Based On Multi Supervision Festure Integration Of Salience Object Detection

Salient object detection of images refers to extracting salient information from images,and processing salient information of images purposefully to reduce the time and complexity of image processing.It is an important research direction in the field of computer vision.Deep learning technology has been gradually applied to image saliency detection.With the help of convolution depth neural network,great progress has been made in image saliency detection.However,the current saliency detection results still can't meet the application requirements.In order to improve the detection accuracy.Image saliency detection model is constructed from the perspective of multiscale feature extraction.The main work includes:Firstly,a new layer multi-scale integration method is proposed based on the iterative aggregation of layers.Aggregation begins at the smallest scale and then iteratively merges larger scales.Iterations are carried out successively through highlevel guidance of low-level,and the scales of the starting positions of iterations are increased successively,multiple iterations are carried out in shallow layers,which effectively extracts the feature information of layers and improves the detection of edges and positions of saliency regions.In view of the previous saliency detection models are supervised at the end of the network,supervision is used in the middle layer to improve the detection accuracy,constantly revises the internal parameters to improve the detection accuracy through correct information guidance.Secondly,based on the fusion method of the loss function of the saliency detection model of skip-layer connections,another iterative integration architecture is proposed,which also starts from the small scale,but makes the scale of the starting position of the iteration unchanged,and the depth of the iteration decreases in turn,By increasing the scale of the beginning of the iteration,the above iteration aggregation is repeated at a relatively high-level.The detection accuracy is improved by deep feature extraction and deep supervision.Finally,the experiments are done on MASR,DUT-OMRON,HKU-IS,ECSSD,the results are compared with those traditional methods,it is shown that the detection accuracy of this method is improved by 1 percentage only through 200 iterations.