OICC Press Majlesi Journal of Electrical Engineering 2345-3796 2026 07 30 An Improved Deep Mask Region-based Convolutional Neural Network for the Detection of Power Transmission Insulator Defects 10.57647/j.mjee.2025.17412 EN Haider Abdulzahra Saad Alsaide Department of Computer Engineering, Institute of Artificial Intelligence and Social and Advanced Technologies, Isf.C., Islamic Azad University, Isfahan, Iran https://orcid.org/0000-0001-5112-3885 Mohammadreza Soltanaghaei Department of Computer Engineering, Institute of Artificial Intelligence and Social and Advanced Technologies, Isf.C., Islamic Azad University, Isfahan, Iran Wael Hussein Zayer Al-Lami Electronic Department, Amara Technical Institute, Southern Technical University, Missan, Iraq https://orcid.org/0000-0002-3796-0393 Razieh Asgarnezhad Department of Computer Engineering, Institute of Artificial Intelligence and Social and Advanced Technologies, Isf.C., Islamic Azad University, Isfahan, Iran Journal Article 2026 07 30 In the power industry, accurate and real-time defect detection in power transmission lines is crucial to reducing accidents and maintenance costs. Although various image processing and AI-based methods have been proposed using unmanned arial vehicles (UAV) images, their efficiency significantly drops under complex environmental conditions, such as high color or texture diversity in backgrounds. To address this challenge, this paper presents an improved version of the mask region-based convolutional neural network (R-CNN) network by redefining the architecture of the head and classification sections to enhance detection accuracy in complex scenarios. The proposed model introduces a novel rhombus-structured fully connected layer in the classification branch for better feature encoding and decoding, alongside a drop-out layer in the bounding box detection branch to prevent overfitting. Additionally, ResNet-101 is employed as the backbone, and transfer learning is used to optimize training and reduce computational complexity. Experimental evaluations demonstrate that the improved Mask R-CNN achieves a mean average precision (mAP) of 92.07% and an accuracy of 92.50% in detecting power transmission line defects, outperforming existing methods. Furthermore, the proposed approach helps lower the cost and time required for fault detection and maintenance, making it a practical solution for real-world power transmission system inspections.