Refine
Has Fulltext
- no (1)
Year of publication
- 2023 (1)
Document Type
Language
- English (1)
Is part of the Bibliography
- yes (1)
Keywords
- Deep learning (1)
- Stacking problem (1)
- Warehouse safety (1)
Institute
Purpose of this paper:
Image processing technology has gained much improvement in recent years. In smart warehousing, most focus is on warehouse operations to improve efficiency and save human labors. However, in put-away, re-warehousing or picking process, after being handled by forklifts or laborers, the stacking of inventory items may change its shape and balance due to deviation, compression, tilting, etc., which may jeopardize the safety of stored goods and even people working in the warehouse. If there is a system that could monitor the stacking of inventory items all the time and could alarm managers should the stacking tilts or deforms to a certain level, inventory safety could be vastly improved. This study is on smart warehouse safety problems emphasizing at goods stacking to avoid the possibility of falling or collapse through image processing via Mask R-CNN. Our experiment is limited to items that form a rectangular block on a pallet, such as wooden bars, cartons or any box-shaped items.
Design/methodology/approach STS:
Mask R-CNN is a deep learning software and can be used in a wide range of applications. In this study, we used Mask R-CNN to detect the images of stacked items to see if the stacking is stable. We took photos from a warehouse and divided them into three classes – training, validation and production. VGG Image Annotator (VIA) was used for object annotation. To detect the tilt degree of stacked items, we derived formulation from book- stacking problems and retrieved dot coordinates of square frames encompassing objects in an image from the result of Mask R-CNN for calculation.
Findings:
Because there is a limited number of photos for training, validation and production, we applied image enhancement techniques such as flipping, affinity, multiplication, etc. to increase the total number of images to avoid over-fitting during the training process. The trained Mask R-CNN network demonstrated its capability in framing objects and after retrieving the dot coordinates of the frames, it could compute the degree of tilt as a reference for block stacking alarm.
Value:
The paper presents a relatively rare study of applying deep learning techniques on images as applied to material stacking safety problems. It contributes to a special smart warehousing domain. The same technique can be applied to inventory counting and help improve labor-saving in counting and in locating items in a warehouse.
Research limitations/implications:
Due to the limited number of images and restricted angles of taking the images, Mask R- CNN may not be so capable in framing objects of special angles of images. More images with diverse angles of them should be obtained for training to improve the capability of the trained network.
Practical implications:
Successful implementation of the technology developed in this study in a warehouse monitoring system can help industries in smart detection of improper stacking to assist human in improving the accuracy of safety judgement and can also save labors in reviewing images or records of a warehouse.