Improve ImageSearch

[iseahound]

ImageSearch has two components:

• Linear search
• Subimage matching

Prioritize:

• Speed up linear search
• Prevent subimage matching

Here is some code to help you:

#include *i ImagePut%A_TrayMenu%.ahk
#include *i ImagePut (for v%true%).ahk
#singleinstance force

pic := ImagePutBuffer("[images]\h.png")                ; Screen capture
pic.show() ; or ImageShow(pic)                         ; Show image
if xy := pic.ImageSearch("[images]\g.png") {           ; Search image
    MouseMove xy[1], xy[2]                             ; Move cursor
    Send "{MButton}"                                   ; MsgBox pic[xy*]
} else Tooltip "no"

https://github.com/iseahound/ImagePut/assets/9779668/3c2587da-d6ed-449d-af78-89633e2a7442 https://github.com/iseahound/ImagePut/assets/9779668/25a63f1b-8fb2-4c2a-84c4-34867b1e5ece

Useful ideas: boyer moore (and string search algorithms in general) multi-dimensional string searcch Has ImageSearch been a priority in the public domain?

Some research papers would be immensely valuable.

Current performance is about 19 fps, compared to pixelsearch's 4000 fps. Aim for 800 fps or higher.

[iseahound]

Benchmarks (on a slow laptop):

Python's template matching (from OpenCV): 3.43 fps First Attempt at ImageSearch: 13.43 fps

[iseahound]

This one does about 200 fps on the test images by preventing the subimage matching.

https://godbolt.org/z/vP679rhYP

The question is: is matching 2/3/4 best? For this script, matching 1,4 is good. But only 4 can be optimized. 2 and 3 may be too correlated.

[iseahound]

This one fastest: https://godbolt.org/z/rGKczzYqs

Rank:

1. focus
2. start pixel
3. range

[iseahound]

#include *i ImagePut%A_TrayMenu%.ahk
#include *i ImagePut (for v%true%).ahk
#singleinstance force

; a := imageputwindow("g.png")

pic := ImagePutBuffer("h.png")                ; Screen capture
hwnd := pic.show() ; or ImageShow(pic)                         ; Show image
if xy := pic.ImageSearch3("g.png") {           ; Search image
    x := xy[1], y := xy[2]
    MouseMove x, y
    WinGetPos &wx, &wy,,, hwnd
    ImagePutWindow("g.png"
       , x ", " y
       , [wx + x, wy + y]) 
} else Tooltip "no"

[iseahound]

or

pic := ImagePutBuffer("h.png")                ; Screen capture
hwnd := pic.show() ; or ImageShow(pic)                         ; Show image
if xy := pic.ImageSearch("g.png") {           ; Search image
    x := xy[1], y := xy[2]
    MouseMove x, y
    ImagePutWindow("g.png"
       , x ", " y
       , [x,y,,, hwnd])
} else Tooltip "no"

with

      try dpi := DllCall("SetThreadDpiAwarenessContext", "ptr", -3, "ptr")
      if IsObject(pos) && pos.HasKey(5) {
         pos[5] := (hwnd := WinExist(pos[5])) ? hwnd : pos[5]
         VarSetCapacity(rect, 16, 0)
         DllCall("GetClientRect", "ptr", pos[5], "ptr", &rect)
         DllCall("ClientToScreen", "ptr", pos[5], "ptr", &rect)
         x += NumGet(rect, 0, "int")
         y += NumGet(rect, 4, "int")
      }
      try DllCall("SetThreadDpiAwarenessContext", "ptr", dpi, "ptr")

?

[iseahound]

Yeah, the [x, y, w, h, r] notation seems to be ideal, as it would allow screenshots relative to a window as well.

[iseahound]

Another note to rank:

1. focus
2. start pixel - if removed with 3, there is a very large speed change.
3. range - if remove, not much speed change

Because the first pixel kind of acts like another check!

[iseahound]

         ; Search for the address of the first matching image.
         address := DllCall(code, "ptr", this.ptr, "uint", this.width, "uint", this.height
            , "ptr", image.ptr, "uint", image.width, "uint", image.height, "uint", image.width//2, "uint", image.height//2
            , "cdecl ptr")

[iseahound]

All of the above has been completed. Only useful data is:

Rank:

1. focus
2. start pixel
3. range