AU2020101729A4 - Continuous labelling assessment of products to improve efficiency of reverse logistics by deep learning model - Google Patents

Abstract

CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL ABSTRACT Reverse Logistics means the method of retrieving products from a buyer to a wholesale or processing origin, which is a progressively significant and inadequately-managed industry task. The development of internet and mobile techniques and its worldwide exponential growth enables online business. The detection of labeled manufactured products in retail shopping often focuses on barcodes, placed a strong emphasis user feedback and being restricted to single item at a moment. This proposal approaches novel deep learning models for assessing the continuous labeling of products. The products are retrieved from the customers and stored into the warehouses and inspected and identified the packed product using combined neural network as Mobilenet v2 with SSD. Further the detected item is classified using deep convolutional network like VGG net as re-sell, scrap, recycle and disposal. This proposal demonstrates the outperforms of continuous labeling assessment with high accuracy. 1| P a g e CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL Drawings A7 714 ch ed dcclo n diulItoa Figure 1: continual labeling assessment of products in reverse logistics using deep learning model 11 P a g e

Description

CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL

Drawings

A7 714 ch ed dcclo n diulItoa

Figure 1: continual labeling assessment of products in reverse logistics using deep learning model

11 P a g e

CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL

Description

Field of the Invention:

This invention relates to improve the efficiency of continuous labeling assessment by using deep learning model in reverse logistics supply chain management. The integrated deep learning convolutional neural networks like MobileNetV2+SSD are employed for feature extraction and detection and the continuous labeling of products using VGG Net model that is used with high accuracy.

Background of the invention:

In latest decades, even though the computer vision discipline has emerged massively. Object detection and classification techniques of processed goods in shopping environment is often in its formative stages, inhibiting the expansion of innovative, more natural living creature-product communication. The significant issue with deep learning is how to select components to be labeled.

Fechteler M et al described the Logic. Cube that nine cameras are used to capture the image data of the parts promptly well after the production process in the large warehouses. The weight of the products was measured and aspects of packaging details of the products are calculated and buffered in a database. Detection and labeling accuracy could be enhanced progressively using unique attributes and yet primarily by using huge amount of images.

Hair et al discussed the research objectives that describe probable interactions only in the most particular context but instead enable multivariate techniques to evaluate

11 P a g e the relationships. In experimental research, the researcher does not seek to verify any association indicated before the assessment, and rather provides the strategy and the information to classify the importance of the interaction. Experimental work is defined as comprehensive observations of deductive logic, and can be either quantitative or empirical.

Aziz et al premised the collections of waste schemes on the image processing of garbage bins have been explored. The optimization techniques identified the bins in the picture and assessed the amount of waste for each bucket utilizing four face mask and a set of support vector classifiers. Perpendicular containers may be flat, partially loaded or packed of waste in different forms and dimensions.

Hsiao recognized the main elements that push and help technologies to improve the reverse logistics capability of an organization. The research results of the empirical scenario assessment show that computer technology, knowledge acquisition and correlation platforms have a beneficial impact on the reversing of the transportation resources of the organization and, correspondingly, on the reversing of the supply chain efficiency in effort to expand its comparative benefit.

Hannan et al implemented Content-based Image Retrieval framework to investigate the potential of predicting the filling amount of waste disposal bins from the surface which is obtained from the pictures. Alternative techniques for the calculation of the resemblance range have been included in the process to measure and analyze the difference among images. The precision extraction techniques was evaluated on a set of 250 sensor data and the analysis indicated that the spacing of the heavy equipment was very precise for computing the correlation of artifacts and presented improved results than other ranges.

21Page

B. Settles discussed emerging deep learning investigation that the training set are inactive, accessible or selected randomly to be classified by domain analysts. Effective in-depth classifying intends at accomplishing the best training outcomes with a restricted collection of classified data. The goal is to know the optimum parameters by choosing a subset of compounds to be labeled and comes under the well-defined context for effective learning. Nevertheless, it remains uncertain how to determine which observations are more descriptive in deep learning model.

Daughterty et al highlight the significance of organization resource budget and assignment for executing the reverse logistics process since reverse logistics systems are expensive and time consuming and organizing inverse flows are very difficult and unpredictable. Enhanced management tools, like managerial focus, top management assistance, improved management expertise, enhancement and deployment of reverse Logistics strategy which have powerful influence on the efficiency of RL. Depending on the outcomes, companies that devote more strategic assets to reversing logistics are performing excellent business on the day-to-day level.

This proposal contributes the deep learning model to improve the efficiency of the classification process of the returned products from the customers.

Objects of the Invention:

• The main aim is to determine the classification accuracy of products in reverse logistics using deep learning models. • Another is to demonstrate the feasibility about the feature extraction and object detection under real-world in reverse logistics environment.

