37results about How to "Rich image features" patented technology

The invention discloses a weld defect recognition method based on the improved LeNet-5 model. Firstly, for the weld gray-scale image, the input of the traditional convolution kernel channel of the LeNet-5 model is improved, and the gray-scale image is passed through pseudo-color The enhancement technology is converted into a color image, and the obtained color image is used as the input of the neural network; then the convolution kernel of the LeNet-5 model is improved, and the convolution kernel channel with the Gabor filter is added; in the sixth layer of the neural network, The features obtained from multiple channels are fused to obtain the feature set T; finally, the SoftMax classifier is used in the seventh layer (output layer) of the neural network to obtain the defect type of the weld and the probability of each category, which is used for evaluation. It provides a reference for film personnel to determine the type of film and formulate on-site repair plans. The invention improves the feature extraction ability of the neural network, thereby improving the correct rate of defect recognition; the recognition result is given by the probability that the defect belongs to a certain category, and provides more sufficient reference information for reviewers.

The present invention provides an image representation method based on phase singularity, which includes: performing two-dimensional complex filter processing on image information to generate a filter response image; detecting image phase singularity; accurately locating phase singularity; generating a localized phase singularity Descriptors serve as image features. The present invention also provides a method for image matching and recognition using the above image representation. The image representation method of the present invention effectively utilizes the information of phase singularity, develops a new image representation method, enriches image features, improves the number of image matching points and the accuracy of matching, and develops a method based on phase singularity packet representation Image classification methods.

The invention provides a liquid state recognition method in liquid separation and a liquid separation system, wherein the liquid separation system executes the liquid state recognition method, and the method includes: starting from entering the liquid separation process, starting to acquire the video of the liquid in the pipeline, and The image frame in the video is extracted every preset period; for each image frame, the image features of the area to be identified in the image frame are obtained, and the image features are input into the trained liquid state recognition model to obtain the corresponding image frame. Predicted state, the trained liquid state recognition model is obtained by training sample image frames with liquid state labels; based on the predicted state corresponding to each image frame, the liquid state in the pipeline is determined. The liquid state recognition model used in this solution is especially suitable for the fast-flowing liquid dispensing process. Using machine learning models to participate in liquid state prediction can make the analyzed image features richer and the algorithm more robust, making liquid dispensing more accurate. Better performance and versatility.

The invention discloses a speckle-based three-dimensional human face reconstruction method and equipment, which can increase the richness of image features and at the same time reduce the amount of matching calculations, and achieve synchronous improvement in three-dimensional human face reconstruction efficiency and reconstruction accuracy. The method includes: inputting a plurality of speckle image pairs conforming to binocular stereo vision; performing epipolar correction on the input speckle image pairs so that corresponding matching points are on the same horizontal line; Face detection, extracting the image pair of the area of ​​interest of the face to be reconstructed; performing the first matching process on the image pair of the area of ​​interest to obtain the first disparity map; setting the matching point search range according to the disparity in the first disparity map, The region image is subjected to a second matching process to obtain a second disparity map; corresponding 3D point cloud data is generated according to the second disparity map and camera calibration parameters to obtain a reconstructed 3D face.

The invention discloses a fried food detection system based on a symbiotic double-stream convolutional network and a digital image, including an image preprocessing module, a fast identification module, a classification and positioning module, a target cutting module and an image analysis module connected in sequence; an image preprocessing module Perform image stylization migration and image filtering processing on the input image in turn to obtain image tensors available to the network; in the fast recognition module, the image is quickly classified by a fully convolutional network composed of a co-occurrence feature extraction network and a recognition network; the classification and positioning module It is a fully convolutional network; the target cropping module uses the optimal frame to cut the target image from the original image; the image analysis module analyzes the target image and gives quantitative analysis results. The invention combines the symbiotic double-stream convolutional network and digital image analysis, and can realize fast and accurate location and attribute identification of fried food.