-
Notifications
You must be signed in to change notification settings - Fork 1
/
EdgeDetector.go
291 lines (243 loc) · 8.83 KB
/
EdgeDetector.go
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
package main
import (
"fmt"
"math"
"image"
"rand"
"os"
)
const (
SudokuGridDimension = 9 // side of board (in squares, not lines)
)
type EdgeDetector struct {
lines []Line
// potential += exp(-sq_dist(point,pixel) / radius)
default_line_radius float64
// potential -= exp(-(angle(a,b) % 90.0) / orientation_sensitivity)
orientation_sensitivity float64
// how many proposals to make at each hill climbing iteration
num_proposals uint
// proposals are chosen with prob: l1_normalize(potentials ^ greedyness).
// 0 is uniform choice, infinity is perfectly greedy
greedyness float64
proposal_variance float64
}
func (ed EdgeDetector) AlignTo(img image.Image) {
// TODO i can just impelment each of these and see which is fastest (all derivative free)
// option 1: draw K transforms, take the best point
// option 2: draw K transforms, take the best point and do line search
// option 3: draw K transforms, if best point isn't "good enough" then drak K more _smaller_ transforms
bounds := NewFloat64Rectangle(img.Bounds())
cur_ed := ed
for iter := 0; iter < 15; iter++ {
// propose some new edge detector positions
minp := math.Inf(1)
proposals := make([]EdgeDetector, ed.num_proposals)
potentials := make([]float64, ed.num_proposals)
for i := uint(0); i < cur_ed.num_proposals; i++ {
proposals[i] = cur_ed.Proposal(bounds)
potentials[i] = proposals[i].Potential(img)
if potentials[i] < minp { minp = potentials[i] }
}
// make sure all potentials >= 0.0, calculate sum
for i,_ := range potentials {
c := potentials[i] - minp + 1 // smallest proposal will have potential = 1.0
potentials[i] = math.Pow(c, ed.greedyness)
}
if len(potentials) != len(proposals) {
fmt.Printf("[wtf] len(pot) = %d, len(pro) = %d\n", len(potentials), len(proposals))
os.Exit(1)
}
i := WeightedChoice(potentials)
cur_ed = proposals[i]
fmt.Printf("[EdgeDetector.AlignTo] accepting pot=%.1f\tfrom [ ", potentials[i])
for _,v := range potentials { fmt.Printf("%.1f ", v) }
fmt.Printf("]\n")
// test this on images to see how fast this should be decreased
//cur_ed.proposal_variance *= 0.9
// print out ED for debugging
outf := fmt.Sprintf("/Users/travis/Dropbox/code/sudoku/img/debug.%d.png", iter)
SaveImage(cur_ed.Draw(img), outf)
}
}
func NewEdgeDetector(b Float64Rectangle) EdgeDetector {
ed := new(EdgeDetector)
ed.default_line_radius = 1.0
ed.orientation_sensitivity = 3.0
ed.num_proposals = 75
ed.greedyness = 2.5
ed.proposal_variance = 4.0 // in degrees
// place some lines
padding := 60.0 //2.0
num_lines := 4 //SudokuGridDimension + 1
dx := (b.Dx() - 2.0*padding) / float64(num_lines - 1)
dy := (b.Dx() - 2.0*padding) / float64(num_lines - 1)
x0 := b.Min.X + padding; xmax := b.Max.X - padding; x := x0
y0 := b.Min.Y + padding; ymax := b.Max.Y - padding; y := y0
//fmt.Printf("[NewEdgeDetector] x0 = %.2f, y0 = %.2f, xmax = %.2f, ymax = %.2f\n", x0, y0, xmax, ymax)
for i := 0; i < num_lines; i++ {
v := Line{Float64Point{x, y0}, Float64Point{x, ymax}, ed.default_line_radius} // vertical
h := Line{Float64Point{x0, y}, Float64Point{xmax, y}, ed.default_line_radius} // horizontal
ed.lines = append(ed.lines, v)
ed.lines = append(ed.lines, h)
x += dx; y += dy
//fmt.Printf("[NewEdgeDetector] v = %s, h = %s\n", v.String(), h.String())
}
//fmt.Printf("[ned] ed.lines = %s\n", ed.lines)
// random perturbation of "perfect"
crappyness := 6.0
ed.proposal_variance *= crappyness
n := ed.Proposal(b)
n.proposal_variance /= crappyness
return n
}
func (ed EdgeDetector) CloneEdgeDetector() EdgeDetector {
e := new(EdgeDetector)
e.default_line_radius = ed.default_line_radius
e.orientation_sensitivity = ed.orientation_sensitivity
e.num_proposals = ed.num_proposals
e.greedyness = ed.greedyness
e.proposal_variance = ed.proposal_variance
e.lines = ed.lines[:]
return *e
}
func (ed EdgeDetector) Proposal(bounds Float64Rectangle) EdgeDetector {
new_ed := ed.CloneEdgeDetector()
// rotations and shifts must be correlated
independent_scale := 0.1
mean_theta := (rand.Float64() * 2.0 - 1.0) * (math.Pi / 180.0 * ed.proposal_variance)
mean_dx := (rand.Float64() * 2.0 - 1.0) * ed.proposal_variance
mean_dy := (rand.Float64() * 2.0 - 1.0) * ed.proposal_variance
for i, l := range ed.lines {
nl := *new(Line) // new line
nl.radius = l.radius
// first rotate the line
