Research on Multi-Class Recognition of Landing Gear Based on an Improved ResNet-50 Approach

To address the current issue of insufficient accuracy in aircraft landing gear state recognition, a recognition model based on ResNet50 and an enhanced attention mechanism is proposed. Firstly, a proprietary image dataset encompassing three states—landing gear deployed, retracted, and partially deployed—is constructed. Sample size is expanded through multiple data augmentation techniques to enhance generalisation capability. Subsequently, the DASAM dual-aligned spatial attention module is embedded within the ResNet50 backbone network to achieve refined feature enhancement in critical structural regions. Comparative experiments against models including GoogLeNet, ResNet34, YOLOv8, ResNet50, and ResNet50-CBAM demonstrate that the proposed model achieves optimal results across metrics such as recognition accuracy, with a validation set accuracy exceeding 95%. Furthermore, this approach effectively enhances landing gear state recognition capabilities under complex viewing angles, low-light conditions, and background interference, offering valuable reference for aviation safety monitoring and intelligent perception systems.