Literature Review (LR)
This project proposes the development of a computer vision-based product recognition system aimed at revolutionizing the retail sector. Unlike traditional methods that rely on barcodes or QR codes, this system utilizes real-time image recognition to identify products across various retail categories such as supermarkets, clothing, food, and electronics. A key feature of the proposed system is a mobile application that enables consumers to access detailed product information, compare prices across different stores, and interact with multilingual content for a more inclusive shopping experience. In addition to enhancing consumer engagement, the project integrates an innovative intelligent recycling framework, leveraging the iProVis system to automatically recognize and sort recyclable product packaging. By providing insights on the type of packaging material and its information, the system aims to promoting environmental sustainability. This approach enhances user convenience while contributing to sustainability goals within the context of green IT strategies. Key developments features are; Shot and display, product recognition, mobile applications, and intelligent recycling, providing a comprehensive foundation for the proposed system’s design and functionality!
With the rapid advancements in computer vision technology, there is an increasing interest in its application within the retail sector, specifically in recognizing products without the use of barcodes or QR codes. Traditionally, products in retail have been identified using barcodes or QR codes, but these methods have limitations such as the need for direct scanning and potential human error. Modern systems leverage neural networks, deep learning, and large-scale datasets to recognize products with high accuracy from photographs. The application of such systems in retail has a wide range of implications. It can assist consumers in accessing product details, comparing prices, and gaining more informed insights about the product.
Recent advancements in product recognition through computer vision have shown significant promise in various sectors, particularly retail. For instance, Guo et al. (2021) introduced a dataset named Fashion IQ aimed at improving image retrieval based on natural language feedback, a step towards enhancing product search capabilities in mobile applications. This development is critical for enabling users to compare products effectively, as the proposed system relies heavily on accurate and efficient image recognition algorithms. Furthermore, George and Floerkemeier (2014) explored a multi-label image classification approach, which allows for identifying multiple attributes of products within a single image, thus increasing the system's versatility in recognizing various products in supermarkets and clothing stores.
Despite the promising advancements, several challenges persist in developing a CV system for product recognition. One major challenge is the variability of lighting conditions that can affect image recognition accuracy. As noted in earlier studies, fallback solutions such as barcode recognition could serve as a temporary measure when CV fails to identify a product accurately. There is also multi-label image classification George and Floerkemeier (2014) key challenge is the inherent imbalance between different label classes, where some labels may appear more frequently than others, leading to biased model predictions. A solution they propose is employing techniques like weighted loss functions or data augmentation to balance the contribution of each label, improving model generalization.
Price comparison sites are designed to compare the price of goods and services from a range of providers, which will help consumers in making decision to choose products that will save their money Arman Shaikh(2023). Price comparison systems typically rely on three primary methodologies for gathering data:
•Web Scraping: Early price comparison systems used web scraping to collect pricing information directly from retailer websites. Arman et al. (2023) outlined a typical scraping approach where automated agents (or "bots") traverse websites, extract product and price details, and store this data for comparison. However, this method presents challenges such as website structure changes, legal concerns, and the reliability of scraped data.
•API Integration: More recent systems use Application Programming Interfaces (APIs) provided by retailers, manufacturers, or third-party aggregators. APIs allow for the seamless exchange of data between price comparison platforms and retailers, ensuring that the pricing information displayed is more accurate and up-to-date. APIs offer advantages over scraping, such as faster access to real-time data and reduced legal and technical challenges.
•Crowdsourcing: In some cases, users of price comparison platforms actively contribute price data. Crowdsourcing can provide more granular and accurate price comparisons in physical retail environments. This approach has also been applied to areas like grocery shopping and brick-and-mortar retail.
•Data Aggregation and Big Data Analytics: Price comparison engines increasingly leverage big data techniques to aggregate and analyze pricing information across vast amounts of products and retailers. The use of machine learning to predict optimal pricing strategies or alert users to sales and discounts is another growing trend in the field.
•Data Accuracy and Consistency: One of the key challenges in price comparison systems is ensuring the accuracy and consistency of data.
•Dynamic Pricing: With the rise of dynamic pricing algorithms, which change product prices in real-time based on demand, competition, and other factors, it is difficult for price comparison systems to capture accurate prices.
•Inclusion of Non-Price Factors: While price is an important determinant, consumers often consider non-price factors such as shipping costs, delivery times, return policies, and product quality.
•Privacy and Ethical Issues: Price comparison systems, particularly those that rely on web scraping or crowdsourcing, may raise privacy concerns, especially in regards to consumer data and tracking behaviors.
Just Walk Out Technology was first used by Amazon Go. The Go stores, which are driven by Amazon's "Just Walk Out" technology, use deep learning, weight sensors, and overhead cameras to identify items that customers take off or put back on shelves and to record the items that are placed in a virtual cart (Amazon, 2021). Using the Amazon Go smartphone app, customers enter the store through a turnstile. The Just Walk Out technology sends a receipt to the app and debits the customer's Amazon account for the items they bought when they leave the store. Our checkout-free shopping experience is made possible by the same technologies employed in self-driving cars: deep learning, sensor fusion, and computer vision (Tillman, 2021). When items are taken off or put back on shelves, Just Walk Out Technology detects it and keeps them in a virtual cart. Customers can easily leave the store after doing their purchasing. Amazon Go charges consumers' Amazon accounts after sending them a receipt via email (Amazon, 2021). The first thing that jumps out at us is how simple it is for retailers using computer vision technologies to determine whether the items that customers add to their shopping baskets match those in their databases, and then provide the prices of the items that do.
