Research And Application On Clothing Collar Classification Based On Convolutional Neural Network

In recent years,with the rapid development of the Internet and the continuous emergence of electronic shopping platforms,online shopping has become more convenient and faster.People gradually choose to buy clothes from offline stores instead of offline stores.In addition,with the development of the clothing industry,in the field of clothing design and clothing sales,higher requirements have been put forward for the management of clothing pictures,especially the classification search,especially clothing attributes and details,such as clothing of color and collars etc.Therefore,it has become a growing demand to use collar types to retrieve clothing on shopping websites.At present,the research in the field of clothing mainly focuses on the feature extraction of clothing categories and attributes,including clothing color,shape,texture,etc.,but only a few researchers pay attention to the detail classification of clothing and the detail classification of clothing,such as collar.At present,the images of garment collar that can be easily collected contain a lot of background noise,and the collar part often has such factors as deformation,different shooting perspectives,and different pantograph ratio.If the traditional method is used to classify the image of collar,the model is difficult to eliminate the interference of background factors,and the difficulty of image classification is increased.In addition,there is currently no targeted collar image data set publicly available for researchers to use.With the opening of the era of intelligence and big data,traditional methods cannot efficiently cope with complex and high-resolution images in actual application scenarios,showing the shortcomings of low recognition rate,low efficiency,and low adaptability.The convolutional neural network has a close correlation among the levels.It is good at mining the local features of the image,and can process the rich spatial information in the image and extract the key features.In order to further improve the classification performance of collar images and meet the needs of real production and life,and in view of the major breakthroughs made by convolutional neural networks in the field of image recognition in recent years,in this paper,two sets of collar data sets and two methods of collar image classification based on convolutional neural network are proposed.The main research contents are as follows:1.In view of the lack of collar image datasets in the current image classification field,this paper constructs two sets of collar image datasets,named Collar-4 and Collar-6 respectively.The Collar-4 dataset includes four categories: round collar,lapel collar,stand collar,and hooded collar,with a total of 39248 images.The Collar-6 dataset includes six categories of round collar,lapel,stand collar,hooded collar,V-neck,and fur lapel,with a total of 18,847 images.The Collar-6 dataset includes six categories of round neck,lapel,stand-up collar,hooded collar,V-neck,and fur lapel,with a total of 18,847 images.Compared with the Collar-4 dataset,the Collar-6 dataset has a smaller number of images per category,and two additional categories are added.Both datasets are collected from the Internet.Collar image has many features,such as multi-pose,multi-noise and small classification area,which brings many challenges to collar image classification.2.In order to solve the image classification problem of the collar dataset containing a lot of noise and improve the ability of existing algorithms to process the classification of real collar images,this paper proposes a classification for real collar images based on the constructed Collar-4 dataset.The algorithm is improved by adding the attention mechanism CBAM block to the Fire module in the Squeeze Net network.The improved module is called Fire CB,and the improved algorithm is called Squeeze Net-CB.In the experimental part,compare the initialization model and transfer model of Squeeze Net-CB with the initialization model and transfer model of traditional convolutional neural networks.Experimental results show that the classification effect of Squeeze Net-CB is better than traditional convolutional neural networks.In the ablation experiment,a simplified comparison experiment of squeezenet-CB was performed,which further proved the effectiveness of the proposed method.The experiments show that the application of Squeeze Net-CB on Collar-4 dataset is feasible,and the algorithm can effectively solve the collar image classification problem with noisy background in the real world.3.The traditional image recognition methods have a good effect on coarse-grained images,but they are unable to accurately distinguish and grasp the main features of images when dealing with complex backgrounds,especially those with overly fine-grained images.In order to solve the above problems and discuss the fine-grained image classification performance of convolutional neural network on small collar images,this paper proposes an M-Res Net50 classification model for garment collar image classification,which is designed on the basis of ECA-Res Net50 and combined with MC-Loss.The datasets used in the experiment include the Coller-6 dataset constructed in this paper and the public dataset Deepfashion.The experimental results show that the improved model has higher accuracy and better feature extraction ability than the existing CNN model.It solves the problem of difficulty in classifying fine-grained collar images and promotes the further development of image classification of clothing commodities.The algorithm and model proposed in this paper have achieved good experimental results in several groups of comparative experiments,which verify that the application of the proposed method in the collar dataset is feasible,and can effectively solve the collar image classification problem with noisy background in the real world.In addition,the distribution of Collar-4 and Collar-6 image data proposed in this paper is verified to be reasonable,which can support the construction of classification model and can be used as the training set data for collar classification task in the real world,which has practical significance for the research on the classification and retrieval of clothing collars.