An Improved Target Detection Algorithm model for Garment image Detection

Abstract

With the rapid development of the Internet platform, users can choose and match clothes according to their personal preferences without leaving home. Merchants have manually sorted and uploaded a large number of clothing images to make it easy for users to shop for clothing online, which consumes a huge amount of labor costs. Such problems can be improved through deep learning related algorithms. However, the conventional deep learning model has a huge amount of computation, resulting in low efficiency of real-time detection of clothing, which limits its application field. Aiming at these theoretical and practical problems, this thesis studies the optimization of clothing image detection and label recognition methods based on deep learning. In view of the real problems of high computational load and slow instant response of existing clothing detection models. This thesis proposes a clothing detection model YOLOv4-GS based on a deep learning framework. Experiments show that compared with the model YOLOv4, this model has a great improvement in detection accuracy and model efficiency. This algorithm first uses the K-means++ clustering method to preprocess the initial dataset DeepFashion2. And construct the GS module based on the deep fusion of Ghost module and SimAM attention mechanism. Then use the GS module to reconstruct the YOLOv4 network to obtain the model YOLOv4-GS, which has higher efficiency and higher model accuracy.