Clothes identification and classification method based on F-CDSSD

Abstract

The invention provides an F-CDSSD-based lightweight clothes identification and classification method, which comprises that a data enhancement technology is adopted to solve the problem that a trainingdata set is too small, and avoiding an over-fitting phenomenon; a traditional VGG-16 network is replaced with ShuffleNet, and fuzzy processing is performed on feature maps output by all convolution layers, so that the complexity of a network structure is reduced, and the network structure can run on a light computing platform; the size, the aspect ratio and the number of default boxes are adjusted, so that the precision and the efficiency of the neural network are improved; a fuzzy theory is introduced for an unclear image, so that the recognition precision of a low-quality clothes image is improved; active learning and iterative learning are realized, a model detection result is fully utilized, and continuous optimization of a neural network model is realized.

Description

technical field [0001] The invention relates to the Internet of Things, clothing identification and classification based on deep learning, active learning, fuzzy theory, concurrent execution and caching technology, and specifically relates to a clothing identification and classification method based on F-CDSSD. Background technique [0002] In recent years, deep learning has developed rapidly, pushing target detection and classification to a new level. The convolutional neural network has strong adaptability to factors such as geometric transformation, deformation, and illumination of the target, and effectively overcomes the recognition resistance caused by the changing appearance of clothing. It can automatically generate descriptions of corresponding features according to the data input into the network, and has high flexibility and universality. At present, target detection in images is mainly divided into single-stage methods (such as SSD, YOLO) and two-stage methods (...

AI Technical Summary

Problems solved by technology

With YOLO as the main body and improved, the proposed SSD increases the detection of multi-scale targets, which largely solves the problems of YOLO's difficulty in detecting small targets and inaccurate positioning.

However, there are too many inappropriate default boxes in the SSD algorithm, which will cause a lot of waste of calculation in the subsequent regression; the neural network obtained by the traditional VGG-16 network structure in SSD is not suitable for light ARM computing platforms. In other words, the structure is too large to run on light computing devices; at the same time, the recognition effect on relatively blurred images is poor

Images