Small Target Detection In Optical Images Based On Deep Learning

Since the birth of artificial intelligence,vision-related research has been the focus of this field,and target detection as the basis and core of computer vision is one of the important topics of research.It has been widely used in people's lives,such as power monitoring and traffic warning,Autonomous driving,etc.After continuous efforts of scientific researchers,target detection algorithms are becoming more and more mature.From traditional target detection algorithms to deep learning-based target detection algorithms,and then from deep-learning-based target detection algorithms are divided into candidate region-based target detection algorithms and regression-based target detection algorithms,the target detection system in the field of vision is increasingly perfect.The target detection algorithm based on candidate regions has the problem of slow detection speed,while the target detection algorithm based on regression has the problem of low detection accuracy.From the perspective of practicability,detection accuracy and detection speed have always been the focus of research on target detection algorithms,and for the detection of small targets in complex scenarios,it has always been the difficulty of target detection algorithms.Due to the small proportion of small targets in the whole image and the difficulty of feature extraction,in the actual detection task,the problem of missed detection is a major problem in small target detection.Facing the problem of small target detection,this paper systematically summarizes the development history of target detection algorithms at home and abroad,the theoretical basis of convolutional neural networks and the latest research results of convolutional neural networks in the field of target detection.Mainly from the standpoint of solving practical problems,elaborating on the technical level of deep learning to solve the problem of missed detection of small targets in power line inspection.Based on the SSD target detection architecture,the MFPSSD(Multidirectional Feature Pyramid Single Shot Detector,MFPSSD)target detection algorithm is proposed.The main work of the article includes the following three parts:(1)As a classic deep learning target detection algorithm based on regression detection,SSD has excellent performance in detection speed and detection accuracy,but it has a high miss detection rate for small targets.In order to solve this problem,this paper introduces the multi-directional feature pyramid structure into the SSD target detection framework and proposes an MFPSSD target detection network.The MFPSSD network makes full use of the location information of the low-level convolutional layer through the structure of the multi-directional feature pyramid,merges multiple paths with high-level semantic information,generates a more robust feature map,and is more sensitive to multi-scale target detection.Through the verification test of the high-voltage tower defect data set and the DOTA public data set,the performance of the MFPSSD detection algorithm is more excellent,which solves the difficulty of small target detection to a certain extent.(2)This paper proposes the default frame setting of introducing K-means algorithm into SSD algorithm.First,the K-means algorithm is used to perform cluster analysis on the training set to obtain the default frame ratio that matches the actual target features,and then set it.Through the verification in the test set,the algorithm can effectively improve the target detection speed.(3)The MFPSSD detection framework proposed in this paper.First,the public data set is used to train the backbone network to obtain a feature extraction network with excellent feature extraction capabilities,and then the feature extraction network is embedded into the overall MFPSSD framework.Then,the overall training of the detection network is performed on the high-voltage tower aerial image data set and DOTA remote sensing data set to obtain the trained model,and compared with the current mainstream target detection algorithm.The MFPSSD target detection algorithm proposed in this paper effectively improves the detection performance.