|
A current approach for performance assessment of imagers is triangle orientation discrimination (TOD). This approach requires observers or human visual system (HVS) models to recognize equilateral triangles pointing in four different directions. Imagers may apply embedded advanced digital signal processing (ADSP) for contrast enhancement, noise reduction, edge sharpening, etc. Unfortunately, applied methods are in general not documented and hence unknown. Within the last decades a vast amount of techniques for contrast enhancement has been proposed. There are some comparisons of such algorithms for few images and figures of merit. However, many of these figures of merit cannot assess usability of altered image content for specific tasks such as object recognition In this work different algorithms for contrast enhancement are compared in terms of TOD assessments by convolutional neural networks (CNN) as models. These models are trained by artificial images with single triangles. Many methods for contrast enhancement highly depend on the content of the entire image. Therefore, the images are superimposed by natural backgrounds with varying standard deviations to provide different signal-to-background ratios. Then these images are degraded by Gaussian blur and noise representing degradational camera effects and sensor noise. Different algorithms are applied, such as the contrast-limited adaptive histogram equalization or local range modification. Then accuracies of the trained models on these images are compared for different ADSP algorithms. Accuracy gains for low signal-to-background ratio and sufficiently large triangles are found, while impairments are found for high signal-to-background ratio and small triangles. Finally, implications of replacing triangles by real target signatures when using such ADSP algorithms are discussed. The results can be a step towards the assessment of those algorithms for generic target recognition.
|
