Search

Salient object detection has achieved great improvement by using deep neural networks. Diverse challenging problems arise from this computer vision task and attract more attention from researchers. This thesis investigates the issues in specific tasks and aims at proposing effective methodologies to tackle the problems and improve detection performance. More specifically, existing network architectures may cause dilution problems in highlevel semantic information during up-sample operations in the top-down pathway of the Feature Pyramid Network (FPN). In order to overcome this limitation, we propose a novel pyramid self-attention module (PSAM) and the adoption of an independent feature-complementing strategy. In addition, a channel-wise attention module is also employed to reduce redundant features of the feature pyramid network and provide refined results. Introducing depth information in a suboptimal fusion strategy may have a negative influence on the performance of SOD. We discuss the advantages of the so-called progressive multi-scale fusion method and propose a mask-guided feature aggregation module (MGFA). We also introduce a mask-guided refinement module (MGRM) to complement the high-level semantic features and reduce the irrelevant features from multi-scale fusion, leading to an overall refinement of detection. RGB-D methods sacrifice the model size to improve the detection accuracy, which may impede the practical application of SOD problems. To tackle this dilemma, we propose a dynamic distillation method along with simple noise elimination. To this end, the final model can significantly reduce the computational burden while maintaining the validity and mitigating the impact of distorted training data caused by low-quality depth maps. Furthermore, in cosaliency detection, we propose a novel adaptive intra-group aggregation (AIGA) method to model the relationship between individual feature representation in a single image and group feature representation. This proposed AIGA can effectively improve the performance without increasing extra network parameters. Open Access

Polyp segmentation is a crucial step towards computer-aided diagnosis of colorectal cancer. However, most of the polyp segmentation methods require pixel-wise annotated datasets. Annotated datasets are tedious and time-consuming to produce, especially for physicians who must dedicate their time to their patients. We tackle this issue by proposing a novel framework that can be trained using only weakly annotated images along with exploiting unlabeled images. To this end, we propose three ideas to address this problem, more specifically our contributions are: 1) a novel sparse foreground loss that suppresses false positives and improves weakly-supervised training, 2) a batch-wise weighted consistency loss utilizing predicted segmentation maps from identical networks trained using different initialization during semi-supervised training, 3) a deformable transformer encoder neck for feature enhancement by fusing information across levels and flexible spatial locations. Extensive experimental results demonstrate the merits of our ideas on five challenging datasets outperforming some state-of-the-art fully supervised models. Also, our framework can be utilized to fine-tune models trained on natural image segmentation datasets drastically improving their performance for polyp segmentation and impressively demonstrating superior performance to fully supervised fine-tuning.

RGB-D salient object detection (SOD) demonstrates its superiority in detecting in complex environments due to the additional depth information introduced in the data. Inevitably, an independent stream is introduced to extract features from depth images, leading to extra computation and parameters. This methodology sacrifices the model size to improve the detection accuracy which may impede the practical application of SOD problems. To tackle this dilemma, we propose a dynamic knowledge distillation (DKD) method, along with a lightweight structure, which significantly reduces the computational burden while maintaining validity. This method considers the factors of both teacher and student performance within the training stage and dynamically assigns the distillation weight instead of applying a fixed weight on the student model. We also investigate the issue of RGB-D early fusion strategy in distillation and propose a simple noise elimination method to mitigate the impact of distorted training data caused by low quality depth maps. Extensive experiments are conducted on five public datasets to demonstrate that our method can achieve competitive performance with a fast inference speed (136FPS) compared to 12 prior methods.