.png){kind=icon}
.png){kind=icon}
Return
ALGORITHM FOR BOOK SPINE SEGMENTATION AND MATCHING BASED ON DEEP LEARNING
.png){kind=icon}
PRE
AI
DOI：10.24507/ijicic.22.02.533.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
. With the advent of smart libraries, real-time book positioning based on computer vision has gained increasing attention. These systems typically rely on segmenting book spines from shelf images and matching them to a reference database. However, due to the complexity of library environments, challenges remain in achieving high accuracy, robustness, and efficiency. This paper proposes a deep learning-based framework for accurate and efficient book spine segmentation and matching. For segmentation, an enhanced DeepLabv3 plus network is developed, where the standard Atrous Spatial Pyramid Pooling (ASPP) module is replaced with Dense Atrous Spatial Pyramid Pooling (DenseASPP) to capture richer multi-scale features. Strip Pooling is introduced to better extract elongated spine structures, while a self-attention mechanism enhances global context awareness. For matching, a deep feature matching algorithm is designed using VGG16 for feature extraction and Euclidean distance for similarity computation. The Facebook AI Similarity Search (Faiss) framework is integrated to accelerate large-scale retrieval. Experiments were conducted on two datasets constructed from segmented spine images: one comprising books from the same series and the other from different series. The proposed method achieved over 95% matching accuracy with an average processing time of 2.33 seconds per sample on both datasets, demonstrating strong robustness and real-time potential.
Keywords:
Book spine segmentation and matching
Smart library
DeepLabv3 plus
DenseASPP
Faiss
Journal
I
IF:
1.1
Papers:
79
Citations:
1.2K