The authors separate datasets were compiled: FruitNet and FruitBox Dataset. They proposed an interlaced deep neural framework that aids in the accurate qualitative (freshness—good/bad) and quantitative (weights in kilograms) predictive analysis of tropical fruits and their origin in wholesale and retail fruit boxes.

Motivation

Fruits have always played a vital role in maintaining a healthy diet, and preserving their quality is paramount for individuals’ well-being. The prevalence of consumerism has significantly impacted the production sector, leading to the packaging of fruits in containers and boxes by manufacturers. Accurately determining the precise weights of these fruit containers is essential to ensure transparency and fairness. Traditional techniques are time and cost-consuming and may even result in weight manipulation due to their inefficiency. Fruits can be graded according to their physical characteristics to help determine their freshness; however, this method requires a lot of manual labor. The issue of estimating the quantity and freshness of eatable produce only from images is complex, and no compelling study has been researched and developed.

The proposed solution employs an integrated deep neural architecture to facilitate precise identification of fruits within wholesale fruit boxes, followed by comprehensive qualitative and quantitative predictive analysis. Furthermore, the algorithm takes into account various fruit classes to determine the weight of the box and ascertain its origin. This adaptable technique holds promise for diverse wholesale sectors. Its primary aim is to supplant manual inspection methods entirely with more efficient and cost-effective alternatives, thus significantly reducing expenses associated with manual inspection. The development of fruit classification systems has been driven by the needs of industrial settings such as factories and hypermarkets, where accurate sorting of different fruit varieties is paramount. Recent progress, achieved by integrating public benchmark datasets with cutting-edge convolutional neural networks (CNNs), has led to the deployment of highly accurate fruit classification models. A comprehensive dataset was compiled by amalgamating high-quality fruit images captured physically with images scraped from public sources, with a focus on including non-curated images in automated systems. To tackle data scarcity, researchers have explored the use of generative models in addition to traditional data augmentation techniques. Initial studies focused on using shape and size features to detect and localize fruits on plants. Gradually, solutions shifted toward the utilization of deep neural architectures.

Dataset description

Separate datasets were curated to develop an accurate quality recognition system, estimate weight measures, and determine geographic origin. The annotated FruitNet dataset was compiled by localizing annotations on fruit classes in the FruitNet dataset. To streamline the process of fruit box detection, the authors enhanced the quality detector training to include box detections, thereby collecting and annotating images of containersand baskets of apples from market traders. The instances were manually labeled using the LabelImg tool within the Python environment. The initial dataset comprised images depicting various qualities of six different fruits, captured from diverse angles and backgrounds. This study focuses on three specific fruit categories that share similar shapes and possess an adequate number of instances. The objective is to discern and categorize the quality of individual fruits within a fruit box as either good or bad. Specifically, the authors chose images containing solitary instances of apples, oranges, and guavas. Moreover, meticulous annotation was applied to the apples within containers and fruit boxes to facilitate grading detection. The FruitBox dataset was meticulously compiled, featuring containerized apples along with their corresponding weights, obtained using a 12-megapixel camera from trading stores and markets. To enrich the dataset, images of fruit boxes were sourced through web scraping. Annotations generated from the object detection model assisted in determining the regions of interest for weight calculation. This comprehensive dataset comprises 2052 instances of fruit boxes, each accompanied by its measured weight in kilograms.

The Fruit360 dataset, containing 131 different fruit types, was utilized for the recognition of fruit variants and their respective origins. Specifically, eight variants of apples originating from six different countries were grouped together to estimate the origins of fruit boxes. These variants were then divided into training and testing sets to aid in the learning process. The model’s versatility allows for easy transferability to different fruit types, depending on the specific fruits contained within the boxes.