The Research On Automatic Extraction Of Clothing Style Based On Image Analysis

The research of clothing attribute includes detecting and identifying clothing parts and contour detection on an image.Due to the limitation of artificially assisted feature extraction algorithm and the lack of public datasets,clothing attribute identify and search problem has enormous challenges in actual applications.We recommendation a model based on fast regions with convolutional neural network(Fast R-CNN)and combine with Grab Cut,and based on the visual mechanism,the edge extraction of enhanced weak contours is proposed.Our contributions can be summarized as:We identifying clothing parts with Fast R-CNN.We merge the labels of street images into 6 categories(e.g.cloth,pants,bag).We present a model based on Fast R-CNN for clothing style classification and retrieval task and combine with Grab Cut.We first establish the our clothing dataset including clothing images of different styles.We train fast regions with convolutional neural network model on clothing dataset and verify on test set.We adjust the parameters of fully connected layer to realize preferable results on identifying clothing parts based on Fast R-CNN.With the expansion of training samples,the accuracy will be higher.Experiments show that our method based on fast regions with convolutional neural network and convolution feature to construct hash index can achieve good results with high accuracy of classification,good effect and fast retrieval.And Grab Cut algorithm is used for segmentation of image,which can locate the positions from complex background images and remove complex backgrounds.Simulating the transmission and processing of visual information in the visual pathway,the main contour information is quickly extracted based on the central and peripheral antagonism mechanisms of the ganglion cells.Then,the difference between the Gaussianfunction and the Gaussian difference function is used to simulate the modulation of the non-classical receptive field of the external geniculate body,and the suppression of the texture background is realized.Then,a multi-directional single-cell receptive field model of V1 region was constructed,and a DOG(difference of Gaussians)response improvement evaluation model based on negative effect was proposed.Finally,the ability of complex cells to characterize visually advanced features in the V1 region is considered.A vision-based visual response fusion model based on parallel processing is implemented to achieve target contour detection and enhancement.The method has good natural contour detection and extraction ability,especially for the detection of partial weak contour edges of images.The new model constructed in this paper will contribute to the understanding of the functional and intrinsic mechanisms of the visual pathways,and will also provide a new idea for image analysis and understanding based on visual mechanisms.