blob: 71822752617a2b1f27c59f8526a1857e06f15a3d [file] [log] [blame]
2017/07/10: release libxcam version 1.0.0
* 360 video stitching performance and quality improvement.
- enable geometry map to improve performance.
- quality tuned on different resolutions (1080P and 4K).
- support CPU and OpenCL path in feature match.
- enable lens shading correction based on fisheye image.
* gyroscope-based video stabilization enabling.
- enable gyroscope 3-DoF (orientation) assist video stabilization.
- orientation (gyro) data should be measured by quaternion as the pose
of target frame reference to base frame
* CL framework refine.
- enable CL argument template instead of member variables in kernel.
- CL kernel support async mode.
- image handler take over input/output buffer management from image kernel.
* prepare libxcam debian package and fix most warnings.
2017/04/01: release libxcam version 0.9.0
* 360 video stitching improvement.
- remove sharp boundary in sphere mode by local/global area resize.
- tune algorithm parameters to preserve blinking.
- integrate 360 stitching pipeline into libxcam module.
- gst-xcam-filter supports 360 stitching case.
* export C APIs for each OpenCL feature.
- support stich/3d-nr/wavelet-nr/defog features by xcam handle.
- be designed for FFmpeg xcam video filter usage.
2016/12/30: release libxcam version 0.8.0
* Digital Video Stabilization (DVS)
- enable DVS smart analysis plugin.
- algorithms used in DVS (feature detect, optical flow, motion model estimation) are based on OpenCV.
- implement image warp perspective algorithm in CL kernel.
* Image stitching to create panorama photography
- implement fisheye calculation which map location to equirectangular.
- enable feature detection and match based on OpenCV optical flow algorithm.
- enable seam mask to improve the quality of blender.
* gstreamer plugin xcamfilter can co-work with xcamsrc
2016/09/30: release libxcam version 0.7.0
* 3D-NR quality and performance improvement.
- use image exposure parameters to adjust global denoise strength.
- calculate image local variation to adjust local denoise strength.
- optimize OCL kernel by shared local memory and sub-group to
improve performance.
* defog/dehaze quality improvement based on DCP(dark channel prior).
- enable dark channel prior to improve hazy image quality.
- fix image corruption in recover step.
- add lateral-range filter to smooth dark channel and remove
halo effect.
* enable multi-band blender for image stitching.
- based on gaussian and laplacian pyramid and blend images in each
band, then reconstruct stitching image from pyramid level.
- blender can smooth seams and prevent ghosting.
* enable geometry correction.
- map 2D image from src to dst image based on coordinate table which
is center aligned.
- prevent sawtooth edge by OCL sampler.
* gst-xcamfilter feature supports.
- xcamfilter can co-work with different camera source.
2016/07/29: release libxcam version 0.6.1
* enable 3D noise reduction, performance stay tuned.
- suppress image noise in both temporal and spatial domain.
- the algorithm applies denoise filter based on pixel-block differences
in neighbor blocks and previous frames.
- the processed image is much more clean and preserves motion objects
edge in low-light conditions.
* support gst xcamfilter plugin which is independent from camera source.
- xcamfilter plugin can work with gstreamer source plugins,
e.g. videotestsrc/filesrc/v4l2src.
- xcamfilter features support defog/wireframe/waveletNR/3d-denoise.
- sink pad supports format NV12, src pad supports DMA buffer-sharing
for performace if links to HW plugin(encoder).
- wireframe/scaler/waveletNR moved to post image processor which can
co-work with isp processor(as pre-image processor)
2016/06/30: release libxcam version 0.6.0
* enable block-split version of histogram based WDR tone-mapping algorithm.
- divide image into several blocks and apply bidirectional log based
function on luminance to improve the contrast inside each block.
- apply weight function based on distance between target pixel and center
pixel of each neighbor block to smooth and eliminate block boundaries.
- apply weight function based on pixel intensity deviation between neighbor
blocks keep differentiation and weaken boundaries.
- image quality is improved and more details in both over-exposured and
under-exposured area can be shown clearly on screen.
* Bayesian wavelet shrinkage for adaptive noise filtering.
- enable adaptive noise filtering based on Bayesian shrinkage estimation
- estimate the images background noise level by analyzing.
wavelet coefficient.
- optimize the threshold T which minimizes the Bayesian risk,
i.e, the expected value of the mean square error.
* Haar wavelet tuning improvement.
