Hmm, I think I have to give up re-working the FP16 part.

_ARM64_DISTINCT_NEON_TYPES

Thanks to @shibayan in #21630, this macro seems to be important.
This macro is added in CMake directly

I found one comment on community of the Visual Studio

Note you are not intended to use it by directly defining the _ARM64_DISTINCT_NEON_TYPES symbol per this blog post.

And the blog post says following.

New ARM64 compiler flags: /Zc:arm64-aliased-neon-types- and /Zc:arm64-aliased-neon-types. When you pass /Zc:arm64-aliased-neon-types- to cl.exe, the compiler will treat NEON intrinsic types as distinct types for ARM64 as defined by the Procedure Call Standard for the Arm 64-bit Architecture, which is consistent with Clang and GCC.

Well, the quoted comment doesn't appear directly, so I "guess" the blog was updated after the comment.
Anyway, the original bug mentioned in #16027 was fixed in Visual Studio 2022 (17.2). and rather directly adding the macro definition, we should use /Zc:arm64-aliased-neon-types-

Still, there are other things to consider.
Changing the option doesn't help the situation.

FP16 instructions disappear

I face the same situation as the previous comment

There is, however, still a compiler issue here: When using /Zc:arm64-aliased-neon-types-, all of the half-precision intrinsics are missing: vcvt_f32_f16, vcvt_f16_f32, etc.

I see the same issue.
Below is a portion of arm_neon.h on my VS 2022 (17.8)

#if !defined(_ARM64_DISTINCT_NEON_TYPES)
#define vget_lane_f16(Dm, lane)     neon_dups16((Dm), (lane))

#define vcvt_f32_f16(src)               __n128_to_float32x4_t(neon_fcvtl_32(__float16x4_t_to_n64(src)))

#define vcvt_f16_f32(src)               __n64_to_float16x4_t(neon_fcvtn_32(__float32x4_t_to_n128(src)))

Please note that specifying /Zc:arm64-aliased-neon-types- is equivalent to define _ARM64_DISTINCT_NEON_TYPES so all f16 functions will not appear and end up undeclared identifier/identifier not found

Reversing the switch neither works

The flag /Zc:arm64-aliased-neon-types- is actually an opt in version of /Zc:arm64-aliased-neon-types . There is a difference at the end of the option.
So reverting this flag will bring back fp16 instructions.
Still, this ends up in no vreinterpret_xx_xx instructions in arm64_neon.h

//vreinterpret
#if !defined(_ARM64_DISTINCT_NEON_TYPES)
#define vreinterpret_f32_s8(a)         (a)
#define vreinterpret_f32_s16(a)        (a)

So at this point, (it's the last thing I want to say, but) the VS + ARM64 is not fully ready, yet.
To check the situation, here's some test from godbolt

workaround

The only way which works is fulfilling all following options

• Set either /D _ARM64_DISTINCT_NEON_TYPES or /Zc:arm64-aliased-neon-types- using cmake
• Specify -DCPU_BASELINE=NEON instead of FP16 when configuring.
  this will allow user to use every NEON instructions but FP16.

Additional notes

I tried both CMake 3.25.0 and 3.25.1, and at least in 3.25.0, there's a bug which fails to handle try_compile on Windows + ARM64.
Who anyone facing same issue, hope it helps you.

@tomoaki0705 I merged the first your patch. Please rebase and fix conflicts.

I pushed the resolved version.
For FP16 issue on ARM64, it's really not clear how the future Visual Studio will handle it in the future, so I left it as-is.
This means, by configuring as

cmake -G"Visual Studio 17 2022" -A ARM64 -DCMAKE_SYSTEM_NAME=Windows  -DCMAKE_SYSTEM_PROCESSOR=ARM64  ..

Will end up as an compile error

error C2065: '__m128': undeclared identifier

and we have to remember that it's a known issue.

I also had to update the ocv_cpu_aarch64_baseline_merge_feature_options

CMake Error at cmake/OpenCVCompilerOptimizations.cmake:494 (string):
  string sub-command REPLACE requires at least four arguments.
Call Stack (most recent call first):
  cmake/OpenCVCompilerOptimizations.cmake:600 (ocv_cpu_aarch64_baseline_merge_feature_options)
  cmake/OpenCVCompilerOptions.cmake:342 (include)
  CMakeLists.txt:656 (include)


CMake Error at cmake/OpenCVCompilerOptimizations.cmake:494 (string):
  string sub-command REPLACE requires at least four arguments.
Call Stack (most recent call first):
  cmake/OpenCVCompilerOptimizations.cmake:600 (ocv_cpu_aarch64_baseline_merge_feature_options)
  cmake/OpenCVCompilerOptions.cmake:342 (include)
  CMakeLists.txt:656 (include)