Autoencoders in Graphical Document Retrieval: Cost Function Influence and Rotation Impact

This study explores autoencoder application in graphical document retrieval, focusing on rotation’s impact. Specific to this paper are scanned floor plan images. We investigate the pivotal role of cost functions—specifically, Mean Squared Error (MSE) and Consistency Loss—in shaping autoencoder behavior and retrieval accuracy. By subjecting autoencoders to the challenges posed by rotated graphical documents, we decipher the extent to which rotation impacts retrieval outcomes. Concurrently, we dissect the nuanced influence of distinct cost functions on retrieval accuracy, shedding light on their diver-gent contributions. Through rigorous experimentation, we show autoencoder adaptability to rotation and cost function effects on retrieval precision. Our findings underline the significance of cost function selection in enhancing retrieval accuracy. Our insights hold implications for optimizing, refining autoencoder-based methods, empowering digital document analysis practitioners.