Abstract: A deep recognition model training method applied to an electronic device is provided. The method includes obtaining a ground plane area by segmenting a first image using a ground plane segmentation network. A projection image of the first image is generated based on the first image, an initial depth image corresponding to the first image, and a pose matrix. A target height loss of a depth recognition network is generated, and a depth loss of the depth recognition network is obtained according to a gradient loss between the initial depth image and the first image and a photometric loss between the projection image and the first image. A depth recognition model is obtained by adjusting the depth recognition network based on the depth loss and the target height loss.

Abstract: A method for training a depth recognition model implemented in an electronic device includes determining test objects from test images, and obtaining a first image and a second image; calculating a test projection slope of each test object according to coordinates of each pixel point of each test object in the test images; generating a threshold range according to the plurality of test projection slopes; recognizing a type of terrain corresponding to a position of each initial object; adjusting an initial ground area in the first image, and obtaining a target ground area in the first image; generating a target height loss of a preset depth recognition network, an initial depth image corresponding to the first image, and the target ground area; and adjusting the preset depth recognition network according to the target height loss and a depth loss, and obtaining a depth recognition model for recognizing depth of images.

Abstract: An assistance method of safe driving applied in a vehicle-mounted electronic device obtains RGB images of scene in front of a vehicle, processes the RGB images by a trained depth estimation model, obtains depth images and converts the depth images into three-dimensional (3D) point cloud maps, determines 3D regions of interest therein, and obtains position and size information of objects in the 3D regions of interest. When the position information satisfies a first preset condition and/or the size information satisfies a second preset condition, the presence of obstacles in the 3D regions of interest is determined and controls the vehicle to issue an alarm. When the position information does not satisfy the first preset condition and/or the size information does not satisfy the second preset condition, the 3D regions of interest are determined as obstacle-free, and permitting the vehicle to continue driving.

Abstract: A method for managing driving applied in an electronic device which assesses distances to objects in a path of autonomous driving obtains RGB images of a scene in front of a vehicle, processes the RGB images based on a trained depth estimation model, and obtain depth images corresponding to the RGB images. The depth images are converted to 3D point cloud maps, 3D regions of interest from the 3D point cloud maps are determined according to a size of the vehicle, and the 3D regions of interest are converted into 2D regions of interest according to internal parameters of a camera. The 2D regions of interest are analyzed for obstacles. Driving continues when the 2D regions of interest have no obstacles, the vehicle is controlled to issue an alarm when obstacles are discovered.

Abstract: A method for detecting defect of images applied in an electronic device inputs flawless sample training images into an autoencoder, and calculates first latent feature by a coding layer of the autoencoder, and calculates first reconstructed images by a decoding layer, and calculates a first reconstruction error by a first preset error function. The electronic device trains the discriminator according to the flawless sample training images and first reconstructed images, and calculates an adversarial learning error, and calculates a sample error, determines an error threshold based on the sample error, and obtains testing sample images, and calculates second latent feature of the testing sample images by the coding layer, and calculates the second reconstructed images of the testing sample images by the decoding layer, and calculate a difference between the testing sample images and the second reconstructed images, thus a detection result of the testing sample images is determined.

Abstract: A vehicle-borne method for recognizing the illumination state of traffic lights even against a backlighting of strong sunlight or other light source obtains a first image of a set of traffic lights in a road traffic environment. A segmentation map is acquired by dividing a first region from the first image, and an illumination region in the segmentation map is extracted by marking RGB pixels in the region which are of a preset threshold in brightness according to a training model. A lit color of the set of traffic lights is recognized according to a position of the illumination region in the segmentation map. By utilizing the method, accuracy of recognition of illumination state of traffic lights is improved.

Abstract: A method for detecting road conditions applied in an electronic device obtains images of a scene in front of a vehicle, and inputs the images into a trained semantic segmentation model. The electronic device inputs the images into a backbone network for feature extraction and obtains a plurality of feature maps, inputs the feature maps into the head network, processes the feature maps by a first segmentation network of the head network, and outputs a first recognition result. The electronic device further processes the feature maps by a second segmentation network of the head network, and outputs a second recognition result, and determines whether the vehicle can continue to drive on safely according to the first recognition result and the second recognition result.

Abstract: A method of detecting defects revealed in images of products obtains sample image training data. An underlying feature dimension of an autoencoder is selected and a score is obtained. By comparing the score with a standard score, an optimal underlying feature dimension is confirmed. A test image is inputted into the autoencoder with the optimal underlying feature dimension to obtain a reconstruction image. A reconstruction error between the test image and the reconstruction image is computed. By comparing the reconstruction error with the predefined threshold a result of analysis of the test image is outputted. An image defect detection apparatus and a computer readable storage medium applying the method are also provided.

