Authors:
S. Gopikha, V. S. Anusuya, Thirumalraj Karthikeyan, Vijilius Helena Raj, Sureshkumar Somayajula
Addresses:
Department of Information Technology, St. Joseph's College of Engineering, Chennai, Tamil Nadu, India. Department of Chemistry, New Horizon College of Engineering, Bengaluru, Karnataka, India. Department of Artificial Intelligence, Trichy Research Labs, Quest Technologies, Tiruchirappalli, Tamil Nadu, India. Department of Mathematics, New Horizon College of Engineering, Bengaluru, Karnataka, India. Department of Computer Science and Technology, Sunlife Canada Financials, Toronto, Ontario, Canada.
Abstract:
Considering the difficulties in identifying pear leaf diseases due to factors such as varying lighting conditions, overlapping leaves, and other green plants in the background, a two-stage framework based on strategies is suggested for segmenting pear leaves and disease categorisation. To begin, the target damaged pear leaf is extracted, and background interference is eliminated by building a DBPN that fuses the low-level feature branch and the semantic branch. Research on transformer-based models has grown in recent years, with several studies showing promising results. Unfortunately, transformers still have issues with edge detail segmentation and small object recognition. To address these issues, the research enhanced the Swin transformer by combining the best features of transformers and CNNs. It then developed an LPSW backbone to boost the network's local perception and classification task discovery accuracy. A framework for an SAIEC system was also developed as part of the study; this contributed to the network's enhanced accuracy. To fine-tune the parameters of the proposed classifier, this paper proposes using improved beetle swarm optimisation. The field dataset DiaMOS Plant, comprising 3505 images of pear fruit and leaves damaged by four illnesses, is used to evaluate the proposed model in this study. The dataset is publicly available and was collected to identify and monitor plant problems. both segmentation and classification.
Keywords: Spatial Attention Interleaved Execution Cascade (SAIEC); Local Perception Swin transformer (LPSW); Edge Detail Segmentation; Beetle Swarm Optimisation; Pear Leaf Disease Detection; Convolutional Neural Network (CNN); Double-Branch Polymerisation Net (DBPN).
Received: 15/08/2024, Revised: 30/09/2024, Accepted: 21/11/2024, Published: 05/09/2025
DOI: 10.64091/ATICL.2025.000227
AVE Trends in Intelligent Computer Letters, 2025 Vol. 1 No. 3 , Pages: 119-131