AMFLW-YOLO: A Lightweight Network for Remote Sensing Image Detection Based on Attention Mechanism and Multiscale Feature Fusion

• Published in: IEEE transactions on geoscience and remote sensing 2023, Vol.61, p.1-16
• Main Authors: Peng, Guili, Yang, Zijian, Wang, Shoubin, Zhou, Yuan
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Attention mechanism; Computation; Computer networks; Convolution; Convolutional neural networks; Deep learning; Detection; Feature extraction; feature fusion; Image detection; Lightweight; lightweight network; Machine learning; Mathematical models; Model accuracy; Object detection; Object recognition; Optical sensors; Parameters; Remote sensing; remote sensing image; Sensors; Task analysis
• ISSN: 0196-2892; 1558-0644
• DOI: 10.1109/TGRS.2023.3327285

The scale of targets in remote sensing images varies greatly and is diverse. It has many small targets that are distributed densely and high complexity of image background. The number of network model parameters and the computation amount of the object detection algorithms based on deep learning is huge. It is difficult to apply them on a platform with fixed performance and limited computing resources. A lightweight remote sensing object detection model is proposed in this article, which called attention and multiscale feature fusion lightweight-YOLO (AMFLW-YOLO). The deep separable convolution, inverted residual, and linear bottleneck structure are employed to replace the standard convolution layer to reduce the model parameters in the backbone network of the model. The coordinate attention (CA) mechanism is introduced into the feature fusion network to capture the direction- and location-aware information across channels at the same time, which improves the accuracy of the network. The bidirectional feature pyramid network (BiFPN) structure is employed to strengthen feature extraction. The learnable weights are introduced to learn the importance of different input features. The multiscale feature fusion is applied to improve the detection effect. The experimental results show that the algorithm achieves satisfactory performance in terms of efficiency and accuracy and has advantages in detection accuracy and model lightweight.