Research On Salient Object Detection Method Based On Deep Learning

Salient object detection aims to detect and extract the most salient objects in an image.Existing salient object detection algorithms can lead to partial loss of image information during feature extraction at coding time.To address the problem of image information loss due to multiple convolutions during feature extraction,this paper proposes an IENet,a salient object detection network based on input enhancement,and applies the proposed IENet to other computer vision tasks.For the proposed IENet model,an Unfold operation is designed in the input enhancement part to achieve the operation of reducing the image size while maintaining all pixel information of the image.The image is then convolved and summed with the feature map output from the upper decoder as a way to compensate for each other’s information.In the feature fusion stage,a feature map blending strategy is designed to rescale and stitch the feature maps from different levels of encoders,and SEASPP is proposed to allocate attention to the stitched images to enhance the fusion effect.Regarding the design of the loss function,binary cross-entropy loss and structural similarity loss are used to enhance the network’s ability to handle salient object boundaries.In the experiments,the proposed algorithm is compared with different algorithms on five datasets and eight metrics respectively,and the results show that the proposed algorithm performs excellently in salient object detection.For other computer vision tasks,this paper applies the salient object detection method to skin tumour classification based on the proposed IENet,and designs a web-based skin tumour classification app.i ENet is mainly used in the image pre-processing part to highlight the lesion region and fade out the effect of the background on the classifier,which uses Dense Net121.In the lesion extraction phase,IENet was compared visually with Ground Truth and numerically with other model algorithms to show that the overall performance of IENet for lesion extraction was good.The effect of using different image processing methods on the classifier’s effectiveness was then tested in the classification phase,and the results showed that the pre-processed images using IENet were more beneficial in improving the classifier’s accuracy.