Small Object Detection Through CNN Based Neural Network

Small-Object detection,one of the leading research directions in computer vision,widely used in intelligent surveillance,robotics,autonomous vehicles,and military applications.Although object detection has a decades-long history of research,its performance has remained unsatisfactory concerning complex scenes,especially if a scene has severe deformation,large areas of occlusion,small objects or overlapping of objects.It is a challenging task in computer vision due to its limited resolution and information.In order to solve this problem,the majority of existing methods sacrifice speed for improving accuracy.In recent years,research on Convolution neural networks(CNNs)has developed rapidly.CNN's have achieved excellent performance in object tracking,speech recognition,and natural language processing.Because of their strong feature extraction ability,CNNs have been utilized as the basis of a substantial number of object detection algorithms.Currently,there are two major object detection algorithms.The first is based on region proposal networks.Such algorithms are represented by R-CNN,FAST R-CNN,and Faster R-CNN.The proposed region is the region where target objects may exist are selected and then feature extraction and object classification are performed on these proposed regions.Such algorithm suffers from enormous computational complexity and poor real-time performance.The second is regression-based object detection algorithm,represent by YOLO and SSD.This type of algorithm can directly return the position information and target category information at multiple positions on the image and the detection speed is fast,but the accuracy is not as good as an object detection algorithm based on proposed region.Both of these algorithms have good accuracy in medium and large object detection,but they are not satisfactory in the accuracy of small object detection.Small object detection is a hot research in the field of computer vision and artificial intellegnece.it is quite challenging task due to their limited resolution and information in the image.Small object detection is even more critical in practical applications.In self-driving,identifying and detecting distant vehicles is necessary for safe driving.In high-resolution images,pedestrians and traffic signs account for only a small percentage of the entire image.In satellite remote sensing,images,buildings,trees,cars,and other objects have few pixels and are almost visible.Many existing methods is used which is discussed in the above to solve this problem but these method have a poor accuracy and speed for the small object.The goal of the my research is to make a system that can determine the objects with high accuracy using alternative feature extraction and varying the other critical parameters such as image size.I split up the images of DOTA data set and applied different annotation technique and augmentation technique on the images which is Horizontal bounding Box and Oriented Bounding Box which is the backbone of my work,after that used the different state of art models Fast-RCNN,YOLO,SSD and Faster RCNN on DOTA dataset for getting the better result and did comparison on the basis of accuracy & detection,i.e.map(mean average precision)and speed FPS(frame per second).