Our Customers
Computer Vision based on Deep Learning
3D object recognition
For a Building Information Management company
Business Problem :
The current state of construction site progress measurement is manual. Building information management (BIM) system implementation provider wanted to integrate object recognition with BIM. This will enable BIM to track site construction progress, it will also detect any anomalies in the site construction.
Technology Used :
Python, OpenCV, Tensorflow, Deep Learning using PointNet++
Solution :
The solution is developed using Pointnet++, which is a pre-built deep learning network for classification & segmentation. Pointnet++ is pre-trained on Shapenet & Modelnet, hence provide transfer learning capability for new data. The customer has provided point cloud datasets for small, medium & large objects captured using LIDAR primarily for pump, bend & valve. This data is then processed using python point cloud library (pcl) and uniformly sampled to make it ready for Pointnet++. The model recognized these objects with more than 90% accuracy.
Scapula & Humerus Bone Detection in DiCOM images
A Senior Researcher @ Top NA University
Business Problem :
The researcher is working on a research project related to the shoulder bones i.e. Scapula, Humerus & Implants. The input data is the 3D DiCOM images of Shoulder i.e. the X-Ray, CT Scan files. The business need was to identify these two bones (Scapula & Humerus) from the DiCOM files accurately. The first phase is 2D object detection and improvise the solution to 3D object detection.
Technology Used :
Python, OpenCV, Tensorflow, Deep Learning using R-CNN & Mask CNN, U-Net
Solution :
The 3D DICOM images were converted into 2D DICOM images. The object labelling was a key task, some of the images had the bones appearing in the image from more than one position. So the labelled bounding boxes were overlapping, and thus detecting the overlapping object was a challenge. We evaluated the deep learning algorithms like R-CNN, Mask R-CNN & U-Net. Both Mask R-CNN & U-Net gave an object detection accuracy of over 90%.
Face Recognition & Object Detection
Product company serving Japanese market
Business Problem :
The customer wanted to automate billing in their cafeteria based on the food served. The requirement was to detect person using face recognition and food ordered using plate/bowl types.
Technology Used :
Python, OpenCV, Tensorflow, Deep Learning
Solution :
The deep learning based solution for object detection (plate/bowl) is developed using Python & Tensorflow. The deep learning model is trained on training data set provided by customer. These models are integrated with restaurant billing software of our client. Evaluated model performance on high end cpu and gpu machine on cost, speed and accuracy parameters. Achieved face recognition and object detection speed in less than 1 second with more than 90% accuracy.
Smart City – Waste Classification
A municipality of an European Country
Business Problem :
The municipality has various kind of Waste Containers which are hosted on roads and localities. The waste could be Glass, Plastic, Liquid, Food Remains, Furniture etc. People dump the waste in the Waste Bin and the bigger waste like Furniture, Mattresses etc are dumped near the Waste Container. For each type of waste, different departments of Municipality as responsible. Citizens can report about the garbage lying near the container by taking the pictures and uploading it on the Smart city app. The business problem was to detect various categories of waste from the upload image and return back the appropriate department to the App.
Technology Used :
Python, TensorFlow, OpenCV, Deep Learning, NLP
Solution :
STY developed custom algorithms to process the text & images uploaded by the citizens on the Mobile App. The algorithms were trained on the training images provided by the Customer. Based on the various type of wastes, the algorithm would return the right department. Along with image, the citizens would also write about the issue. Using NLP based Customer algorithms, we could classify the text, identify the correct department and return it to the App. The accuracy achieved was over 94% which is a great accuracy to have.
Leather Defect Detection for a Shoe Manufacturer
A Shoe Manufacturer in India
Business Problem :
A shoe manufacturer in India was looking to automate visual inspection of leather to detect ten different type of defects. The current approach is manual and error prone.
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
STY developed a solution of deep learning and computer vision to automatically detect defect on leather surface. Around 50 images were provided for each type of defects. The high resolution images were used for better model accuracy. The AI model analyses the image and detects & localize defect present on the leather surface. This solution worked at very high accuracy of more than 90% and helped to automate visual inspection process for the defect detection.
