Many machine learning algorithms have been introduced to solve different types of problem. Recently, many of these algorithms have been applied to deep architecture model and showed very impressive performance. In general, deep architecture model suffers from over-fitting problem when there is a small number of training data. In this paper, we attempted to remedy this problem in deep architecture with regularization techniques including overlap pooling, flipped-image augmentation and dropout, and we also compared a deep structure model (convolutional neural network (CNN)) with shallow structure models (support vector machine and artificial neural network with one hidden layer) on a small dataset. It was statistically confirmed that the shallow models achieved better performance than the deep model that did not use a regularization technique. However, a deep model augmented with a regularization technique-CNN with dropout technique-was competitive to the shallow models.
15 February 2020 in Journal of Ambient Intelligence and Humanized Computing
Evaluation of car damages from an accident is one of the most important processes in the car insurance business. Currently, it still needs a manual examination of every basic part. It is expected that a smart device will be able to do this evaluation more efficiently in the future. In this study, we evaluated and compared five deep learning algorithms for semantic segmentation of car parts. The baseline reference algorithm was Mask R-CNN, and the other algorithms were HTC, CBNet, PANet, and GCNet. Runs of instance segmentation were conducted with those five algorithms. HTC with ResNet-50 was the best algorithm for instance segmentation on various kinds of cars such as sedans, trucks, and SUVs. It achieved a mean average precision at 55.2 on our original data set, that assigned different labels to the left and right sides and 59.1 when a single label was assigned to both sides. In addition, the models from every algorithm were tested for robustness, by running them on images of parts, in a real environment with various weather conditions, including snow, frost, fog and various lighting conditions. GCNet was the most robust; it achieved a mean performance under corruption, mPC = 35.2, and a relative degradation of performance on corrupted data, compared to clean data (rPC), of 64.4%, when left and right sides were assigned different labels, and mPC = 38.1 and rPC = $$69.6\%$$ when left- and right-side parts were considered the same part. The findings from this study may directly benefit developers of automated car damage evaluation system in their quest for the best design.