Research On High Performance And Lightweight Saliency Detection Algorithms

The goal of the saliency detection is to simulate the human eye’s attention mechanism and discover salient areas and objects in the picture.Saliency detection plays an important role in many practical applications.Although saliency detection models based on deep learning methods have made significant progress in recent years,there are still many problems worthy of research.On the one hand,how to improve boundary quality and effectively suppress the background noise is the core issue to improve the performance.On the other hand,most of the existing methods focus on the detection performance,while ignoring the computing operations and storage overhead of the model.Therefore,how to achieve a balance between model performance and overhead is also an issue worthy of research.For the first question,this paper proposes a saliency detection method combined with an edge detection network.Firstly,the features of the pre-trained edge detection model are integrated into the decoder of the saliency detection model to enrich the edge features.Secondly,the low-resolution output images of the edge detection model and the saliency detection model are input into a super-resolution optimization network to further optimize the result while increasing the output resolution at the same time.Lastly,this article starts from the perspective of supervision information,combines multi-layer supervision mechanism and attentionenhanced multi-scale pooling method to enrich the global context information of the network.For the second question,this article proposes a lightweight saliency detection model based on lightweight modules and model compression techniques.First,a lightweight backbone and decoder structure are proposed.On this basis,two model compression routes are proposed.One is to use a pre-trained base network for model sparse training and pruning.First,this paper proposes a dynamically decreasing sparse rate setting to solve the model degradation phenomenon.After the sparse training,this paper proposes a greedy pruning strategy combined with the mean change point method,which can achieve lightweight demands through gradual pruning while ensuring performance.In the second method,this paper first uses a randomly initialized model on the target task for sparse training and pruning to obtain an optimal lightweight model that is more in line with the target task.After that,the parameters are reinitialized,and the small network is retrained using the method of model distillation.The models proposed by the above two methods have been tested on multiple public test sets,and compared with the existing advanced algorithms quantitatively and qualitatively.The experimental results show that the algorithms in this paper have achieved good results.