theta := mean_theta + independent_scale * (rand.Float64() * 2.0 - 1.0) * (math.Pi / 180.0 * ed.proposal_variance)
v := PointMinus(l.right, l.left)
z := v.Rotate(theta)
// scale back up to the correct length
// new and old vecs share a midpoint, add/subtract half of the difference
z.Scale(0.5)
nl.left = PointMinus(Midpoint(l.left, l.right), z)
nl.right = PointPlus(Midpoint(l.left, l.right), z)
// now apply left-right and up-down shifts
// TODO the indepented scale for dx dy shifts should be higher to allow for when
// the original distance between lines is too great or small
dx := mean_dx + independent_scale * (rand.Float64() * 2.0 - 1.0) * ed.proposal_variance // left-right movement
dy := mean_dy + independent_scale * (rand.Float64() * 2.0 - 1.0) * ed.proposal_variance // up-down movement
nl.left.X += dx
nl.left.Y += dy
nl.right.X += dx
nl.right.Y += dy
// now make sure it's in the bounds
nl.ProjectInto(bounds)
new_ed.lines[i] = nl
}
// scalings (shrinks and stretches) in x and y directions
// TODO write variance struct that includes L/R, U/D shift amounts in (0,1)
stretch := (rand.Float64() * 2.0 - 1.0) * ed.proposal_variance
center := Float64Point{0.0, 0.0} // find center of all lines, stretch to/from this point
for _,l := range new_ed.lines {
center := PointPlus(center, Midpoint(l))
}
center.Scale(1.0 / len(new_ed.lines))
amount := 0.0 // TODO
for _,l := range new_ed.lines {
nmp := ShiftedMidpoint(l, center, amount)
half := PointMinus(l.right, Midpoint(l))
l.right = PointPlus(half, nmp)
l.left = PointMinus(nmp, half)
}
return new_ed
}
func (ed EdgeDetector) Draw(img image.Image) image.Image {
width := img.Bounds().Dx()
height := img.Bounds().Dy()
fmt.Printf("[EdgeDetector.draw] about to copy original image (%d, %d)...\n", width, height)
output := image.NewRGBA(width, height)
for x := 0; x < width; x++ {
for y := 0; y < height; y++ {
output.Set(x, y, img.At(x, y))
}
}
fmt.Printf("[EdgeDetector.draw] about to draw %d lines: ", len(ed.lines))
for _, l := range ed.lines {
l.Draw(output)
fmt.Printf("*")
}
fmt.Printf("\n")
return output
}
func (ed EdgeDetector) Potential(img image.Image) (p float64) {
// put a "sparse prior" on random steps
// steps should usually be mostly in one direction
// compare to coordinate descent and no prior
// does it make sense to have extra benefit for getting a cross at two intersecting lines?
// this could get fooled on the numbers
// probably not...
// activation for each line and pixel
var delta, dist float64
add := 0.0
b := img.Bounds()
for x := b.Min.X; x < b.Max.X; x++ {
for y := b.Min.Y; y < b.Max.Y; y++ {
for _, line := range ed.lines {
// TODO may need to play with this formula
dist = line.SquaredDistance(float64(x), float64(y))
delta = DarknessAt(img, x, y) * math.Exp(-dist / line.radius)
p += delta; add += delta
if math.IsInf(add, 1) {
fmt.Printf("[Potential] hit inf!\n")
os.Exit(1)
}
}
}
}
p /= float64(len(ed.lines)); add /= float64(len(ed.lines))
// orientation of the lines
remove := 0.0
num_pairs := 0 // man up: N * (N-1) / 2
N := len(ed.lines)
for i := 1; i < N; i++ {
for j := 0; j < i; j++ {
num_pairs += 1
dist = ed.lines[i].Angle(ed.lines[j])
delta = math.Exp(-math.Fmod(dist, 90.0)) * ed.orientation_sensitivity
/*p -= delta;*/ remove += delta
}
}
remove /= float64(num_pairs)
p -= remove
fmt.Printf("[EdgeDetector.Potential] potential = %.2f\t(+%.2f, -%.2f)\n", p, add, remove)
return p
}
/**********************************************************************************************/
func main() {
base := "/Users/travis/Dropbox/code/sudoku/img/"
img := OpenImage(base + "clean_256_256.png")
ed := NewEdgeDetector(NewFloat64Rectangle(img.Bounds()))
// draw out ED right after creating it
SaveImage(ed.Draw(img), base + "after_ed_init.png")
ed.AlignTo(img)
SaveImage(ed.Draw(img), base + "output.png")
}
func test_draw() {
base := "/Users/travis/Dropbox/code/sudoku/img/"
inf := base + "clean_256_256.png"
outf := base + "output.png"
img := OpenImage(inf)
// convert to grayscale, make mutable
m_gray_img := Convert2Grayscale(img)
// draw a line on it
ed := new(EdgeDetector)
b := NewFloat64Rectangle(m_gray_img.Bounds())
for i := 0; i < 500; i++ {
mid := RandomPointBetween(b.Min, b.Max)
lo := RandomPointBetween(b.Min, mid)
hi := RandomPointBetween(mid, b.Max)
radius := rand.Float64() * 5.0
ed.lines = append(ed.lines, Line{lo, hi, radius})
}
m_col_img := ed.Draw(m_gray_img)
SaveImage(m_col_img, outf)
}