A Framework of Detection, Counting, Classification, and Advanced Processing Techniques for Objects in Multi-Resolution Imagery

Abstract

Detecting, counting, and classifying objects represent the most primary and challenging tasks in the field of computer vision. In the signal, image, and video processing domain, it is crucial to improve detection performance and apply evaluation metrics when evaluating algorithms and post-processing schemes. In this Ph.D. Dissertation, we have conducted research studies on single-class vehicle detection, counting, and classification, as well as multi-class object detection, classification, and recognition using aerial datasets of variable resolutions. For single-class vehicle detection and classification, major contributions are categorized in three manifolds: i) adapted a variety of object detection and segmentation algorithms and five existing post-processing schemes, quantified and compared their performance to analyze their advantages and shortcomings via two sets of evaluation metrics; ii) derived three post-processing schemes on object detection and classification using low-resolution wide-area aerial datasets, conducted quantitative analysis based on two sets of classical metrics and some enhanced measures; iii) applied the multi-stage learning based methods for large-scale datasets where the test images are of variable resolutions and oriented for different topics. For multi-class object detection, classification, and recognition, we have conducted research studies in the following aspects: i) performed multi-class detection on small objects by designing optimal post-processing module; ii) applied two-stage learning and updated learning schemes for aerial objects; and iii) supplemented machine learning based quantitative results and ablation study on some related research topics. Experimental results have proved the validity and efficiency in these scenarios: a) improved average F-score to be more than 0.8 and achieved average reduction of 83.8% in false positives for five object detection and segmentation methods; b) applied FC-DenseNet-103 model for two-stage machine learning in the online vehicle detection in aerial imagery (VEDAI) dataset, where an average of 0.855 for initial object detection accuracy was achieved via Google Colab Pro using adjusted learning rates and some other optimized parameters for both training and validation; optimal post-processing results were achieved by the proposed MTAP scheme (0.891 on object detection accuracy and 82.8% on mean average precision (mAP)), which may progressively improve the performance of multi-class object detection and classification; and c) achieved relatively better performance for vehicle license plate recognition (VLPR) when comparing to a few classical approaches.