Lightweight Small Target Key Part Object Detection For Remote Sensing Images

In object detection applications for remote sensing images,there is a significant research value in object detection for key parts of important objects,and in the military,key part object detection can be used to discover the specific parts to be struck for precise strikes.After testing,many missed detections occur with existing object detection algorithms for key parts of small target objects,and the reason for this is that the detection of key parts of small objects is more challenging.In order to solve the problem of poor accuracy of key part object detection by existing techniques,this thesis focuses on neural network lightweight technology,key part object detection technology,neural architecture search technology,and small object detection technology.The main work and innovations of this thesis are.（1）To address the problem of insufficient accuracy of key part object detection,a two-stage key part object detection SKP-YOLO is proposed.it consists of two stages,the first stage is ST-YOLO for small object detection,and the second stage is KP-YOLO for key part object detection.by splitting the key part object detection into two stages,the accuracy of key part object detection is improved compared with the common object detection algorithm.By splitting the detection of key parts into two stages,the accuracy of key part object detection is improved compared to the generic object detection algorithm.（2）To address the problem of network lightweighting,a neural architecture search algorithm SKPNAS is proposed for object detection.the neural architecture search algorithm adopts a supernet training method with strict fairness characteristics by not putting back sampling strategy and multiple training with one parameter update to improve the average training accuracy of the supernet.After the training of the hypernetwork of SKPNAS is completed,the search strategy of genetic algorithm is used to efficiently search for superior network structures in the search space compared to the random search algorithm.（3）To address the problem of missing detection of key parts of small object,this problem can be improved by splitting SKP-YOLO into two stages of key part object detection,and to further improve the key part object detection accuracy of small objects,two feature fusion modules,feature pyramid network and path aggregation network,are also used to improve the small object detection capability and multi-scale object detection capability of the object detection network.After testing,the two-stage key part object detection SKP-YOLO proposed in this thesis for visible remote sensing image dataset has significantly improved the key part object detection accuracy with a speed difference of only 8.5 fps compared with the traditional object detection method.Compared with the original network structure,the Mobile Net V2-NAS-4 for the visible remote sensing image dataset searched by the neural architecture search algorithm SKPNAS for object detection proposed in this thesis reduces MACC by 3.82 M,but improves the accuracy by 0.17 %.