OpenCV 4.14.0

What's new in this version:

Core module:
- Fixed copyTo on empty fixed-type matrices
- Improved precision of RotatedRect::points
- Fixed UB in cv::error when breakOnError set to true
- Fixed cv::mul overflow for U16 type
- Revert "Disable IPP with AVX512 in cv::compare because of performance regression"
- Fixed UBSan function pointer type mismatch in countNonZero
- Fixed heap-buffer-overflow in YAML parseKey for empty keys
- Optimized flip Horizontal
- Added AVX VNNI support
- Fixed meanStdDev on Windows for ARM
- Added NEON implementation for rotate function
- Fixed inplace transpose race by enforcing LLSS ordering via local barrier in OpenCL
- Fixed incorrect results for in-place flip on strict OpenCL implementations
- Added platform-specific SIMD for cv::reduce REDUCE_SUM
- Implemented safe handling of compressed file level extension (.gz[0-9])
- Extended AutoBuffer to mimic std::vector
- Fixed signed overflow UB in Point/Point3i::dot()
- Added SIMD intrinsics support for norm_mask function
- Add ARMPL support for DFT Function
- Implemented OpenCL kernel for UMat::diag to avoid aliasing races
- Optimized norm with RVV HAL
- Deprecateed MatCommaInitializer_
- Fixed v_matmul/v_matmuladd scalable semantics and expand lane-group test coverage
- Siwtched to better Durand-Kerner initialization in solvePoly
- Fixed numerical instability and out-of-bounds store in VBLAS SIMD
- Added convertScale data type conversions to RVV HAL
- Fixed inverted continuity check in cvReshapeMatND()
- Fixed memory leak in OpenCL extensions gathering
- Added check to reject nd-matrix dim count above CV_MAX_DIM in FileStorage read
- Moved IPP math functions from core module to HAL
- parallelize sort_ using parallel_for_
- Fixed potential stack overflow in FileStorage
- Vectorizeed meanStdDev
- FileStorage reads integers above INT_MAX into float/double without truncation
- Optimized minMaxIdx/minMaxLoc
- Added fast-path transposeND for identity and 2D transpose orders
- Migrateed MomentsInTile_SIMD to universal:scalable intrinsics
- Fixed Unicode temp path handling on Windows
- Vectorized cv::reduce
- Enabled ipp calls for sort and sortIdx
- Fixed use-after-scope when Mat::mul() is given a scalar
- Optimized countNonZero
- Enabled AVX-512 dispatch for sum
- SIMD optimizations for norm, distance and Hamming APIs
- Optimized gemm implementation with Universal SIMD

Imgproc module:
- Added findTRUContours - lock-free parallel contour extraction
- Fixed race condition in GaussianBlur
- BilateralFilter optimization for AVX512
- MedianBlur improvement with AVX512
- Fixed integer overflow in medianBlur
- Fixed Laplacian kernel precision for double inputs
- Fixed minEnclosingCircle O(n^3) worst case by adding Welzl shuffle
- Added SIMD support for distance transform function
- Fixed fitEllipseDirect/AMS determinant threshold for near-circular data
- Optimized tiled parallel box filter with TLS buffers and custom border fill
- Fixed ROI handling in 2D-tiled parallel execution in FilterEngine
- Fixed Apple conversions build without AVFoundation
- Fixed floodFill simple path
- Speed up first BGR2Lab and BGR2Luv call by building the color LUT faster
- Fixed missing tipLength validation in arrowedLine
- Optimized EMD (Earth Mover's Distance) solver performance
- Improved Parallelism and Memory Access Efficiency for Hough Transform Accumulator
- Fixed remap nearest rounding
- Fixed matchTemplate TM_SQDIFF precision loss for CV_8U images
- Improved color convert and AVX512 dispatch added
- Optimized LUT and equalizeHist with SIMD
- Optimized remap interpolation
- Optimized PyrDown
- Fixed IPP warpAffine
- Disabled ipp path for 32FC1 in ipp_hal_warpPerspective
- Disabled IPP implementation for warpPerspective F32C4 as it triggers exception on Windows
- Restored IPP integration for Sobel and Schaar
- Extracted following IPP integrations as HAL: resize
- Filter2D IPP migration to HAL for 4.x
- Extracted IPP integration as HAL function for box_filter
- Moved IPP-based Sobel and Scharr to HAL
- Made RVV HAL resize bit-exact for 8U INTER_LINEAR
- Added Laplacian hal for riscv-rvv:
- Added RVV HAL for spatialGradient
- Re-enabled RVV integral for safe cases
- hal/riscv-rvv: Fix test failures on hardware that implements tail-agnostic as all 1s

Imgcodecs module:
- Allow TIFF compression schemes for 32F
- imgcodecs: fix integer overflow in BMP and SunRaster decoders
- imgcodecs(webp): improve IMWRITE_WEBP_LOSSLESS_MODE to support exact lossless compression
- Fix invalid PAM decoding
- Added support for cICP metadata in imreadWithMetadata()
- Fixed JPEG chroma configuration & Standardize component settings
- Fixed UBSan load-invalid-value in SunRasterDecoder::readHeader
- Improved EXIF profile handling
- Fixed out-of-bounds write when reading 2-channel PAM

Flann module:
- Helped compiler vectorize loops in meanSplit
- Fixed Index::radiusSearch output size mismatch
- Used per-thread storage for heap pool to remove lock contention
- Bound kd-tree node indices when loading a saved flann index

