Large-scale complex images contain a large number of crowded and fine-grained features that are prone to information loss during Deep Learning's (DL) downsampling process. While most recent contributions focused on discriminative tasks like small object detection, fewer papers tackled complex image processing for generative tasks. This paper presents a new pipeline for extracting and clustering significant sub-images from complex, unannotated image datasets. The pipeline is composed of three main steps: (1) identifying clusters of keypoints using K-means, (2) extracting and refining sub-image frames with a region-based expansion algorithm, and (3) classifying extracted frames. Subsequently, it enables data augmentation, simplifies domain complexity, and supports conditional training, tailoring the learning process to specific contexts, and enhancing the quality of image synthesis models. Furthermore, we analyze the structure of complex image domains by investigating intra-domain gap between extracted clusters.

Generative Adversarial Networks (GANs) have shown potential for generating images, but have limitations when applied to complex datasets. To address these limitations, class-conditional training is employed, as it performs better and maintains a high level of semantic diversity. In this work, we propose a new method for training generative models on complex images by extracting rules defining the relationships between objects in the image, cropping significant sub-regions based on these rules, and training the models in a conditional setting using the extracted rules as labels. The proposed approach is evaluated and the results demonstrate its effectiveness by increasing the training dataset size, and then feeding it to conditional training. As a result, synthesized samples maintain asset fine-grained details and the visibility of small instances.

Computer vision applications such as object detection and recognition, allow machines to visualize and perceive their environments. Nevertheless, these applications are guided by learning-based methods that require capturing, storing and processing large amounts of images thus rendering privacy and anonymity a major concern. In return, image obfuscation techniques (i.e., pixelating, blurring, and masking) have been developed to protect the sensitive information in images. In this paper, we propose a framework to evaluate and recommend the most robust obfuscation techniques in a specific domain of application. The proposed framework reconstructs obfuscated faces via deep learning-assisted attacks and assesses the reconstructions using structural/identity-based metrics. To evaluate and validate our approach, we conduct our experiments on a publicly available celebrity faces dataset. The obfuscation techniques considered are pixelating, blurring and masking. We evaluate the faces reconstructions against five deep learning-assisted privacy attackers. The most resilient obfuscation technique is recommended with regard to structural and identity-based metrics.