Research On Multispectral Image Salient Object Detection Based On Deep Multi-level Cascade Network

Salient object detection plays a crucial role in the task of accurately determining the location of the objects in the scene and searching the corresponding contents.Although in recent years,lots of salient object detection algorithm and research methods have been proposed,especially salient object detection algorithm based on deep learning,but the precision of the corresponding tasks and salient features of utilization rate still have space for improvement,as well as salient object detection under complicated conditions and scenes tasks still didn't get satisfactory solutions.According to the above all kinds of problem,this paper puts forward the salient object detection based on multi-level cascade network architecture with multispectral data,in order to improve the accuracy of salient object detection and image salient features of utilization,and solving complex conditions and scenarios of salient object detection results the improvement of robustness and accuracy.The main work is as follows:Fully convolutional neural network(FCNs)plays an important role in the current salient object detection task,because its multi-level structure describes the depth features of images at different scales.In order to gather and utilize the salient features of each level reasonably and effectively,a new end-to-end multi-level convolution feature cascade model is proposed.The model consists of two modules: one is the multi-level deep feature extraction module realized by the improved FCNs;The other module is multi-level feature fusion module,which effectively combines the global contour features and local fine features corresponding to multiple pooling layers into a whole through cascade,upsampling,deconvolution and other operations.Finally,the output of the fusion module is used to predict the saliency map through further learning.The model can effectively and flexibly aggregate the features of multi-level convolution layer and provide accurate saliency prediction.(2)On the basis of the previous work,the network architecture module of deep feature extraction is improved,and the dense module is combined with the previous deep learning model to make the overall model more lightweight.With fewer parameters,the deep learning model training can be converged more quickly,and the accuracy and accuracy of the significant target image results under single modal can be improved,and improve the efficiency of generating the final significance prediction image.Use multimodal data set for the double side synchronous training,under different modal will significantly predict figure images in the visible light spectrum and hot infrared spectrum images of significant prediction graph through the consistency constraint cross modal fusion to produce the final significant prediction images,make model in responding to the scene for better performance in the complex interference,this will deep learning a combination of the overall algorithm and traditional algorithm has higher robustness,in response to a real life all kinds of challenging scenarios.