目标检测与分割是计算机视觉的核心问题之一，但在真实世界复杂应用场景下，尤其是当不同目标类别之间存在较大尺度差异，或者由于成像距离导致目标出现大尺度变化的情况下，现有方法的性能还难以满足实际应用的需求。本项目针对大尺度变化目标的检测与分割问题，从方法框架、特征提取、假阳性目标滤除等角度进行系统的研究。研究图像局部区域自适应缩放机制，构建适用于大尺度变化目标检测与分割任务的新型深度神经网络训练与推理框架；设计既能突出局部物体又能整合全局信息的轻量级特征提取方法，提高网络模型效率；研究高效的可端对端学习的假阳性目标滤除方法，提高网络模型定位精度；最终实现一套高效准确的大尺度变化目标检测与分割系统，为智能目标自动定位与识别技术在复杂场景下的应用提供理论和方法支持。

Object detection and segmentation are the key problems in the field of computer vision. However, the current state-of-the-art methods cannot fulfill the requirements of complex real-world applications, especially when objects exist extreme scale variations. The extreme scale variation problem usually arises when there is large scale discrepancy between different object categories or the same kind of objects are imaged at different distances. This project aims to investigate the challenging problem of extreme scale variation object detection and segmentation from the aspects of algorithm framework, feature extraction and false positive suppression. We intend to cope with the extreme scale variation problem first by designing an adaptive image region zooming mechanism and develop the corresponding deep neural network training and inference framework. In order to improve the network efficiency, a lightweight feature extraction network model is proposed which can reveal both the local and global image features. We will further study the end to end learning method for false positive suppression to improve the localization precision. Finally, a prototype system will be developed for algorithm evaluation and validation. Our study hopes to provide deep insights and forward-looking ideas for the application of intelligent object detection and recognition techniques in complex real-world environments.