Major package refactor

[lfenzo]

• Introduced new src/ organization for image hashing implementations
• Removed inner modules AverageHash, DifferenceHash and PerceptualHash; now all functions are exported under ImageHashes.
• Moved test dependencies - eg. TestImages, Test - from Project.toml to test/Project.toml
• Reduced code duplication

This PR does not:

• edit docs
• change the public interface

[davidbp]

Hi @lfenzo, thanks for taking a look.
Did you check if there are any performance regressions out of the proposed refactoring?

[lfenzo]

Hi @lfenzo, thanks for taking a look. Did you check if there are any performance regressions out of the proposed refactoring?

Hello @davidbp,

I benchmarked it against main. As it turns out, the difference_* functions had some slight slowdown. I'll try to address that shortly!

main:

average_hash  1.173 μs (4 allocations: 416 bytes)
average_mathash  1.124 μs (7 allocations: 528 bytes)
difference_hash  1.268 μs (4 allocations: 432 bytes)   <-----
difference_mathash  1.204 μs (9 allocations: 704 bytes)  <-----
perceptual_hash  30.049 μs (60 allocations: 20.42 KiB)
perceptual_mathash  29.882 μs (65 allocations: 20.86 KiB)

PR:

average_hash  1.169 μs (4 allocations: 416 bytes)
average_mathash  1.123 μs (7 allocations: 528 bytes)
difference_hash  1.360 μs (10 allocations: 688 bytes)  <-----
difference_mathash  1.246 μs (10 allocations: 688 bytes)  <-----
perceptual_hash  30.190 μs (60 allocations: 20.42 KiB)
perceptual_mathash  30.232 μs (65 allocations: 20.86 KiB)

Benchmark Code

using ImageHashes
using TestImages
using BenchmarkTools

function run_benchmark(f, img)
    print("$(f)")
    @btime $f($img)
end

function main()
    functions = [
        average_hash,
        average_mathash,
        difference_hash,
        difference_mathash,
        perceptual_hash,
        perceptual_mathash,
    ]

    img = testimage("lighthouse")

    for f in functions
        run_benchmark(f, img)
    end
end

main()

[lfenzo]

Hello @davidbp!

I was able to reduce allocations only in difference_mathash, the difference_hash still allocates more. Haven't figured out a way to address that yet. Open to suggestions!

Performance difference:

                    PR (c64a155)                                main
difference_hash     1.299 μs (7 allocations: 560 bytes)  # vs.  1.268 μs (4 allocations: 432 bytes)
difference_mathash  1.187 μs (7 allocations: 560 bytes)  # vs.  1.204 μs (9 allocations: 704 bytes)

[davidbp]

The difference is negligible, good job!

[lfenzo]

Thanks, @davidbp!