Architecture for improved machine learning operation

#The National Imagery Interpretation Rating Scale (NIIRS) is graded from 0 to 9, where 0 is the coarsest resolution and 9 is the finest resolution. The NIIRS was developed using human perception of image quality to perform imagery analysis tasks. The image quality is determined based on one or more attributes of the image. The imagery attributes are the physical characteristics of the image such as the resolution, the relative edge response, and signal-to-noise. The analysis tasks can include, for example, determining whether a door is open or closed, identifying a type of automobile, distinguishing individual steps, or identifying grill detail or a license plate number on a license plate of a vehicle.

#FIG. 1 illustrates, by way of example, a diagram of an embodiment of a system for improving a machine learning (ML) technique.

#FIG. 2 illustrates, by way of example, a diagram of an embodiment of a method for improving an ML technique.

#FIG. 3 illustrates, by way of example, a block diagram of an embodiment of a machine (e.g., a computer system) to implement one or more embodiments.

#Some embodiments relate to automatically identifying an image quality. The image quality that is automatically identified can be used to remove images of insufficient quality for a specified purpose. For example, training a ML target model with images below the quality for the specific recognition task will generate poorly performing models regardless of training set volume. Likewise, inferencing images of insufficient quality to perform a given task will likely lead to spurious results generated by the trained target model.

#Given some quantitative image quality metrics (e.g., resolution, signal to noise ratio (SNR), or the like) a NIIRS value can be predicted using the a general image quality equation (GIQE). The analysis tasks possible using the image can be determined using a National Imagery Interpretability Rating Scale (NIIRS). NIIRS is a subjective scale used for rating the quality of imagery acquired from various types of imaging systems. The NIIRS was developed using a large number of trained imagery analysts. Individual NIIRS have been developed for visible band images, radar images, infrared images, and multispectral images. With NIIRS, analysis tasks are associated with a minimum image quality required to perform the analysis task.

#Automatically determining the quality of an image can include using object recognition technology, sometimes a convolutional neural network (CNN). Since the NIIRS is a subjective scale that defines quality of imagery based on objects that are distinguishable in the image, an object recognition technique (or multiple object recognition techniques) can be used to determine the quality. The following table summarizes the rating using a NIIRS rating scale on visible imagery. Sometimes, sufficient information to determine the GIQE rating for a image is provided with the image, such as part of a header (sometimes called metadata) of the image.