Leveraging CNNs with Auto-Encoders for Detection of Shadows in Optical Images

Shadow detection is a crucial and evolving problem in the domain of computer vision. It has high implications in autonomous vehicles to prevent fatalities and accidents due to detection of shadows as real objects. This paper introduces a novel framework for shadow detection in images, inspired by auto-encoder principles and based on encoding-decoding techniques. The proposed architecture employs handcrafted Convolutional Neural Networks (CNNs) rather than relying on pre-trained networks, allowing it to automatically learn significant features of shadow areas from input images. The proposed approach is trained and tested on publicly available SBU dataset. The proposed approach achieved an accuracy of 87.20% and SSIM of 0.85. The proposed CNN and autoencoder-based shadow detection method can be used with driverless vehicles for autonomous detection of shadow regions and prevent accidents due to false detection of shadows as a real object.