Detection And Recognition Methods For Movable Ground Targets With Few Samples

Object detection and recognition technology has always been the key research direction in the field of computer vision.The main applications in the civilian field include automatic driving,face recognition and intelligent transportation,etc.In the military field,its main applications are reflected in photoelectric reconnaissance,surveillance,warning and precision guidance.This paper mainly studies the detection and identification of movable ground targets under complex backgrounds,focusing on time-sensitive targets such as tanks,armored vehicles and special transport vehicles that may change their positions and attitudes.This kind of target has the characteristics of small but many targets,diverse target types,complex background and scarce samples.In order to solve these problems,this paper proposes the use of visual saliency theory and deep learning method to study the detection and recognition technology for movable ground targets under the condition of scarce samples.The main work and achievements of this paper are as follows:1.Aiming at the problem of small target detection under complex background,we proposed a new high-quality dataset that can reflect the characteristics of reconnaissance and surveillance applications,with less center bias,balanced distribution of salient object size ratio,diverse image resolution,and scenes with multiple objects.It includes 300 color images containing multiple targets and multiple scales which are from the reconnaissance and surveillance scenario.We quantitatively evaluated 12 representative visual saliency models on this new dataset and analyzed the performance of different algorithms on small target detection tasks.Based on our analysis,we proposed a new salient object detection method that combines Boolean map and grayscale rarity.Our method can effectively solve the problem of multiple target and small target detection in complex background.In addition,our method is computationally efficient on images of various sizes.2.Aiming at the problem of object recognition under the condition of scarce samples,we proposed a method that utilizes simulated samples to train the deep CNN models and then recognize the real samples.Firstly,a set of dataset establishment methods are proposed,including data simulation based on 3D model,image acquisition based on Web Crawler,automatic annotation based on saliency map and data enhancement based on reserved annotation.Secondly,the simulation samples were used as the training set to retrain the Faster R-CNN.We tested the retrained model on the real samples of four kinds of movable ground targets,and the recognition accuracy(m AP)could reach 70%.3.According to the characteristics of infrared data,we pre-processed the original infrared data to obtain the grayscale-enhanced image and pseudo-color image.Then the Faster R-CNN model is trained and tested.The recognition accuracy(m AP value)is81%.In addition,according to the size characteristics of the infrared target,the modification of the anchor parameters of the RPN network is proposed and verified by experiments.