Research On Camouflaged Object Detection Based On Guide-Learning And Scale-Aware Augmentation

Camouflaged object detection based on deep learning is an emerging computer vision task that aims to detect camouflaged objects with a high similarity to their environment,and therefore has important applications and development prospects in areas such as intelligent transportation,medical diagnosis,military reconnaissance and rescue search.However,the low differentiation of features such as colour,texture and edge between camouflaged objects and their environment,and the large scalevariation of camouflaged objects,bring huge challenges to existing research on camouflaged object detection.This paper proposes a detection network based on guide-learning and a detection network based on scale-aware augmentation,starting from enhancing the feature distinction between camouflaged objects and the background and building the scale perception function of the network,the former precisely guiding the refinement of the network from a global perspective and the latter giving a network the ability to perceive large scale-variation of objects.The main research of this paper is as follows;(1)The high similarity between the features of the camouflaged object and the environment makes it difficult for existing methods to accurately segment objects from the cluttered background.To address the problem,this work analyzes that the weak semantic association within the camouflaged objects is one of the reasons for the high similarity between objects and their environment,and uses this as a starting point to propose a multi-scale camouflaged object detection network GLNet based on guide-learning.The network first constructs a guide learning module through the proposed guide learning strategy to collect high-level semantic features to generate a high-quality guide map,which strengthens the weak semantic associations within the camouflaged objects and guides the network to refine the prediction results on a global level.A multi-scale feature enhancement module is then designed which provides a stable and effective feature for multi-scale detection in a cascade format.Finally,a parallel attention mechanism is constructed,which is deployed in the guide learning module and the multi-scale feature enhancement module,and works to model long semantic dependencies within the camouflage object in terms of space and channels,thus deepening the feature contrast between the camouflage object and the environment.Experiments show that the proposed GLNet method not only offers better detection performance but also more stable noise immunity than the existing state-of-the-art camouflaged object detection methods.(2)In order to address the problem of large scale-variation of camouflaged objects,this paper proposes a cross-level interaction network based on scale-aware augmentation,CINet,which first uses parallel branches with different receptive fields to build a scale-aware augmentation module,which adaptively learns the weight ratios between the receptive field branches to make the network capable of perceiving scale-variation of objects.Then,a cross-level interaction module is proposed to facilitate the fusion of features between levels in an interactive learning approach and to enrich the contextual information of the features.A dual-branch feature decoder is designed in the decoding stage,which facilitates the strengthening of the connection between predictions at each scale.Experiments have demonstrated that the method can significantly mitigate the negative impact of large scale-variation of objects on performance,and effectively improve the recall of extreme-scale camouflaged objects.In this paper,we propose two deep learning-based camouflaged object detection networks to address two key problems in the field of camouflaged object detection,which effectively improve the detection performance.At the same time,this paper conducts anti-noise experiments and application discussions,which play a positive role in promoting the application and development of camouflaged object detection technology.