Unsupervised Image Semantic Analysis Via Deep Learning

Recently,deep learning has achieved remarkable successes in various semantic analysis tasks of computer vision.This paper focuses on two tasks,i.e.image segmentation and face detection,and accordingly proposed two unsupervised image semantic analysis techniques based on deep learning.Specially,image segmentation aims at dividing an image into a number of meaningful parts without intersection semantically by some special rules in pixel level.Image detection(i.e.,object detection)can locate and classify the objects within proposals precisely in the meantime.The research in this paper is face detection(i.e.,reducing the range of object detection to face level),which detects faces in complicated circumstances and provides further works with face pre-processing.Image segmentation and face detection are two of the most popular research topics.However,existing methods have several weaknesses.Specifically,existing unsupervised image segmentation methods have several limitations such as bad precision,bad robustness and low efficiency of computational time.Face detection methods suffers from undetected error,high fallout ratio of non-face,the trading off computational time costing and precision and so on.The purpose of this paper are 1)boosting the efficiency and robustness of models in unsupervised image segmentation,and 2)optimizing leak,false detection error,and trading off between the computational complexity and precision.To address these concerns,two works are presented in this paper:1)First,this paper presents a new unsupervised image segmentation algorithm based on deep learning,named SDAHPI(Stacked Denoising Auto-encoder with Hierarchical Patch Indexing).In SDAHPI,we first extract deep-level feature representations from non-overlapped patches and then reduce dimension using Stacked Denoising Autoencoder(SDA).Afterwards,we perform hierarchical k-means clustering on these feature representations and build an indexing tree structure.SDAHPI shows significant improvements in both the computational efficiency and the segmentation accuracy on standard datasets.2)Second,we propose a face detection framework based on Faster R-CNN.First,we achieve one point improvement in the segmentation accuracy by online mining difficult negative samples and combining focal loss.Second,to reduce the timeconsuming,we proposes an improvement framework,which support parallel operation rather than testing images one by one.The proposed method can not only improve single GPU's throughput(i.e.,4 times by origin),but also improves the efficiency.Besides,the proposed method considerably reduces the undetected error.In particular,for cartoon faces,the proposed method reduces 90% undetected error samples than existing approaches.In conclusion,this paper proposes a novel unsupervised approach for object segmentation and face detection,respectively.The proposed approaches are novel in theory and achieve considerable performance improvement.This paper makes certain contributions to the area of unsupervised image semantic analysis.