Research On Multi-scale Salient Object Detection Algorithm Based On Convolutional Neural Network

Salient object detection is a popular direction in the field of computer vision,which aims to identify the most noticeable parts of an image.After years of development,multi-scale feature extraction and fusion have become the basic issues of salient object detection.It has been used for detecting,segmenting,and classifying objects of various scales.In the task of salient object detection in a single modality(RGB SOD),there exist problems such as the inability to effectively extract multi-scale features from images and the failure to fully utilize global information to guide low-level features to focus on target areas for feature representation and learning using convolutional neural networks.With the introduction of depth images,the problem of how to handle cross-modal information arises.In processing multiple modal clues,i.e.,cross-modal salient object detection(RGB-D SOD),the optimal feature aggregation and multimodal learning strategies become a challenging problem.To address the aforementioned problems,this paper proposes two salient detection algorithms based on fully convolutional neural networks,which more effectively capture valid semantic information in images and highlight foreground target areas,thereby effectively extracting and fusing multi-level,multi-scale features in both single and multiple modalities.The main research work of this paper is as follows:A new globally-guided multi-scale feature convolutional neural network is proposed for salient object detection,which combines global information directly with multi-level local features.This model can better utilize global information,thus improving detection performance.Based on commonly used evaluation metrics,algorithm performance comparison and analysis are conducted on four datasets to demonstrate the effectiveness and robustness of the proposed algorithm.A new interleaved hierarchical fusion network is proposed for RGB-D salient detection,which can be used for RGBD-SOD processing in complex scenes.The information processing methods and fusion methods are different between different scales and modalities,which can effectively suppress noise in student features.The effectiveness of the model is validated by comparing it with 14 state-of-the-art methods on 6 publicly available benchmark datasets.