An app for removing compression and noise from photos

this is not another "AI upscaler" but a compression artifact remover and denoiser using open-source weights such as FBCNN and SCUNet

features:

• remove compression artifacts
• denoise
• before/after view
• fully offline
• image descaling (see BRISQUE under in-app FAQs)

models:

you can download models here

examples

check out examples to get an idea of what DeJPEG can be used for

limitations:

• processed locally, minimum 4gb ram and 4 threads recommended.
• large images might cause issues
• no GIF, TIFF or RAW support

other platforms

please use chaiNNer which should work well with these models (regardless if PyTorch or ONNX)

for FBCNN, which chaiNNer does support but in a limited fashion, install this custom node and use the original PyTorch models, not the mobile onnx.

building

this app includes OpenCV for BRISQUE assessment

required:

• android NDK 27.3.x
• cmake 3.x or newer
• git
• opencv + contrib

create a directory called opencv after cloning dejpeg, then clone opencv + contrib inside that folder

git clone https://github.com/opencv/opencv.git
git clone https://github.com/opencv/opencv_contrib.git

and make a folder called build_android

cd opencv
mkdir build_android && cd build_android

structure should be like this

$ tree -L 2
...
├── opencv
│   ├── build_android
│   ├── opencv
│   ├── opencv_contrib
...

rm -f CMakeCache.txt && rm -rf CMakeFiles/

cmake \
  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_HOME/build/cmake/android.toolchain.cmake \
  -DANDROID_ABI=arm64-v8a \
  -DANDROID_NATIVE_API_LEVEL=21 \
  -DANDROID_STL=c++_shared \
  -DCMAKE_BUILD_TYPE=MinSizeRel \
  -DCMAKE_CXX_FLAGS_MINSIZEREL="-Os -DNDEBUG -fvisibility=hidden -fvisibility-inlines-hidden -ffunction-sections -fdata-sections" \
  -DCMAKE_C_FLAGS_MINSIZEREL="-Os -DNDEBUG -fvisibility=hidden -fvisibility-inlines-hidden -ffunction-sections -fdata-sections" \
  -DCMAKE_SHARED_LINKER_FLAGS="-Wl,--gc-sections -Wl,-z,max-page-size=16384 -Wl,-z,common-page-size=4096" \
  -DOPENCV_EXTRA_MODULES_PATH=../opencv_contrib/modules \
  -DBUILD_SHARED_LIBS=ON \
  -DBUILD_LIST=core,imgproc,imgcodecs,ml,quality \
  -DBUILD_TESTS=OFF -DBUILD_PERF_TESTS=OFF -DBUILD_ANDROID_EXAMPLES=OFF -DBUILD_DOCS=OFF -DBUILD_EXAMPLES=OFF -DBUILD_PACKAGE=OFF -DBUILD_opencv_java=OFF \
  -DWITH_JPEG=ON -DWITH_PNG=ON -DWITH_TIFF=OFF -DWITH_WEBP=OFF -DWITH_OPENEXR=OFF \
  -DWITH_IPP=OFF -DWITH_EIGEN=OFF -DWITH_LAPACK=OFF -DWITH_TBB=OFF -DWITH_PTHREADS_PF=ON \
  -DWITH_GSTREAMER=OFF -DWITH_V4L=OFF -DWITH_GTK=OFF -DWITH_QT=OFF -DWITH_OPENCL=OFF -DWITH_CUDA=OFF -DWITH_VTK=OFF \
  -DBUILD_PROTOBUF=OFF -DBUILD_TBB=OFF -DBUILD_IPP_IW=OFF -DBUILD_ITT=OFF \
  ../opencv && make -j$(nproc)

you can skip stripping and just copy the libs from lib/arm64-v8a to there and the next operation will strip them but you'll need to build a Release instead of Debug (and sign)

copy:

cp -r lib/arm64-v8a ../../app/src/main/jniLibs && cd ../../app/src/main/jniLibs/arm64-v8a

strip debug symbols:

llvm-strip libopencv_{core,imgproc,ml,imgcodecs,quality}.so

build:

rm app/src/main/jniLibs/arm64-v8a/libbrisque_jni.so
BUILD_BRISQUE_JNI=ON ./gradlew clean assembleDebug

note: due to IzzyOnDroid's apk filesize limit to 30mb or less, the native binaries in the official release (including libbrisque_jni.so) are compressed using upx --best --lzma --android-shlib after being stripped of debug symbols (excluding libc++_shared.so). you can skip this step if needed.

compres

upx --best --lzma --android-shlib libopencv_{core,imgproc,ml,imgcodecs,quality}.so

additionally, i also compress libonnxruntime.so with UPX as well (using upx --ultra-brute --android-shlib), so to get the original version, extract libonnxruntime.so from your compiled version after you've comment out the following config, then extract from the apk, copy to jniLibs, pack using UPX, uncomment and build. this is how i make all releases as to not exceed IzzyOnDroid's 30mb apk size limit but only needs to be done once unless i change something with opencv/brisque.

...
    packaging {
        jniLibs {
            pickFirsts += "lib/*/libonnxruntime.so"
        }
    }
...

credits and license

I am by no means a professional developer and only do this in my spare time, the code is not perfect and quite janky.

Credits to @adrianerrea for a starting point, FBCNN and SCUNet creators plus all other model owners.

This is a GUI for a select amount of 1x ONNX processing models, used under their respective licenses (Apache 2.0). You are welcome to embed parts of this app in your own project as long as it remains free (as in beer) and abides to the GPL v3 license

disclaimer:

DeJPEG is not affiliated or related with Topaz DEJPEG or any other similarly named software/project. although ive wondered if the term 'JPEG' is copyrighted/trademarked due to it literally being the acronym for Joint Photographic Experts Group, for this reason i might need to change the app's name if legal issues start to occur.