Damage Detection Algorithm Based On Object Shape And Attention Mechanism Characteristics

In recent years,airplanes have become one of the main means of transportation.Due to the long-term flight in service,aircraft surface will be affected by various external factors and cause various types of damage.The damage poses a huge threat to flight safety and should be discovered in time.At present,the commonly used visual detection has low maintenance efficiency and large manual workload.And the automatic damage detection algorithm could improve the maintenance efficiency to a certain extent.Traditional algorithms usually use artificially designed features,but artificially designed features rely more on human experience,feature expression and generalization capabilities are relatively low.In recent years,object detection algorithms based on deep learning have received widespread attention and have been widely cited in many fields.Aiming at the problem of aircraft scar detection,this paper analyzes the reasons for the low accuracy and makes corresponding improvements.The main research contents include:Aiming at the problem of low correlation between the positioning accuracy of detected boxes and their confidence,this paper designs an IoU branch and a two-stage regression branch to predict the positioning accuracy.And the classification confidence is recalculated by predicted IoU and the probability distribution of location in the first stage.The designed branch could reduce the difficulty of regression,and make the detection frame more accurate in NMS,so as to retain the detected boxes with higher quality,and improve the positioning accuracy.Aiming at the constant gradient of loss for regression and the sample imbalance of regression,this paper designs sample balanced loss function with adaptive gradient to accelerate model convergence and enhance the positioning accuracy.And in order to improve the feature expression ability of backbone,this paper designs a multi-dimensional channel attention module to enhance the position information contained in the model feature map to improve the detection accuracy of the algorithm.The algorithm proposed in this paper was tested on the aircraft damage dataset.The experimental results show that the model designed in this paper can enhance the positioning accuracy and detection accuracy.The mAP reaches 55.02%,and improved by 1.67%compared to ATSS algorithm.