- fine tuning the threshold statistic data for ultra-low light condition.
to prevent edge from over smoothing.
* face detection supports.
- enable face detection framework and support FD plugins.
- enable wire-frame to track faces.
* fog removal improvement based on Retinex algorithm.
- support multi-scale algorithm to reduce halo effects.
- tune parameters to keep brightness and clear edge.
- improve saturation based on chroma enhancement.
* image processing service framework.
- add pipe-manager to support smart-analysis and post image processor.
- separate major code into different modules, e.g. xcore/isp/ocl/3a.
- support NV12 stream as fake input instead of camera sensor.
2016/03/28: release libxcam version 0.5.0
* enable new WDR tone mapping based on histogram adjustment.
- apply bidirectional log based function on luminance to improve
image contrast.
- transmit wide dynamic range image data to 8-bit color data by
novel histogram adjustment algorithm.
- details in both over-exposured and under-exposured area clearly
show up on screen.
* enable wavelet-based algorithm for noise reduction.
- decompose image into multiple-scales by wavelet transform with
low pass and high pass filters.
- perform soft threshold on smaller coefficients(high-pass)
to reduce noise.
- threshold should be suitably decided by tuned parameters.
- reconstruct image from scaled low-pass and high-pass images.
- enable Haar-wavelet and Hat-wavelet NR filters.
* enable fog removal feature based on retinex algorithm.
- base on retinex single-scale algorithm.
- scale down image for gaussian blur to improve performance.
- scale up blured image and amplify log differences with original
image.
* improve sharpness in normal light with extreme profile.
* support cl post processing with isp mode.
* support swap-buffer on nv12 format if only single plane need to
process.
2015/12/30: release libxcam version 0.4.0
* improve performance on OCL pipeline.
- design new formats and pipeline to save memory IO.
- improve bayer noise reduction with bilateral filter.
- support edge enhancement into bayer pipe.
- make gpu/cpu working together for 3a-statistics calculation.
- move tonemapping kernel from RBG to bayer format.
* improve WDR-tonemapping algorithm for quality.
- utilize an new adaptive local tone mapping algorithm.
- provide the mapping between 16-bit wide dynamic range images to
8-bit images to have nice looking”.
- preserve more information on luminance values of the scene
especially for high contrast images.
* tonemapping parameters calculated to increase the contrast between
high light and low light.
- tone mapping regularization parameters are calculated according
to the distribution of brightness histogram.
- the local adaptive parameter are calculated by weighting the
neighbor pixels with a Gaussian blur filter.
* add bayer raw video input to simulate camera sensor.
2015/10/19: release libxcam version 0.3.0
* improved OCL pipeline on basic and advanced/extreme pipeline
- add new bayer noise-reduction into demosaic
- merged yuv-tnr into yuv-pipe kernel
- merged TNR-yuv and TNR-rgb together
* support WDR(wide dynamic range) feature
* add 3a analysis tuner framework for more features tunning, e.g TNR
* add hybrid 3a analysis framework
- support partial customized 3a algorith, e.g AWB/AE
* add smart analysis framework
- generate small scaled picture
- support customized to analyze small picture and feedback results
- support loading user-defined method on smart analysis
2015/07/31: release libxcam version 0.2.1
* improved OCL pipeline on basic 3a image processing
- merged bayer kernels with blc, wb, gamma, demosaic and 3a stats.
- merged yuv kernerls with rgb2yuv color conversion and macc.
- support async framework as option on cl features
- support OCL buffer allocation which can choose tiling mode.
* support different AIQ versions.
* fix capture of dead-loop and sensor format setting.
2015/07/02: release libxcam version 0.2.0
* add gstreamer plugin 'xcamsrc' for Linux media framework support
* support loading user-defined 3a algorithm lib dynamically
* support more manual 3a features on ISP, e.g multiple-ae-window, ae/awb speed...
* add OpenCL pipeline for 3a image processing
- bayer format features, black level correction, defect pixel
correction, whitebalance, 3a statistics calculation,
demosaic, gamma correction, HDR(high dynamic range)
- RGBA format features, bilateral NR(noise reduction), simple NR,
temporal NR, macc, color correction
- YUV format features, temporal NR, edge enhancement, color conversion
- support any user-defined 3a algorithms(e.g AIQ)
* rich test cases, e.g. test-device-manager, test-cl-image
2015/01/15: Initial xcam version 0.1