A novel approach of boundary preservative apparel detection and classification of fashion images using deep learning

Abstract

Visual analysis of fashion images gain much attention in the fashion industry due to its commercial and social importance. In recent years, deep learning techniques offer overwhelming progress in improving the accuracy of fine‐grained apparel segmentation with accurate bounding box prediction. The baseline pixel‐based masking techniques show excellent performance in object detection and segmentation but sometimes ignores the boundary of objects, resulting in uneven and complicated segmentation masks. Moreover, it is time taking to generate a multi‐scale feature map against each anchor box. To remedy this problem, a more accurate, faster, and suitable deep learning architecture is proposed that accurately detects, classify, and performs fine‐grained segmentation of cloth products in a single platform. In this paper, initially, an Object Class Head Detector model is proposed in which the baseline Mask‐RCNN model is used as a reference model. Here, we replace the Region Proposal Network with the proposed modified YoloV2 model to locate apparel products with its class prediction. The modified YoloV2 model has more capability to detect tiny objects because of local and high‐level feature fusion. The goal of this step is to accurately locate the objects in minimum time intervals. Furthermore, the predicted bounding box is converted to object shape offsets using deep snake architecture that tightly fits onto the apparel shape. It can improve the accuracy of cloth shape segmentation by preserving object contours. The proposed architecture is empirically validated on various existing fashion image datasets. The experimental results illustrate that the proposed architecture performs better on the Deepfashion2 dataset with mAP of 86.86%, as compared to other state‐of‐the‐art deep learning models.