quality analysis of shopee seller portal and semi-supervised learning approach for automatic detection and fashion product category prediction using fc -yolov4

Abstract

Consumers have shifted away from traditional transactions to online shopping as a result of the COVID-19 pandemic, digital social life, and staying at home. This significantly impacts the volume of transactions for several e-commerce platforms, including Shopee, which must continue improving customer satisfaction. One way to accomplish this is to improve the accuracy of the automatic category recommendation, which currently contains a significant number of errors made by the seller. Over the past few decades, object identification in images has evolved quickly. Achieving fast and accurate detection of fashion products in the ecommerce environment is essential for selecting the right category. Nowadays, e-commerce sites provide the purchase of both new and second-hand clothing. Therefore, when categorizing fashion clothing, it is essential to be able to categorize it precisely, regardless of the cluttered background. The Shopee Data Scraper was used to collect data, and two qualitative analyses, namely content and thematic analysis were used to determine how many instances of automatic category recommendation fraud were committed by the seller. This research also proposes the Fashion Category detector FC-YOLOv4 algorithm for the detection of multi-class fashion products and accessories categories. We present recently acquired tiny product images with various resolutions, sizes, and position datasets from the Shopee E-commerce (Thailand) website. We used the semi-supervised learning approach to reduce image labeling time, and the number of resulting images is then increased through the use of brightness, mosaic, rotation, and CLAHE augmentation.

According to the content analysis, there is a 29 percent error in product category selection. Meanwhile, 75.1 percent of products should be classified separately but are instead classified as 'others.' Following that, 72.7 percent of product titles contain the same words as existing categories. From this research, we analyzed the percentage of errors in the automatic category recommendation mechanism from the Shopee platform, which causes sellers to place their products in the wrong category so that they can be used as suggestions for improvement or further advance of research. And our FC-YOLOv4 approach results in reasonable Average Precision (AP), Mean Average Precision (mAP), True or False Positive (TP/ FP), Recall, Intersection over Union (IoU), and reliable object detection. According to experimental findings, our model increases the mAP by 0.07 percent and 40.2 percent increment compared to the original YOLOv4 and YOLOv3. Experimental findings from our FC-YOLOv4 model demonstrate that it can effectively provide accurate fashion category detection for properly captured and clutter images compared to the YOLOv4 and YOLOv3 models. Then, we utilized TensorFlow Lite to implement our FC-YOLOv4 model on the android platform. Then, we compared the detection findings of our FC-YOLOv4 Android software with those of the Shopee Thailand Seller Web Portal.