Industrial Pond Surface Abnormality Detection
A Manufacturing company in Japan
Business Problem :
A large manufacturing company in Japan is looking at automating surveillance of Pond located withing Industrial premise. They need to ensure that industrial waste or other type of waste is not present on the water surface. Currently this is inspected manually
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
STY developed a solution of deep learning and computer vision to automatically detect surface abnormality due to waste material on water surface. The surface abnormality was inspected at varying lighting conditions. The solution continuously analyses the feed coming from cctv camera’s installed around pond. The AI model analyses the image and detects & localize waste present on the water surface. This solution worked at very high accuracy of more than 90% and helped to automate surveillance of industrial pond for waste detection.
Channel & Rebar Counting Pilot project for a Steel Company
A Steel Company in North America
Business Problem :
A major steel company in Canada was looking for inexpensive solution to count channels, rebar post stacking them in the bundle. The current solutions did not give them good accuracy and hence were doing it manually
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
STY developed a solution of deep learning and computer vision to automatically count channels and rebar from the image. The end use can taken picture of the stacked rebars & channels and upload it to the cloud based platform. The AI model analyses the image and detects channels & rebars withing in the image and count them. This solution worked at very high accuracy of more than 95% and helped to uncover inconsistencies in object count between different stages of manufacturing
Product Recognition for Personalized Premium Quote
A Start up in PropTech in North America
Business Problem :
A startup based in PropeTech space wanted to build innovative mobile app where smart mobile camera to be used to recognize household items. This will enable it to list down inventory of common household objects present in the home. This will be used to prepare the personalized premium quote based on items present in the home.
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
Initially customer team tried using Aws sagemaker for detecting household objects, the solution did not scale for custom objects which were not present in the Aws Sagemaker pre-trained model. Also the lack of sufficiently large training data was another challenge. Here we have benchmarked state of art Object Detection models to identify the best fit models. There was a challenge to detect smaller household items specially present in kitchen and similar looking objects. Used more specialized model to classify similar looking items and deeper network to recognize smaller items. The solution achieved overall accuracy of greater than 85%
Brand Detection for Visual Product Search – PoC
A Retailer in Japan
Business Problem :
A retailer was looking for the solution to build smart visual product search for better in store customer experience. The customer will show the image of product from the mobile phone and the solutionwill guide user about product availability and where it is located in the store
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
The major challenge was the brand images of the product looked very similar. The standard object detection does not work where products looks very similar. Adopted the approach of fine grained image recognition where subtle difference between two brand images is recognized with much higher accuracy. In this approach, we used triplet loss to identify subtle difference between brand images. AI model was deployed as an API which was invoked from the mobile app.
Signature Verification
A Financial Institution in India
Business Problem :
A leading financial institution in India wanted to automate the process of signature verification to reduce the chance of fraud and digitize the process.
Technology Used :
Python, TensorFlow, Django, Deep Learning
Solution :
The ground truth signature is extracted from PAN or Aadhar card. The customer upload the policy application form or claim form in the web application. The signature verification tool extract the signature from the document. This signature is then compared with the ground truth signature. The robust image pre-processing pipeline was developed to handle noise and image quality in ground truth and submitted document. This enabled solution to achieve very high level of accuracy
Landmark Detection & Noise Removal
A Deep Health Tech start up In North America
Business Problem :
As the product is improved continuously, the phase 3 of the project focused on scattered noise removal from dicom data. Also there is need to correctly identify landmark on the bones. This is essential step for developing surgery simulation platform
Technology Used :
Python, OpenCV, Tensorflow, Deep Learning using 3D U-Net
Solution :
The paired CNN is used for denoising the effect of scatter noise in dicom image. We generated synthetic data, as very less data was available for training. The model is trained with a pair of denoised image and noisy image. We have got very good accuracy with medical image using this approach
Other part of this phase was to correctly detect landmarks on the bone. This is important for the surgery simulation platform. We have used heatmap regression to correctly identify landmark on the shoulder bone