33results about How to "Improve detection and recognition accuracy" patented technology

The invention discloses a method and a system for detecting fatigue driving based on head and neck movement feature recognition of a driver. The method comprises the following steps: (1) arranging a camera on the side face of a driving position in a car, and collecting images of the head and neck part of the driver from the side face; (2) storing a group of images collected by the camera in the step (1) through an image processing unit at intervals of time T, and abstracting contour lines of head and neck side images of the driver; and (3) forming head and neck movement features of the driver according to a plurality of groups of continuous contour lines of the head and neck side images of the driver, wherein the continuous contour lines of the head and neck side images of the driver are obtained from the step(2), comparing the head and neck movement features of the driver with the head and neck movement features when a human body is in a sleepy state, carrying out judgment, if the head and neck movement features of the driver conform to sleepy state requirements, considering the driver to be fatigue driving and sending an alarm prompt, or else, repeating the step (2), and keeping operation continuously and circularly. The method and the system for detecting fatigue driving based on head and neck movement feature recognition of the driver have the advantages that detecting recognition accuracy is high, the requirements for an image processing algorithm and corresponding hardware are low, and implementation is easy.

The embodiment of the present invention discloses a training method and device for a target classification model. The method includes: obtaining a training image and its corresponding calibration information; The first binary feature extraction operation obtains the corresponding output image; if the convolution layer is a binary convolution layer, the first binary feature extraction operation includes: the first binary feature extraction operation and preset dimension transformation Operation; the input image corresponding to the first convolutional layer is the training image, and the input image corresponding to the other convolutional layers is the output image corresponding to the previous convolutional layer; use the feature classification layer of the model and the last convolutional layer For the corresponding output image, determine the prediction detection information of the training image corresponding to the output image; combine the calibration detection information corresponding to each training image, adjust the model parameters, and determine the target binary classification model to achieve high recognition and detection accuracy and calculation A small target classification model.

The invention discloses a semi-supervised hyperspectral sub-pixel target detection method based on an enhanced constraint sparse regression method. According to the enhanced constraint sparse regression method, the precision and the stability for decomposing a hyperspectral mixed-pixel are improved; a statistical model for distinguishing a target pixel and a background pixel is constructed by applying a generalized likelihood radio checking theory; and furthermore, the enhanced constraint sparse regression method and the generalized likelihood radio checking theory are combined, so that the target distribution can be quantitatively detected, and the detection efficiency can be improved effectively, and the mis-alarm probability is reduced.

The invention provides a SAR remote sensing image oil spill detection and recognition method, the specific process is as follows, using the Gamma MAP filter to filter the SAR image, and then performing Sobel filtering on it; performing a watershed algorithm on the gradient map obtained after the Sobel filtering to realize sea and land segmentation ;Use the mean value of the sea area image to fill the land area, and then use the C-V algorithm to segment and extract the target area in the homogeneous area of ​​the filled image; extract the gray level co-occurrence matrix of the target area, the texture characteristics of wavelet decomposition, gray The visual frequency histogram is constructed using the degree and shape features, and the trained SVM classifier model is used to classify the visual frequency histogram, and the suspected oil spill area is removed from the target area to realize the initial false alarm elimination; the result of the initial false alarm elimination As the initial label field, based on the initial label field, the feature field in the MRF context model is used to further eliminate false alarms, so as to realize the oil spill detection and recognition method of the SAR remote sensing image.

The invention discloses a cloud system for automatic identification and detection of underground pipe networks based on deep learning, including an underground pipe network defect collection module, a public data set collection module, a typical defect sample library of the underground pipe network, a typical defect analysis and classification module, and an underground network management installation. Environmental geological state data collection module, deep learning module, automatic identification and detection module of underground pipe network and decision-making module; by establishing a typical defect sample library of underground pipe network, using the known multiple underground pipes contained in the typical defect sample library of underground pipe network The typical defects of the network and the data collected by the public data set collection module are used as the data set for model training. The automatic identification and detection of the underground pipe network obtained after training The LexNet network model has high identification and detection accuracy, and the automatic identification and detection speed of the underground pipe network is fast. The test results are of stable quality and high reliability.