Research And Implementation Of Pixel-wise Labeling Algorithm Based On Convolutional Neural Networks

Pixel-wise labeling is one important task in the field of computer vision.Image semantic segmentation and saliency detection are two important pixel-wise labeling tasks.Semantic segmentation aims to assign a categorical label to every pixel in an image,which plays an significant role in autonomous navigation,medical image processing and virtual reality.Saliency detection aims to identify and isolate the most visually distinctive objects or regions in an image,which can bring positive assistance for weakly supervised semantic segmentation and other computer vision tasks.Existing semantic segmentation methods based CNN mainly focus on learning pixel-wise labels from an image directly,while we advocate tackling the pixel-wise segmentation problem by considering the image-level classification labels.Theoretically,we analyze and discuss the effects of image-level labels on pixel-wise segmentation from the perspective of information theory.In practice,an end-to-end segmentation model,termed as I2PNet,is built by fusing the image-level and pixel-level networks.A generative network is included to reconstruct the input image and further boost the training of segmentation model with an auxiliary loss.Extensive experimental results on benchmark datasets demonstrate the effectiveness of the proposed method.In order to reduce the cost of data annotation,weakly supervised semantic segmentation has attracted researchers'attention.For reason that the saliency detection can bring positive assistance for weakly supervised semantic segmentation,we proposed a new architecture for saliency detection.The proposed GDPNet(Gating for Double Pyramidal Networks)contains two pyramidal structures:Feature Pyramid Network(FPN)and Pyramid Pooling Module(PPM).FPN has the capability to capture the inherent multi-scale and pyramidal hierarchy.while PPM can exploit the global context information by different-region-based context aggregation.However,existing irrelevant information corresponding to non-salient objects or backgrounds may suppress the model's performance.Therefore,we introduce Cross-gating for FPN and Single-gating for PPM to filter out the existing irrelevant information in hidden layers.Experimental results show that our method achieves new state-of-the-art performance on five benchmark datasets.