Summary of the Invention: Reverse logistics is the strategy of formulating, executing and managing the reliable, efficient distribution of raw materials, in-process production, manufactured products

31Page and relevant details ranges from point of sale to the point of source for the objective of recovery or proper waste disposal. Object detection and identification are being broadly utilized in automation, electronic components, road safety, automated vehicles, intelligent transportation systems and product identification systems. This proposal is mainly deployed in reverse logistics since class labeling of products consumes more time and very complex process. Therefore the deep learning models are employed for continuous labeling in reverse logistics process. The products are collected from the end customers after delivering due to some defects. The collected packed products are bringing back to warehouses. In warehouse the products are captured by using cameras and preprocessed using some preprocessing techniques like image resampling and contrast enhancement for improving the quality of image and reduce unwanted noise. Further the combined convolutional neural network called Mobilenet v2 and SSD is proposed for feature extraction and object detection. The mobile net v2 is defined as base infrastructure and output is feed into SSD network for object detection. The purpose of the MobileNet v2 layers is to translate pixel values of the source images to characteristics that identify the elements of the picture and transfer into the upcoming layers. MobileNet is exploited as a feature map for a SSD neural network. Based on the detected object, the products are classified as re-sale, recycle, scrap and waste disposal by using deep learning methods like VGG Net convolutional neural network and labeling the products in the continuous manner. This proposal yields the better labeling accuracy using deep learning models, provides less cost and time consumption. Detailed Description of the Invention: Figure 1 explains the deep learning model for continuous labeling of products in reverse logistics. The defected products are collected from the customers and delivered back to the warehouses using transportation facility. The products are captured by using the camera and preprocessed to remove the noises and enhances the quality of the image. The Mobilenet v2 and SSD as the combined neural network model is promoted for feature extraction and object detection. Further VGG Net is employed for classifying the labels as different labels like re-sale, recycle, scrap and disposal. Figure 2 illustrates the architecture of Mobilenet v2 and SSD for object detection. MobileNetV2 incorporates the original fully convoluted layer with 32 filters, trailed

41Page by 19 residual bottleneck layers. ReLU6 function is calculated as a non-linearity since its reliability when exploited in low-precision processing and uses kernel size 3 x 3 as requirement for modern networks, and average pooling during training phase. The first layer yields a constant growth rates across the network. The MobileNet v2 is considered as a base structure for extracting features as feature maps and SSD is deployed in the architecture for object detection. The SSD approach is premised on a feed-forward convolutional network which generates a fixed-size gathering of bounding boxes and ratings for the appearance of products training points in the boundary boxes, preceded by a minimum concealment process for ultimate detections. Figure 3 promotes the novel deep learning model for continuous labeling of products in reverse logistics. The VGG Net 16 is proposed for classifying labels of products. The contribution to the cov Ilayer is a predefined size image of224 x 224. The image is transferred via a bundle of convolutional layers that the filters were utilized with a very simple layer: 3x3. It also deploys 1x1 convolution filters which are defined as configuration for linear transformation of the input channel. The spatial resolution of the input image is maintained after convolution process. Spatial pooling is decided to carry five max-pooling layers that are preceded by convolutional layers. The VGG Net16 consists of number of hidden layers which are configured with rectification non-linearity function. The fully connected layer classifies the products and provides continuous labeling. The final layer called softmax layer that determines the accuracy of the classification and label the products as resale, recycle, scrap and disposal.

51Page

CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL

CLAIMS: 1. The method of proposal comprising, Electronic devices with display and high network connections like 5G enabled and Wi-Fi. 2. Collecting the defected products from the customers and again stored into the warehouses. 3. The products are captured by cameras and processed by some preprocessing methodologies. The preprocessing methodologies like contrast enhancement and image resampling for enhancing the quality of the images. 4. Following claim 3, the learning model also comprising The MobileNet v2 along with SSD is employed for feature extraction and object detection. 5. For claim 4, MobileNet and SSD utilizes the depthwise separable layers and converts the input image pixels into features and act as feature extractor and yields feature map for object detection with less computational cost. 6. For claim 4, the deep learning model comprising Single Shot model uses MobileNet v2 as a base network. SSD utilizes distinct feature maps, from the base network for product detection. A batch of the predefined boundary boxes allocated to specific tiny cell in the feature maps. Further SSD estimates the indexing for features and four bounding box offsets for each feature map. SSD exploits more than one feature maps with various sizes to enhance the high and low level details. 7. The claim 4, also comprising The convolutional neural network called VGG Net 16 is promoted for continuous labeling of products with high accuracy. The VGG Net 16 contains convolutional layer, max-pooling layer and average pooling layer for classification and softmax layer is finally attached to classify the labels as resale, recycle, scrap and disposal with low cost and time consumption.

1 Pag e

CONTINUOUS LABELLING ASSESSMENT OF PRODUCTS TO 08 Aug 2020

IMPROVE EFFICIENCY OF REVERSE LOGISTICS BY DEEP LEARNING MODEL

Drawings 2020101729

Figure 1: continual labeling assessment of products in reverse logistics using deep learning model

1|Page

Figure 2: Mobilenet v2 with SSD architecture

2|Page

Figure 3: VGG Net Architecture

3|Page