Research On Object Detection And Semantic Segmentation Based On Representation Learning

Deep learning,which is a significant branch of representation learning,arises a new urge of a large number of visual recognition tasks development.Although it has been studied for dozens of years,visual recognition tasks are still immature facing the complicated real world and applied platforms with relatively scarce computing resources.As the key component of visual recognition tasks,object detection and semantic segmentation are always stuck with the tradeoff between the model accuracy and computation cost.A main idea of this thesis is combining the theory of representation learning to design visual recognition models with low intrisic computation complexity and inference latency.The main research achievements are as follows:We review various works in deep convolution neural network based on representation learning for visual recognition tasks,introduce the representative architecture and light-weight neural networks.The thesis categorizes representation learning into three groups: multi-scale feature learning,contextual feature learning and relation network.For multi-scale feature learning,we first divide it into four paradigms: image pyramid,prediction pyramid,integrated features and feature pyramid.Next,we summarize the contextual feature learning of two families: global contextual feature learning and local contextual feature learning.Then we give a detailed introduction of relation networks in visual recognition tasks,including graph convolution networks and self-attention neural networks.Existing skin lesion segmentation networks based on a large amount of floating point operations and long runtime,it is difficult to deploy models to medical devices.To tackle this issue,we employ contextual feature to extract multi-level context aware information around skin lesions.Further feature filtering with self-attention modules improves the discriminability of lesions information.Therefore,we combine context aware local features with attention modules as the basic unit to propose a fast light-weight model for skin lesion segmentation.In particular,our model achieved the Jaccard index of 80.9% on ISBI 2017 Skin Lesion Segmentation dataset.Furthermore,our model has less than 0.5M parameters,and can process a dermoscopy image with 768 × 1024 resolution at a speed of 20 FPS on only one NVIDIA TITAN X.Extensive experiments well demonstrate the effectiveness of our proposed work for skin lesion segmentation tasks.Shallow layer features in convolution neural networks have less semantic information,which hinder detecting small face targets in images.Under the premise of not signigicantly increasing the cost of computation,we combine astrous convolution filter with spatial pyramid module to design light-weight feature enhance block.One stage detection framework is exploited to refine multi-scale feature information.Moreover,anchor density strategy is introduced to improve the recall rate of tiny faces.Extensive experiments on datasets show that our light-weight face detector infer images of 1024 × 1024 resolution at 64 FPS.Specially,the mAP of our framework surpasses two stage detector named Faster RCNN on FDDB dataset.The results on multiple face detection benchmarks indicate that our detector improve efficiency and performance simultaneously.Working on the multi-scale feature learning and contextual feature learning,two frameworks of convolution neural networks are designed for skin lesion segmentation and face detection tasks.These kinds of visual recognition models achieve high performance and efficiency.We believe that our work is useful for promoting the final application of visual recognition models.