int(float32(size.X)-width*scale) / 2
|
||
|
|
roi.Min.Y = 0
|
||
|
|
roi.Max.X = int(width * scale)
|
||
|
|
roi.Max.Y = size.Y
|
||
|
|
}
|
||
|
|
|
||
|
|
Resize(img, &square, roi.Max, 0, 0, InterpolationDefault)
|
||
|
|
|
||
|
|
if swapRB {
|
||
|
|
CvtColor(square, &square, ColorBGRToRGB)
|
||
|
|
}
|
||
|
|
|
||
|
|
fp32Image := NewMat()
|
||
|
|
defer fp32Image.Close()
|
||
|
|
|
||
|
|
square.ConvertTo(&fp32Image, MatTypeCV32F)
|
||
|
|
|
||
|
|
if mean != 0 {
|
||
|
|
// subtract mean
|
||
|
|
fp32Image.SubtractFloat(mean)
|
||
|
|
}
|
||
|
|
|
||
|
|
if scaleFactor != 1.0 {
|
||
|
|
// multiply by scale factor
|
||
|
|
fp32Image.MultiplyFloat(scaleFactor)
|
||
|
|
}
|
||
|
|
|
||
|
|
fp16Blob := fp32Image.ConvertFp16()
|
||
|
|
defer fp16Blob.Close()
|
||
|
|
|
||
|
|
return fp16Blob.ToBytes()
|
||
|
|
}
|