The Recognition And Similarity Matching Of Women’s Trousers Silhouette Based On CaffeNet Models

Clothing style classification identification and similarity matching have gradually become one of the research hotspots in the field of computer vision with the explosive growth of image information and the diversification of clothing categories in recent years.It is difficult to classify and identify the same style of clothing because of the similarity of external profile in the research on the classification and identification of clothing styles,and the results are often not ideal;Most similarity matching algorithms use low-level visual features to measure the similarity between images,and manual features need to be set manually,How to improve the classification recognition rate of the same style clothing and match the similar clothing image across semantic gap is the key problem.The silhouette is the basic feature of clothing,therefore,the style of women’s trousers was taked as an example in this project,similar matching was conducted based on the classification and recognition of the silhouette of women’s trousers,and a visual interface for similar matching of the silhouette of women’s trousers was built.Firstly,according to the overall appearance silhouette difference of women’s trousers and the classification of women’s trousers version by e-commerce platform merchants,1525 images of women’s trousers with different silhouette were collected in this project,and a sample database of women’s trousers silhouette including saggy pants,straight pants,flared pants,broad-legged pants and pencil pants was created.Second,CaffeNet model of convolutional neural network was applied to the classification and identification study of the women’s trousers silhouette,the network weight parameters were updated step by step using the back propagation algorithm,the network model structure suitable for the silhouette classification and recognition of women’s trousers was obtained combining with the pre-training experiment,including five convolution layers,three down sampling layers,three full connection layer and a Softmax layer,the silhouette features of women’s trousers were extracted by alternating convolution and descending sampling;Thirdly,median filter was used to preprocess the image of women’s trousers to eliminate the noise of the image.CaffeNet model was fine-tuned by modifying the network’s super-parameter file,and gradient descent method was used to minimize the loss function.The improved model was used to classify different silhouette trousers,and the classification recognition rate was more than 95%.Then,the perceptive hash feature of the image of women’s trousers image was extracted based on discrete cosine transform to retain the low-frequency information of the image of women’s trousers and obtain the overall framework,the feature vector of women’s trousers image was quantized and encoded into binary form and finally mapped into perceptual hash sequence,and the hamming distance between sequences was calculated to measure the similarity of perceived hash sequence values;Finally,a visual similarity matching interface of women’s trousers silhouette was built by using Java language,the image data of women’s trousers was read at the front end,and the the perceptual hash fingerprint of the queried women’s trousers image and the corresponding silhouette sample database image was extracted automatically at the back end,the similar silhouette images of women’s trousers were returned to the front end by comparing the perceptive hash binary sequence and realized the similarity matching.To sum up,this topic aimed at the research on the classification and identification of women’s trousers silhouette,CaffeNet model was improved and a method applicable to the classification and identification of women’s trousers silhouette was proposed,which provided an effective way for the visual classification and identification of clothing commodities;The visual similarity matching interface of women’s trousers silhouette is built to search for images,which can adapt to the rapidly growing demand of image data and reduce the manual workload.