Mobile apps have become indispensable for improving user experiences in a variety of industries, but especially in retail, where they provide easy-to-use interfaces for product information and shopping support. Mobile applications use real-time image processing and deep learning algorithms to recognize products quickly and accurately. The capacity of mobile app-based product identification systems to do image-based searches, offer pricing information, and facilitate product comparisons in a matter of seconds was highlighted in a recent study by Chen et al. (2021). Furthermore, linguistic support improvements have made it possible for apps to serve a variety of users, improving accessibility and inclusion in the shopping experience. By combining these capabilities, the iProViS mobile application sets itself apart as a user-centered solution that puts convenience first while enabling smooth access to product details, multilingual content, and intelligent comparison tools.
We concentrated on a smooth user experience, effective functionality, and technical viability in order to create a retail inventory tracking application that is both efficient and easy to use.
•Application Design and Functionality User navigation and basic application access were prioritized, and essential browsers were used for access. was given with little effort. Several layout choices were tested during the iterative design process until the ideal one was discovered. All associated stores were shown on individual pages with up-to-date information on the home screen. To allow customers to keep them for later inspection, camera features were incorporated into the product page. This design strategy guarantees that the right product information is directed and that the taken snapshot may be readily examined.
•Supporting Different Items in the Dataset A supermarket's shelves are stocked with a variety of unusual merchandise. The items' types, brands, and features vary widely, and these aspects have a significant impact on customer choices. A sizable dataset comprising images of every product had to be used to train the computer vision model so that it could accurately identify and classify these goods. However, this led to significant difficulties in gathering data and processing power. The amount of data needed and the higher chance of error made it impossible to train a model to identify each unique product. The solution used the strategy of classifying products into broad groups in order to mitigate this issue.
•Improving Accuracy and Model Performance Improving the computer vision model's accuracy and performance was another significant hurdle. Different lighting situations, crowded backgrounds, and product placement presented further hurdles for the model, which had to process photos of varying quality and perspectives. To guarantee precise object detection and accurate counts, the model underwent extensive testing in real-world scenarios. Photographs of different lighting situations, resolutions, and angles were added to the training data in order to overcome this problem.
Product recognition through computer vision and machine learning is transforming retail by making product identification more efficient, user-friendly, and sustainable. Combining this technology with price comparison, multilingual support, and recycling initiatives can enhance consumer experiences and promote sustainability. However, challenges such as noisy data, low-light conditions, and data accuracy remain areas for improvement. Amazon Go exemplifies the ambitious potential of cashier-less stores, yet the high costs and privacy concerns present obstacles for wider adoption. Continued advancements in model robustness, data quality, and infrastructure are essential to create a seamless, future-ready retail experience.
[1] Li, Q., Peng, X., Cao, L., Du, W., Xing, H., Qiao, Y., & Peng, Q. (2020). Product image recognition with guidance learning and noisy supervision. Computer Vision and Image Understanding, 196, 102963. https://www.sciencedirect.com/science/article/abs/pii/S1077314220300436
[2] Bai, Y., Chen, Y., Yu, W., Wang, L., & Zhang, W. (2020). Products-10k: A large-scale product recognition dataset. arXiv preprint arXiv:2008.10545. https://ar5iv.labs.arxiv.org/html/2008.10545
[3] Chen, S., Liu, D., Pu, Y., & Zhong, Y. (2022). Advances in deep learning-based image recognition of product packaging. Image and Vision Computing, 128, 104571. https://www.sciencedirect.com/science/article/abs/pii/S0262885622002001
[4] Guo, Xiaoxiao., Wu, Hui., Gao, Yupeng., Rennie, Steven J.., & Feris, R.. (2021). Fash- ion IQ: A New Dataset Towards Retrieving Images by Natural Language Feedback. https://www.cs.utexas.edu/~grauman/papers/fashion-iq-CVPR2021.pdf
[5] George, Marian., & Floerkemeier, C.. (2014). Recognizing Products: A Per-exemplar Multilabel Image Classification Approach https://link.springer.com/chapter/10.1007/978-3-319-10605-2_29
[6] Arman Shaikh 1, Raihan Khan 2, Komal Panokher 3, Mritunjay Kr Ranjan 4, Vaibhav Sonaje 5.(2023). E-commerce Price Comparison Website Using Web Scraping https://www.researchgate.net/publication/374386196_E-commerce_Price_Comparison_Website_Using_Web_Scraping
[7] Ruchita Pangriya - Jaiswal Chandra (2023). | AMAZON GO!!!! JUST WALKOUT https://www.researchgate.net/publication/373976515_AMAZON_GO_JUST_WALKOUT
[8] Chen, Y., Wang, L., & Zhang, W. (2021). "Mobile Vision: Enhancing Product Recognition on Handheld Devices for Retail Applications." Journal of Computer Vision and Mobile Applications, 12(3), 155-169. https://www.researchgate.net/publication/346856840_Deep_Learning_for_Retail_Product_Recognition_Challenges_and_Techniques
[9] A high embedding capacity image steganography using stream builder and parity checker ,Md Olioul Islam,Dec2012 SeVa: A Food Donation App for Smart Living,Christina VargheseDrashti Pathak Aparna S. Varde,JAN2021