Deepak Singh

Delineated farm boundary on planetscope satellite images using Deep Learning. Performed data cleaning, preparation, normalization of Satellite images for feature engineering. Applied Semantic Segmentation and achieved 94.9% accuracy using DenseNet (Keras). Implemented Object-based Crop classification on Sentinel satellite images. Analyzed and planned ground point label collection using historical data and spatial distribution. Executed data check, verification and checked class imbalance using SMOTE Technique. Random Point generation in each farm and classified 10 crops using SVM with 82.8% accuracy. Video Analytics (Automatic Number Plate Recognition System and Face Recognition System). Trained YOLO V5 for Number Plate Detection and Paddle OCR for Recognition over a custom dataset. Implemented face recognition system to identify/track person using Openface algorithm. Developed Crop health pattern identification on Sentinel satellite images using Google Earth Engine API. Selected cloud free data and calculated time series NDVI images using Google Earth Engine API in python. Overlayed stacked NDVI images on farm boundaries vector layers and extracted NDVI values. Used linear regression on temporal series NDVI and predicted crop health with accuracy of 86.7%. Designed and implemented near optimal algorithm for NP-Hard problem (i.e Crew Scheduling) for NMMT (Navi Mumbai Municipal Transport) and AICTSL (Atal Indore City Transport Service Ltd).

Double JPEG Compression Detection: Classified JPEG image for forgery and Localize the tempered region in the forged image. Trained ResNet, AlexNet, InceptionNet and DenseNet models with 94.6% average accuracy. Audio Source Identification: Three audio (.wav) files were recorded using 19 smartphones. Transformed raw audio data into DFT domain we trained Modified DenseNet and our classifier was able to classify 1-second audio with 97.1% accuracy.