Research On Salient Object Detection Algorithm Based On Multi-layer Feature Fusion

Salient object detection is to extract the most attractive objects from images,which is often used as an important preprocessing or intermediate step of various visual tasks,is an important research topic in the field of computer vision.At present,deep convolution network-based salient object detection(SOD)has achieved impressive performance.However,it is still a challenging problem to make full use of the multi-scale information of the extracted features and which appropriate feature fusion method is adopted to process feature mapping.This thesis mainly studies the SOD algorithm based on multiple feature fusion.The main research work and achievements are as follows:(1)Aiming at the blurred detection results of image target region caused by only single-layer features in most current detection algorithms based on full convolution neural network,we propose a new adjacency auxiliary network(AANet)based on multi-scale feature fusion for SOD.Firstly,we design the parallel connection feature enhancement module(PFEM)for each layer of feature extraction,which improves the feature density by connecting different dilated convolution branches in parallel,and add channel attention flow to fully extract the context information of features.Then the adjacent layer features with close degree of abstraction but different characteristic properties are fused through the adjacent auxiliary module(AAM)to eliminate the ambiguity and noise of the features.Besides,in order to refine the features effectively to get more accurate object boundaries,we design adjacency decoder(AAM＿D)based on adjacency auxiliary module(AAM),which concatenates the features of adjacent layers,extracts their spatial attention,and then combines them with the output of AAM.The outputs of AAM＿D features with semantic information and spatial detail obtained from each feature are used as salient prediction maps for multi-level feature joint supervising.Experiment results on six benchmark SOD datasets demonstrate that the proposed method outperforms similar previous methods.(2)In the third chapter,better detection results have been achieved through the fusion network.However,only the single stream framework is used to extract and fuse features,and the information of images is relatively single.In order to fully explore and make full use of the information in images to detect saliency targets,we propose an end-to-end dual-stream co-dec network(DEDNet)for salient object detection.Firstly,the original image(rgb)stream and the processed coarse saliency(C-sal)stream are input into the encoder to obtain multi-layer features.In the process of decoding the model,the C-sal stream can guide the feature extraction of the model and help to locate quickly;On the other hand,the information in the rgb stream can also supplement the coarse saliency stream.In order to effectively capture the main salient areas and suppress the background noise,we design the global attention guide module(GAGM)to enhance the global features.In addition,when considering the interaction of two streams,we design the information interaction guide module(IIGM)to solve the problem of interaction and fusion of the two streams.When they all have effective information,they combine with supplementary information,add more reliable relevant pixels,suppress deviations and make consistent predictions;when storing information deviations,they synthesize useful information to make correct choices.In addition,the decoding part is formed by concatenating multiple feature fusion decoding modules(IFM),which fuses the features of the corresponding coding layer,the upper decoding layer,the features after interaction with another stream,and the global information.Finally,the features after the fusion of the two streams are used to output the final prediction maps.On six commonly used benchmark data sets,we compare our proposed DEDNet method with the other 13 advanced methods on F_βand MAE evaluation indicators to prove the effectiveness of the network.