Abstract: An image defect detection method is used in an electronic device. The electronic device determines a training image feature set, and trains a Gaussian mixture model by using the feature set to obtain an image defect detection model and a reference error value. An image for analysis is input into the autoencoder to obtain a second implicit vector and a second reconstructed image, and to calculate a second reconstruction error. The electronic device obtains a test image feature of the image for analysis according to the second reconstruction error and the second implicit vector, and inputs the test image feature into the image defect detection model to obtain a prediction score. The image for analysis is determined to reveal a defect when the prediction score is less than or equal to the reference error value.

Abstract: An image defect detection method used in an electronic device, calculates a Kullback-Leible divergence between a first probability distribution and a second probability distribution, and thereby obtains a total loss. Images of samples for testing are input into an autoencoder to calculate a second latent features of the testing sample images and the second reconstructed images. Second reconstruction errors are calculated by a preset error function, as is a third probability distribution of the second latent features, and a total error is calculated according to the third probability distribution and the second reconstruction errors. When the total error is greater than or equal to the threshold, determining that the images of samples for testing reveal defects, and when the total error is less than the threshold, determining that the images of samples for testing reveal no defects.

Abstract: In a method for defecting surface defects, a trained weighting generated when defect-free training samples are used to train an autoencoder and pixel convolutional neural network is obtained. A test encoding feature is obtained by inputting the trained weighting into the autoencoder and pixel convolutional neural network and a weighted autoencoder of the weighted autoencoder and pixel convolutional neural network encoding a test sample. The test encoding feature is input into a weighted pixel convolution neural network of the weighted autoencoder and pixel convolutional neural network to output a result of test. The test result is either no defect in the test sample or at least one defect in the test sample. Inaccurate determinations as to defects are thereby avoided. An electronic device and a non-transitory storage medium are also disclosed.

Abstract: A method applied in an electronic device for detecting and classifying apparent defects in images of products inputs an image to a trained autoencoder to obtain a reconstructed image, determines whether the image reveals defects based on a defect criterion for filtering out small noise reconstruction errors. If so revealed, the electronic device calculates a plurality of structural similarity values between the image and a plurality of template images with marked defect categories, determines a target defect category corresponding to the highest structural similarity value, and classifies the defect revealed in the image into the target defect category.

Abstract: In a method for defecting surface defects, a weighting generated when defect-free training samples are used to train an autoencoder and pixel convolutional neural network is obtained. A test encoding feature is obtained by inputting the weighting into the autoencoder and pixel convolutional neural network and encoding a test sample with a weighted autoencoder of the weighted autoencoder and pixel convolutional neural network. The test encoding feature is divided into many sub-test encoding features. The sub-test encoding features are input into a weighted pixel convolution neural network of the weighted autoencoder and pixel convolutional neural network one by one to output a result of test, the test result being either no defect in the test sample or at least one defect in the test sample. Inaccurate defect determinations are avoided, and accurate determinations even of fine defects improved. An electronic device and a non-transitory storage medium are also disclosed.

Abstract: A model input size determination method, an electronic device and a storage medium are provided, the method includes acquiring a plurality of test images and a defect result; and encoding each test image to obtain an encoding vector. The encoding vector is decoded to obtain a reconstructed image, then a reconstruction error and a plurality of sub-vectors are calculated; the plurality of sub-vectors is inputted into a Gaussian mixture model, then a plurality of sub-probabilities, an estimated probability and a test error are determined; a detection result in the test image according to the test error and the corresponding error threshold are obtained; an accuracy according to the detection result and the defect result are determined, and an input size is selected from the plurality of preset sizes according to the accuracy. An accuracy of defect detection in manufacturing can be improved.

Abstract: A defect detection method based on an image of products and an electronic device can accurately determine the error threshold by determining the reconstruction error generated during image reconstruction and by determining the estimated probability generated by the Gaussian mixture model. The test error can then be compared with the error, since the test error and the error threshold are compared numerically, the existence of subtle defects are revealed in the product image, thereby improving the accuracy of defect detection.

Abstract: A method for detecting defects in multi-scale images and a computing device applying the method acquires a to-be-detected image and converts the to-be-detected image into a plurality of target images of preset sizes. Feature extraction is performed on each target image by using a pre-trained encoder to obtain a latent vector, the latent vector of each target image is inputted into a decoder corresponding to the encoder to obtain a reconstructed image and then into a pre-trained Gaussian mixture model to obtain an estimated probability. Reconstruction error is calculated according to each target image and the corresponding reconstructed image. A total error is calculated according to the reconstruction error of each target image and the corresponding estimated probability, and a detection result is determined according to the total error of each target image and a corresponding preset threshold, thereby improving an accuracy of defect detection.