| Automatic target detection based on deep learning is one of the hot research fields of remote sensing image interpretation.Because deep networks are usually large in scale and have many parameters,overfitting tends to occur when the number of training samples is small;even if the training data is sufficient,the existing deep learning methods still have generalization capabilities in the face of open and complex environments.The problem is poor,so in many practical applications,artificial target detection is still the main method of remote sensing image interpretation.Humans have a strong learning ability,only a few samples can master how to identify a certain target,and have a strong generalization ability for environmental changes,but traditional manual interpretation methods are difficult to meet the target detection of large amounts of data and strong real-time Task requirements.Brain-Computer Interface(BCI)technology realizes the communication between internal information of the brain and the external environment by interpreting the electroencephalography(EEG)signals of the human cerebral cortex,and provides a new technology for improving the efficiency of human interpretation way.The performance of the BCI-based target detection method mainly depends on the two factors of human target reading ability and the performance of the EEG signal classification method.The enhanced training of human visual ability can effectively improve the efficiency and robustness of its target detection.Compared with the deep neural network model,the artificial detection model has its own advantages and disadvantages in learning ability,generalization ability,and detection efficiency.In the target detection field,two target detection models have difficulty in training,learning ability,generalization ability,and detection efficiency.And the lack of detailed quantitative analysis of the model stability.Aiming at the above problems,this paper designs an artificial target detection experiment and a deep network-based target detection experiment,and conducts a detailed comparative analysis of the experimental results,providing a basis for the training of the target detection model and the study of related issues.The main research results of this article are summarized as follows:(1)This article selected 10 healthy,normal-vision,right-handed students from Xidian University of Science and Technology to participate in the experiment,with an average age of 23 ± 2.7 years.They have never conducted target detection experiments and related field research,and have not been exposed to EEG-related issues.experiment.The VEDIA database containing 2540 remote sensing images is selected as the experimental data set,which contains 1270 natural light images and infrared images each.120 images are selected as test sets to test the learning ability and generalization ability of the model,and the remaining 1150 natural light images are selected.As training set.(2)This paper selects Faster R-CNN as the deep learning target detection experimental network,and controls the number of training samples to quantify and analyze the target detection performance of the deep learning network.The experimental results show that the deep network has a very high model training efficiency,and it only takes 1 hour,15 minutes,and 21 seconds to train the model on the full sample data.The learning ability of the deep network increases with the number of training samples.When the ratio is less than 26% of the full sample,its target detection performance is poor,and when the number of training samples reaches 43% of the full sample,its target detection performance starts to saturate,and the amplitude of the improvement is significantly reduced;the generalization ability of the deep network It is more dependent on the number of training samples.When the number of training samples reaches 61% of the full sample,its target detection performance in the infrared scene surpasses the random classifier,and gradually enters saturation,and the generalization ability is poor.(3)This paper carried out a hand-labeled target detection experiment based on large-format images.The training experiment lasts for two days: 5 natural light images are selected as training data,and experimental training and self-training are performed once a day;human fatigue will cause its performance to be significantly reduced under large data tasks,so the test data will be equivalent.It was randomly divided into five parts.In the test experiment,the natural light scene and the infrared scene experiment were performed in the morning and afternoon each day.The experimental results show that the human sample has a strong learning ability in this mode.The average accuracy and recall rate of target detection by 10 subjects in the test experiment finally reached 93.92% and 91.04% respectively;while the performance of human target detection in the infrared scene is relatively The drop in natural light scenes is similar to that of deep networks,but because of its higher target detection accuracy in natural light scenes,the average accuracy and recall of target detection in infrared scenes is 79.25% and 81.26%;in this way The subject’s target detection efficiency increased steadily as the experiment progressed.The fastest large-format image detection in the experiment reached 12671 ms,which was far lower than the deep network.In addition,the subjects showed human’s unique self-learning ability in experiments,and their target detection performance continued to improve with the progress of the test experiments.(4)This article carried out a BCI target detection experiment based on image slices.Based on the model of BCI combined with Rapid Serial Visual Presentation(RSVP)experimental paradigm,the specific EEG generated by the subjects during the image recognition process was interpreted.Using support vector machine(SVM)to build a subject-specific target detection calculation model,to achieve rapid interpretation of image information.The experiments are divided into offline experiments and online experiments,which verify the performance of offline target detection based on BCI and the performance of online realtime target detection,respectively.This paper verifies through two-stage training experiments: the combination of key and non-key is a more effective EEG training mode,and the best training ratio of the classifier is the training set and test set 4: 1.Due to the difficulty of EEG training,this article expands the training data to 20 natural light scene maps.After cutting,980 small-format stimulus slices are obtained.After training,8 subjects have the potential to induce N170 on the target slice;test experimental time settings Similar to the manual labeling experiment,the participants showed similar self-learning ability to the manual labeling experiment,and their target detection performance showed continuous improvement as the experiment progressed.In the offline experiment,the learning ability of the subjects was lower than that of the manual labeling experiment,and the target of the relatively deep network model had a strong ability to recall the target.However,because of the difficulty of identifying the interference source under the RSVP sequence,the accuracy of the target capture The performance of the target in the infrared scene in this mode is the lowest compared to the natural light scene in the three kinds of experiments,and it has the best generalization ability.In online experiments,in this paper,the target is detected in real time,but The pre-processing of the data in this mode is relatively simple,and the lack of steps such as complete removal of ophthalmology and electromyography,de-baseline drift,and full data re-reference makes the target detection accuracy lower than that of offline experiments.This paper verifies human’s fast visual target detection capabilities.Experiments show that as the slice presentation frequency accelerates,the subject’s target detection performance will decrease slightly.Researchers can sacrifice some detection accuracy to improve target detection efficiency according to actual needs.(5)This paper compares and analyzes the results of three kinds of target detection experiments based on deep networks,manual tagging,and BCI technology.Through the parameters such as model training time and difficulty,target detection accuracy and recall,and target detection efficiency,Detailed analysis and comparison of different practical application scenarios provide intuitive quantitative comparative analysis data.The research results provide a basis for the research and application of model training on related issues. |
PDF Full Text Request