OpenCV 4.8.0 was officially released in the early morning of June 29, 2023. In this release, there are many highlights such as the support for the TFLite model in the DNN module and the support for FP16 precision inference. Click on the bottom to read the original text to access the original English version of ChangeLog. The ChangeLog translation is as follows (contributions from our team and the Chinese community are marked in red):
DNN:
Supports reading and reasoning of TFLite models (including int8 quantized models)
.Support compiling DNN modules in a way that does not depend on Protobuf.
More operators and models are supported:
ONNX: Layer Normalization, GELU和QLinearSoftmax。
CANN backend: supports ONNX operators Split, Clip(Relu6), Slice, Sub, PRelu, Conv(auto_pad) and ConvTranspose.
Refactor the implementation of the Reduce operator to achieve better robustness and better performance.
Support Meta's Segment Anything Model
.Fix broadcast problem of nary element wise operator on CPU and CUDA backend.
Accelerate the running speed of Conv operators on ARM and x86, optimize the speed of Conv operators in 1D and 3D input situations, and support the fusion of Conv operators and element-wise operators.
The FP16 computing branch is supported on the ARMv8 platform, which can be enabled through net.setPreferableTarget(DNN_TARGET_CPU_FP16); compared to FP32, it has a 1.5 times faster speed
.Refactor the Vulkan backend to achieve better robustness and performance, which is 4 times faster than before refactoring
.Added blobFromImageParam API for building network input and preprocessing.
Added support for newer versions of OpenVINO.
Objdetect:
Update FaceDetectorYN, the speed and accuracy have been further improved, and OpenCV Zoo has been updated simultaneously
.Add a new QR code detection algorithm based on ArUco.
The barcode detection and decoding functions are migrated from opencv_contrib to the main module.
Added GraphicalCodeDetector, a general API for graphic codes that can call barcodes and QR codes.
Add support for ChAruco boards released in pre-4.6.0.
Fixed some bugs in QR code detection and decoding, and improved performance.
Fixed several ArUco bugs and improved performance.
Calibration:
Improvements to the USAC framework.
Fix stddev estimation in camera calibration process.
Fix incorrect pixel grid generation in icvGetRectangles, improve the accuracy of getOptimalNewCameraMatrix, stereoRectify and other proofreading functions.
Support template generation and interactive proofreading tools for Charuco board, as well as corresponding examples.
Image processing:
Fixed several issues in LineSegmentDetector.
Fix the problem that resize returns incorrect results when the mode is set to INTER_NEAREST_EXACT and the input scale is even.
Optimize the local cost calculation of IntelligentScissorsMB::buildMap.
Inlier of remap() that retains BORDER_TRANSPARENT mode.
Fix distransform to support large image input.
Features2d:
Improve the accuracy of SIFT algorithm.
Core:
Add REDUCE_SUM2 mode of cv::reduce.
Added cv::hasNonZero function.
Update the version of the dependent library IPP to 20230330.
Improved support for RISC-V RVV vector extension instruction set:
Support for RVV v0.11 inrinsics available in LLVM 16 and GCC 13.
Support Pingtouge RISC-V toolchain (RVV 0.7.1 and 1.0).
Several OpenCL vendor and version handling improvements.
Multimedia:
Added support for AVIF by relying on libavif.
Support Obi Zhongguang Femto Mega camera.
HEVC/H265 support for ViedeoWriter with MS Media Foundation backend.
Fix FPS calculation for FFmpeg backend.
Support CAP_PROP_AUTO_WB and CV_CAP_PROP_WHITE_BALANCE_BLUE_U of VideoCapture in DShow backend.
Fix capture of OBS Virtual Camera.
CV_32S encoding supports tiff.
Introduced stricter decoder checks and dynamic resolution support in the cudacodec module of opencv_contrib.
Added basic VideoCapture property support in opencv_contrib's cudacodec module.
Python Bindings:
Fix the problem of IDE auto-completion for opencv python.
Fix reference calculation error in registerNewType.
Fix detector binding for ChAruco and diamond board.
Add bindings to support GpuMat and Steam objects initialized from other libraries.
Supports internal conversion of np.float16 to Mat.
Added Python bindings for RotatedRect, CV_MAKETYPE and CV_8UC(n).
JavaScript Bindings:
Allow to turn off wasm in opencv.js.
Add JS bindings for Arcuco, ChArcuco, QR codes and barcodes.
Platforms:
Android: no longer rely on deprecated tools
iOS: Fix framework build issue with CMake 3.25.1+.
other:
Fix several serious problems in wechat_qrcode of opencv_contrib module.
Download OpenCV 4.8.0:
https://github.com/opencv/opencv/releases/tag/4.8.0
OpenCV中国团队是OpenCV在中国的开发团队，非营利目的，致力于OpenCV的开发、维护和推广工作。
长按下方QR码关注我们获取最新动态