Features2d module:
- Added missing KAZE DIFF_CHARBONNIER support
- Added OpenCL acceleration for BFMatcher cross-check
- OpenCL BFMatcher optimization for crosscheck case
- Added OpenCL support for AKAZE features
- Optimized FAST for RISC-V RVV
- Improved detectors parameter validation
- Fixed UB left-shift of negative value in BRISK subpixel2D

Objdetect module:
- Fix ~0.5px systematic offset in CharucoDetector subpixel refinement
- Optimize duplicated computation in QR code error correction
- Identify ArUco markers based on threshold to reduce false positives
- Rewrite getBitsFromByteList to avoid harmless buffer overflow
- Clear stale outputs in CharucoDetector::detectDiamonds
- Fixed ArUco maxCorrectionBits in predefined dictionaries
- Use validBitIdThreshold for Aruco refineDetectedMarkers
- Fixed speed regression in Aruco identify
- Optimized ArUco to get border errors in one pass for both normal and inverted markers
- Bound alphanumeric values in qr decodeAlpha before map lookup
- Bound marker string length in aruco readDictionary

DNN module:
- TopK layer K boundary check allows K == input_dim
- Parallelized Resize layer implementation
- Parallelize Slice layer implementation (4.x)
- Fixed Squeeze to remove all size-1 dims when axes is empty
- Fixed OpenVINO 2026 build failure due to ov::Tensor::data() const change
- Bound einsum subscript index before reading input shape
- DNN/ONNX: Preserve axis attribute in GatherCastSubgraph fusion
- Fixed Resize initNgraph for two-input case
- Fixed BatchNorm bias blob index in validation
- Fixed Flatten axis=rank bug
- Enable 1D int8 AvgPool with OpenVINO with 2025.3
- Vectorize SigmoidFunctor using universal intrinsics
- Fixed RandomNormalLike
- Added SIMD imp for transcendental activation layers (15 functors)
- Bound tile axis and repeats length in onnx parseTile
- Added RVV LMUL=1 kernel for convBlock_F32 (ConvolutionLayer, rv64gcv)
- Added DynamicQuantizeLinear ONNX layer support

Calib3d:
- Optimize calibrateCamera with Schur‑complement LM and parallel Jacobian accumulation
- Fixed potential crash in FMEstimatorCallback::runKernel if solveCubic cannot find solution
- docs(calib3d): clarify coordinate systems in projectPoints, undistort Points, and fisheye::undistortPoints
- Use Delaunay triangulation in computeRNG for findCirclesGrid

Video module:
- Optimized ScharrDeriv
- Added Multiscale ECC
- Added more optimizations for accumulate

VideoIO module:
- Added alpha channel support to VideoWriter and VideoCapture
- Replaced unnecessary get() usage with integer types
- Fix timestamp drift and color negotiation on Apple with GStreamer
- Fixed AvFoundation fractional fps seek
- Added CAP_PROP_IMAGE_SEQ_START open-only property for setting the first frame index when a VideoCap is opening with pattern
- Fixed crash in videoInput::start() when pVih is NULL in DSHOW backend
- Fixed heap-buffer-overflow in own MJPEG implementation
- Fixed SIGSEGV crash when releasing VideoWriter on iOS
- Fixed audio sample-rate validation and dropped last frame

Highgui module:
- highgui(cocoa): fix moveWindow Y conversion when Dock is visible
- Preserved UTF-8 window names in fallback paths with QT backend

ML module:
- SIMD optimization for KNN brute-force findNearest distance
- SIMD optimization for SVM kernel reductions

G-API:
- Added Universal Intrinsics SIMD for InRange kernel
- Optimized Fluid Select kernel using Universal Intrinsics
- Added u8 N-D mats onnx layout workaround

Python bindings:
- Fixed tests with modern numpy
- Fixed: mark distCoeffs, cameraMatrix as optional in calibrateCamera function
- Fixed inRange type signature to accept Scalar values
- Fixed HoughCircles Python return type to allow None in python
- Fixed segfault on 0-channel numpy array input
- Add context manager compatibility for VideoCapture and VideoWriter
- Fixed python wrapper around ECCParameters to include itersPerLevel

Build:
- Fixed build for Power achitecture with SIMD
- Fixed IPP management in HAL
- Enabled FP16 conversions in Windows on ARM configuration
- 3rdparty(itt): support LOONGARCH64
- add NEON support for cvFloor in fast_math.hpp on Windows ARM
- Remove bitcode support from Apple framework build scripts
- Clean up stale typing stubs during incremental builds
- Fixed IPPIW binaries search
- Fixed bundled protobuf headers being overridden by system-installed protobuf
- cmake: find and link libcblas separately for LAPACK/Generic
- js: fix Emscripten version detection and macro usage
- Modernized Doxygen comments to support for v1.15.0
- Fixed Clp library detection for vcpkg on Windows
- Updateed IPP ICV integration to IPP 2026.0.0 (20260630)
- Updated KleidiCV to verison 26.03
- Added Ubuntu 26.04 support
- Fixed OpenCV World module build with QT 5
- Used Xcode DocC tool for buiding Apple platform docs + add quick-start tutorial
- Added support for reproducible builds by making host system version optional
- Relaxed CUDA version check to major.minor in OpenCVConfig
- Fixed macOS dynamic framework RPATH for relocatable install
- Enabled non-exact IPP optimizations if build-time algorithm hint allows it

Platforms:
- Fixed resource leaks in Android Utils.java
- Fixed out-of-bounds read in WASM SIMD type conversion
- Fixed CMake crash when CUDA arch autodetection comes up empty

Applications:
- Ignore exif orientation and add option for the calibration