4results about How to "Avoid quantization error" patented technology

The present application relates to the vertical shaft cage detection technical field, specifically, it relates to a kind of industrial intelligent vision detection and diagnosis system for vertical shaft cage, acquisition unit synchronously obtains cage flange surface image and cage car pitch angle, coal dust shadow quantization unit uses bimodal adaptive threshold segmentation combined with region growing algorithm to extract coal dust area, introduce space weight factor to calculate average gray value and coverage area proportion, dynamic coupling calibration unit relies on three-dimensional look-up table and nonlinear interpolation of multi-working condition calibration, match the optimal deviation compensation coefficient of the combination of jitter and coal dust, generate dynamic gray scale segmentation threshold, defect diagnosis decision unit is through zoned adaptive segmentation, multiscale morphological filtering denoising, combined with gradient amplitude and curvature double-constraint subpixel edge detection accurately calculate wear depth, according to real-time depth and historical trend output grading control instruction, eliminate coupling interference, improve cage wear detection accuracy, provide reliable basis for safe operation.

The application discloses a precoding matrix determination method, device and equipment and a storage medium. The method comprises the following steps: sending a first instruction to a terminal; the first instruction is used for instructing the terminal to report at least a first matrix; the first matrix is used for determining a precoding matrix; receiving a first uplink reference signal sent by the terminal; obtaining a prediction result by using the first uplink reference signal and a pre-trained neural network model; the prediction result comprises a first downlink channel matrix between a network device and the terminal; and determining the precoding matrix based on the first downlink channel matrix.

This application provides a road network trajectory generation method, apparatus, and storage medium. The method includes: initializing a target sequence; performing a reverse iterative denoising step based on a denoising network trained through a discrete diffusion forward process. The iterative denoising step includes: if the current time step is not 0, inputting the current time step and the corresponding target sequence into the denoising network so that the denoising network outputs probability distribution data of each mask position in the target sequence corresponding to the current time step corresponding to a preset set of road signs; sampling each mask position in the target sequence corresponding to the current time step according to the probability distribution data to obtain the target sequence corresponding to the previous time step, and updating the current time step; if the current time step decreases to 0, determining the target sequence corresponding to the time step as the road network trajectory sequence. This application can improve the topological compliance of the generated trajectory, ensure the rationality of the generated result, avoid quantization errors, and generate high-fidelity trajectory data.

The invention discloses a bridge crack segmentation method based on hierarchical curvature mask and neural topology evolution, and the method comprises the steps: obtaining a bridge image, executing two-dimensional frequency domain decoupling preprocessing, and separating a low-frequency illumination background through a band elimination filter; calculating a Hessian matrix of the feature tensor to extract a tubular main curvature, generating a layered mask to reserve the fracture skeleton features and executing sparse reconstruction; global priori of the visual basis large model is injected into the backbone network by using a frequency domain low-rank feature adapter, and global context features are extracted in combination with a discretized selective state-space equation; introducing continuous implicit neural representation, and converting discrete pixel segmentation into probability field regression of a continuous coordinate space; the continuous coherence theory in algebraic topology is introduced in the fine tuning stage, and penalty constraint is applied to fracture disconnection errors by calculating the Wasserstein distance of a topological persistence graph. According to the method, discrete quantization errors are effectively overcome, and segmentation connectivity and precision under a complex background are improved.