Abstract:

Objective Color recognition of human clothing has become a topic with widespread interest recent years due to its potential to address various issues such as clothing retail, smart security, and fashion recommendation. An accurate color recognition model for human clothing can greatly enhance user experience and service quality. However, accurately recognizing the color of clothing on a human body in images can be challenging due to the multi-angle and multi-pose nature of the clothing, as well as the influence of the material and texture of the clothing, the color and texture of the human body background, and other factors. Therefore, constructing an efficient and accurate human clothing color recognition model is important and necessary to be solved in order to improve recognition accuracy and stability.

Method To enable color recognition of human clothing, a human clothing attribute dataset called pcaparsing was first created. Then, an end-to-end convolutional neural network model called ClothResNet was constructed, which used ResNet18 as the backbone network. The model also featured an optimized pyramid pooling module responsible for capturing multi-level semantic information and a coordinate attention mechanism with a focus on the contour information of human clothing. Additionally, atrous convolution was used to improve network efficiency. The dataset was split into training and testing data at an 8∶2 ratio, with 80% of the dataset used for training the ClothResNet model and the remaining 20% used for testing its effectiveness. Comparative experiments were conducted between the traditional clothing color recognition methods of K-means and Hog-KNN, the deep learning clothing color recognition method of CNN, and the method proposed in this paper. Ablation experiments were also conducted to demonstrate the effectiveness of selecting ResNet18 as the backbone network, expanding the dataset, and adding each module to the model. Overall, the study aimed to improve the accuracy and efficiency of color recognition for human clothing.

Results This study aimed to recognize the color of human clothing in natural scene images using a convolutional neural network algorithm. The use of proposed network model, ClothResNet, achieved 94.49% accuracy rate in recognizing 12 different colors, demonstrating the feasibility and effectiveness of the deep learning method compared to traditional methods. To evaluate the effectiveness of the ClothResNet model, comparative experiments were conducted with traditional methods, and ablation experiments were designed and carried out. Expanding the dataset significantly improved the evaluation indicators of each network (Tab. 1 and Tab. 2). Furthermore, the addition of the pyramid pooling block and coordinated attention module further enhanced the performance of the model. These ablation experiments demonstrated the effectiveness of the extended dataset and network modules, laying the foundation for future work on automatic color recognition of human clothing. Overall, this study highlights the potential of deep learning methods for accurately and efficiently recognizing clothing colors in natural scene images.

Conclusion In this study, a novel human clothing color recognition model called ClothResNet was proposed, which utilizes ResNet18 as the backbone network and incorporates an improved pyramid pooling module and a coordinate attention mechanism. By combining the strengths of an end-to-end convolutional neural network, this model has led improved recognition of various human clothing colors. Through a series of experiments, we verified the feasibility and effectiveness of our proposed method. This approach provides new ideas for the development of smart clothing, although there is still room for improvement. For instance, there are far more than 12 colors of clothing in reality, so further research is needed to develop methods for recognizing a broader range of colors.

Key words: human clothing color recognition, residual network, optimized pyramid pooling, ClothResNet model