Detection of Dust Deposition Using Convolutional Neural Network for Heritage Images

Abstract

This paper presents a vision-based approach for heritage image classification and condition monitoring to preserve the historical facts. The proposed approach uses convolutional neural network for classification. The approach interprets the heritage condition in terms of dust level. Initially, real-time scene image is preprocessed using image processing operators such as dilation, erosion, region filling, and binarization. Resultant image is segmented and enclosed by bounding boxes. The enclosed segments are fed to CNN for classification. The proposed approach also provides the dust level in image by comparison of probability score of the classified image with ideal one. The dust is interpreted as Gaussian noise in the image. The dust level, greater than an acceptable tolerance level, generates a notification for heritage maintenance. Results show that the proposed approach is able to classify the heritage image in the presence of noise.