基于可见光遥感图像的目标检测是战场监测、应急抗灾等军民领域的核心技术。当前，遥感图像已迈进“大数据”时代，高分辨率、快速更新的海量遥感图像为目标检测带来严峻挑战：一方面，由于在轨存储资源和星-地传输带宽的制约，大部分遥感图像以压缩形式存储和传输，必须解压缩才能进行后续处理，影响目标检测的实用性和实时性；另一方面，由于星上处理能力有限，现有的目标检测方法难以平衡其准确性和复杂度。为解决上述问题，本文改进传统遥感图像压缩和目标检测相互分离的方式，提出了压缩域特征提取与目标检测方法，即挖掘遥感图像压缩数据特性，直接在少量解压或不解压的压缩数据上进行特征提取与目标检测，缩短目标获取时间，提高检测效率。本文的主要工作及创新点如下： (1)针对熵解码和反量化后的压缩数据，提出了基于JPEG2000小波域的机场检测方法。本方法利用小波低频系数进行区域分割和局部特征点匹配，并提取大尺度纹理、结构特征，粗略判定机场候选区域；利用小波高频系数提取候选区域多尺度、多方向的直线结构特征，精确定位机场位置。与传统目标检测方法相比，本方法可达到较好的检测率和虚警率，运算速度提高60%。 (2)针对完全不解压的原始压缩码流，提出了基于JPEG2000码流的特征提取方法。本方法自动获取JPEG2000码流中的包头和包体信息，利用包头信息直接提取图像的强度、纹理等特征；通过推演JPEG2000算术编码，将包体信息重构为小波梯度强度图，并通过小波梯度强度图提取精细、有效的特征。综合利用上述码流特征，构建了机场和海面舰船的快速检测框架。相比于传统的图像域方法，本方法检测结果准确且检测速度显著提升。 (3)为提高压缩域特征的表征能力，提出了一种基于条件随机特征映射的压缩域深度表征方法。本方法改进了典型的超限学习机，并将其用于深度特征表征。为解决超限学习机表征能力不足的问题，采用有效的标签对齐度量函数，训练得到条件随机特征映射，并且扩展为深层表征框架。相较于现有深度表征方法，本方法复杂度更低，且具有良好的鲁棒性和泛化性。 (4)为改进现行的遥感图像压缩处理性能，进一步减少压缩数据量，提出了一种生成式图像压缩方法，并基于本方法获取压缩域特征。本方法采用生成式对抗网络训练得到编/解码器，训练过程利用对抗损失函数对局部像素重构误差进行优化、对全图内容进行恢复，通过剔除更多视觉冗余信息来获得更高压缩比。并且，生成式压缩数据可直接作为特征来辨别遥感图像目标。在高压缩比条件下，本方法比现行压缩方法的质量更好，且压缩域特征具有良好的判别能力。 关键词：遥感图像；压缩域处理；特征提取；目标检测；深度特征表征；生成式压缩

 Efficient object detection from optical remote sensing images (RSIs) is one of the key technology in the military and civil fields, e.g., battlefield monitoring, disaster resistance and so forth. Nowadays, with the increasing spatial resolution and updating velocity of optical images (OIs), the RSI processing systems have been into the "Big Data" era, offering serious challenges for the traditional object detection on the OIs. On the one hand, due to the limitation of on-orbit storage and down-link bandwidth, the OIs are typically stored and transmitted in the compressed format. It is unavoidable to analyze the OIs after decompression steps, which negatively impacts on the feasibility and timeliness of RSI processing systems. On the other hand, caused by the constrained resources of satellite platforms, the existing objection detection algorithms, which are towards ground environment, is unattainable to tackle massive OIs on satellites. To address these issues, this thesis propose the compressed-domain feature extraction and object detection methods, modifying separated compression and detection in the traditional RSI procedure. The compressed-domain methods means visual features are directly extracted from partial decompressed data or raw codestream to detect objects instead of pixel values, benefitting from unique characteristics of compressed data. The main contributions of this thesis can be summarized as follows: (1)In terms of intermediate compressed data obtained by partial decoding, we propose a fast airport detection based on discrete wavelet transform (DWT) from JPEG2000. In this work, we utilize low-frequency DWT coefficients to segment regions and match local features, and then, the large-scale features are extracted for locating candidate regions coarsely. Moreover, the high-frequency DWT coefficients are applied to detect runways in multi-scale and multi-orientation to determine the airport more accurately. Experimental results verify that, compared with the traditional image-domain algorithms, the proposed method has the comparative detection results with faster speed by over 60%. (2)Regarding the fully-compressed codestreams in on-board storage, we propose feature extraction method based on JPEG2000 codestream. Our method automatically captures the header and body information in JPEG2000 codestream, and the headers are directly exploited to present texture and intensity features. Furthermore, by analyzing the arithmetic codec in JPEG2000, the Wavelet Gradient Magnitude Maps (WGMM) can be reconstructed from the bodies, so that the compact and advantageous features are designed based on WGMM. By taking advantages of the obtained codetream-domain features, we develop two rapid detection framework for airport and ship respectively. Experimental results on the real IOs in compressed format demonstrate that the proposed approaches outperform the image-domain ones with superior detection accuracy and lower latency. (3)In order to improve compressed-domain discriminative representation, a deep compressed-domain representation model based on Conditional Random Feature Mapping (CRFM) is presented. Most of the existing deep neural network (DNN) based representation models are unable to achieve balance between performance and complexity. Unlikely them, we aim to extend Extreme Learning Machine (ELM) for deep representation. To resolve representation limitations in the ELM, we utilize the label information for optimizing the CRFM by an efficient label alignment metric. Moreover, we develop the CRFM to the supervised deep representation architecture. Extensive experiments on various datasets demonstrate our model is more effective than original ELM-based and other existing deep representation models with rapid training/testing speed. (4)For purpose of reducing data volume produced by existing OIs compression methods, we proposed a generative OIs compression, and extract features from this compressed domain. The proposed method apply Generative Adversarial Network (GAN) to optimize codec. In the training process, adversarial loss function not only reconstruct each pixel value, but also preserve global image content, leading to reduce more redundant information and achieve a higher compression ratio. Moreover, the codestream from the trained codec can be directly exploited as the compressed-domain features for inference tasks. Testing results indicate that our compression approach has a significant quality improvement against the popular compression standard, and its compressed-domain features have good robustness and generalization. Key Words: Remote Sensing Image Processing, Compressed-Domain, Feature Extraction, Object Detection. Deep Representation, Generative Compression