Methods and systems of generating pseudo lesion masks. One system includes an electronic processor configured to receive an annotated medical image, the annotated medical image including a bounding box annotation positioned around at least one lesion of the medical image. The electronic processor is also configured to generate, using a ground truth generator, a pseudo-mask candidate, the pseudo-mask candidate representing a pseudo lesion mask for the at least one lesion of the medical image. The electronic processor is also configured to train a segmentation model using the pseudo-mask candidate as ground truth.

Techniques are described for inpainting of image data with a missing region. In an embodiment, at each iteration, the process determines a corresponding missing boundary region of the missing region and generates a collection of boundary patches for the missing boundary region. Based on comparing a boundary patch from the collection to source patches from a known source region of image data, the process generates replacement patches for the missing boundary region. When a boundary pixel data unit corresponds to multiple replacement pixel data units from different replacement patches, the process aggregates the multiple replacement pixel data units to generate an updated boundary pixel data unit. In an embodiment, the process performs convolution using the updated and previously known region of the image data.

A method for predicting a time to a failure condition of an asset receives image pairs of a healthy condition asset and historical image pairs of the asset from inspections and at a failure condition of the asset. A first model is trained to rebuild input images to resemble a healthy condition asset. Similarity coefficients are generated for the historical image pairs of a regular and infrared image by use of the rebuilding of images by the first model as a similarity base and historical image pairs include a timestamp of image capture. A second model is trained to predict a time to an asset failure condition based on timestamps of image capture, timestamps of an asset failure condition, and similarity coefficients of the historical images. Responsive to receipt of a real-time image pair, the method computes a predicted time to a failure condition of the asset.

Methods and systems for training computer-aided condition detection systems. One method includes receiving a plurality of images for a plurality of patients, some of the images including an annotation associated with a condition; iteratively applying a first deep learning network to each of the images to produce an attention map, a feature map, and an image-level probability of the condition for each of the images; iteratively applying a second deep learning network to each feature map produced by the first network to produce a plurality of outputs; training the first network based on the attention map produced for each image; and training the second network based on the output produced for each of the patients. The second network includes a plurality of convolution layers and a plurality of convolutional long short-term memory (LSTM) layers. Each of the outputs includes a patient-level probability of the condition for one of the patients.

A computer vision-based system provides for load distribution estimation and load rating and vibration serviceability assessment of structures. The system integrates evaluates the structural load carrying capacity, the diagnosis and prognosis of performance and safety, and vibration serviceability. Cameras record images of a structure, and regions of interest are monitored in those images for their displacement and velocity as loading varies. Where the displacement determined exceeds a predetermined threshold, or where the acceleration determined exceeds predetermined limits, or where the distribution of displacements of parts of the structure deviates substantially from an estimated displacement distribution, an output indicating potential problems with the structure is output.

A method for detecting defects in images, is employed in a computer device, and stored in a storage medium. The method trains an autoencoder model using unblemished images, inputting an image to be detected into the autoencoder model, and obtaining a reconstructed image. An image error is calculated between the image to be detected and the reconstructed image, and the image error is inputted into a student's t-distribution and a calculation result is obtained. In response that the calculation result falls within a preset defect determination criterion range, the image to be detected is determined to be an unblemished image. In response that the calculation result does not fall within the preset defect determination criterion range, the image to be detected is determined to be a defective image. The method improves the efficiency and accuracy of defect detection.

Methods and systems for training computer-aided condition detection systems. One method includes receiving a plurality of images for a plurality of patients, some of the images including an annotation associated with a condition; iteratively applying a first deep learning network to each of the images to produce a segmentation map, a feature map, and an image-level probability of the condition for each of the images; iteratively applying a second deep learning network to each feature map produced by the first network to produce a plurality of outputs; training the first network based on the segmentation map produced for each image; and training the second network based on the output produced for each of the patients. The second network includes a plurality of convolution layers and a plurality of convolutional long short-term memory (LSTM) layers. Each of the outputs includes a patient-level probability of the condition for one of the patients.

A digital video computing system receives two or more frames depicting an environment from a camera system of a vehicle. For a salient image feature identified in the two or more frames, a global motion vector is calculated that is indicative of movement of the feature at least partially attributable to movement of the vehicle. A local motion vector is calculated that is indicative of movement of the feature independent from the movement of the vehicle. Based on the local motion vector, the salient image feature is determined to have an apparent motion relative to the environment that is independent from the movement of the vehicle. A candidate image patch is identified including the salient image feature. The candidate image patch is analyzed to output a likelihood that the candidate image patch depicts a second vehicle.

Images of optical codes are mapped to an overview image to localize optical codes within a space. By localizing optical codes, information about locations of various products can be ascertained. One or more techniques can be used to map the images of optical codes to the overview image. The overview image can be a composite image formed by stitching together several images.

A method for selecting one or more targets for non-invasively treating a cardiac arrhythmia in a patient includes receiving a mapping associated with the patient's heart and generating a segmented model of the mapping associated with the patient's heart. The segmented model divides the mapping into a plurality of segments. The method includes identifying one or more abnormality in the segmented model of the mapping associated with the patient's heart, determining which segment or segments of the plurality of segments include the identified one or more abnormality, and selecting a target for non-invasive treatment of the cardiac arrhythmia based on the determined segment or segments of the plurality of segments that include the identified one or more abnormality.