Tricolor Attenuation-Based Shadow Detection and Removal Using Feature Descriptors

The suggested work presents TAM-FD, a creative extension of the tricolor attenuation model designed to tackle the difficult problem of shadow detection in pictures. Using restricted contextual reasoning via pair wise potentials in Conditional Random Fields, conventional shadow detection techniques often emphasize examining the local appearance of shadow areas. The suggested method, on the other hand, captures global scene features and higher-level relationships. The shadow detector has been trained to operate as a generator of a conditional TAM (Type Attention Model) with an aim to improve shadow detection accuracy by combining the usual TAM loss with another data loss term based on feature descriptors. Shadows occur as a result of certain objects blocking direct illumination which usually comes from the sun. They are divided into cast shadows which are created by projections from an object toward a light source, and self shadows which form parts of the object not directly illuminated. Cast shadows can be sub-divided into umbra which is the region of total blockade of light and penumbra which is a region of partial blockade of light resulting in blurred boundaries between shadowed and non-shadowed areas. Shadowed regions encompass a loss of irradiance, and amplify the difficulty of imaging tasks such as interpretation, object detection and recognition, and change detection due to the loss of information presented. In all cases evaluated for performance, the proposed method yielded improvements in quantifiable indices such as 1.76% in the shadow detection index SDI, 9.75% in the color component index for preserving color contrast during shadow removal, and an increase of 1.89% in the normalized saturation value detection index NSVDI which indicates precision in recognizing shadow regions.