Custom Lightweight Convolutional Neural Network Architecture for Automated Detection of Damaged Pallet Racking in Warehousing & Distribution Centers

Abstract

This paper proposes a Convolutional Neural Network-Block Development Mechanism (CNN-BDM) enabling the development of a lightweight deep learning architecture for the detection of damaged pallet-racking, within the manufacturing/warehousing environment. The developed CNN architecture consisted of only 6.5 Million learnable parameters, making it the first custom designed CNN architecture for the pallet racking domain. Architectural training was based on a real dataset collected from various warehouses after implementation of several data modelling strategies for scaling and increasing the variance within the dataset, in a representative manner. Additionally, after achieving a baseline accuracy of greater than 90%, various regularization strategies were applied for further enhancing the performance and generalizability of the network. Dropout at a drop rate of 50% provided the highest performance during training, achieving 99% precision, recall and F1 score. The effectiveness of the proposed methodology was manifested by the fact that the architecture was able to maintain high performance on the test data achieving an overall F1 score of 96%.