Deep Learning Projects

With the advancement new technologies in Artificial intelligence the scope of Deep Learning is playing an important role in solving multiple problems and is being applied in various domains. This technology aims to imitate the biological neural network, that is, of the human brain. For Engineering students, the Practical approach will play a more vital role than theoretical knowledge ,we have an extensive list of multiple projects which cover all application areas in a real time scenario. Let us explore some real-time project ideas with very interesting deep learning project ideas which beginners and professional can work on to put their knowledge to the test and further help to get good hands-on experience in deep learning.

01. Person age & gender recognition using deep learning techniques.

Today’s smartphone cameras have incorporate AI. They are even capable of determining a person's gender and age. Deep learning can be used for this, but we'll need a lot of data to build the model for detection of age and gender.

02. Slumped Driving Detection of a driver using deep learning model.

Napping of a driver can be determined using driver drowsiness model built by deep learning technique. This model when put into practise for drivers will help to avoid accidents.

03. Pose estimation for people

The art of determining a person's body alignment by calculating various body joints is known as human pose estimate. In order to apply a filter to a person, Snapchat uses posture estimation to determine where the individual's head and eyes are situated to fix a filter. Similar to this, we can determine a person's stance and apply filters to them in real time.

04. computer vision technique for viability of quantifying the treatment for stroke patients a deep neural network approach

Training in limb rehabilitation speeds up the recovery process and enhances quality of life for stroke patients with hemi paralysis. Both doctors and patients must be aware of the patient's progress in rehabilitation. The computer vision method, which can identify a patient's training action, movement trajectory, and activity status, can be used to monitor rehabilitation more precisely and effectively than wearable sensors or deep cameras. In the clinic, it is difficult to quantify the dynamic change of different training sessions to assess the progress of the rehabilitation, with the exception of static measures of real-time behaviour.

In this study, we suggested a computational method to compare the upper limbs' motion change. The upper limb joint points were first identified using OpenPose to preprocess the video data, and the positions of each joint point were then specified using Cartesian coordinates. Second, in order to determine the rehabilitation progress, we computed the similarity of the limb's lift angle and time in various training periods using the dynamic time warping algorithm. The outcomes demonstrate that our approach can measure data and analyse the effectiveness of rehabilitative actions using a basic monocular camera, which has a promising future for use.

05. Yolo V4 & Yolo V5 approach for recognition of bird and drone.

Due to their popularity and rising prevalence, it is now difficult to spot drones, so we built this project to identify them. In addition to their beneficial uses, their potential for use in malevolent actions has given rise to concerns about physical infrastructure security, safety, and surveillance at airports. There have been numerous complaints in recent months of different types of drones being used improperly at airports, which has disrupted airline operations and left some people unable to locate the drones or birds.

To effectively recognise and identify two different types of drones and birds, we implement this project utilising a deep learning-based method. Comparing the suggested method to existing detection systems, evaluation of the proposed methodology using the generated image dataset shows greater efficiency. Additionally, because to their similar physical characteristics and behaviours, drones and birds are frequently mistaken for one another. The suggested method is not only capable of determining if drones are present or absent in a given area, but it can also identify worrisome and differentiate between two different types of drones and tell them apart from birds.

06. RESNET50, GoogleNet, Squeezenet, and AlexNet for the diagnosis of grape fruit illness.

Due to the constantly changing climatic and environmental factors, illnesses are relatively common in crops. Crop diseases are typically difficult to control and have an impact on the growth and output of the crops. Accurate illness detection and timely disease control measures are essential for ensuring high production and good quality. The widely cultivated grape plant in India is susceptible to various diseases that can harm the fruit, stem, and leaves. Fruit diseases that are the early signs of sour rot, powdery mildew, healthy grapes, grey mould, and healthy grapes. Therefore, it is necessary to create an automatic system that can be utilised to identify the different types of diseases and offer relevant treatments.

07. Classification of Brain Tumors Using a Highly Accurate Attention-Based Convolutional Neural Network

One of the most frequent tumours that poses a threat to human life safety has always been brain tumours. Currently, China still has a dearth of computer-aided diagnostic tools designed particularly to identify particular brain tumour situations and conduct relevant research. Brain magnetic resonance imaging (MRI) datasets that are publically accessible were gathered for this work, and data preparation like normalisation was applied to them.

This study provided a method of merging attention mechanism with Convolutional Neural Network (CNN) to lessen the influence caused by irrelevant background information features in images, taking into account the peculiarities of medical image complexity of brain MRI. The experiment's findings using the suggested methodology were contrasted with self-described traditional models like VGGNet and MobileNet.

Testing on the dataset reveals that the CNN model performs significantly better on the brain tumour recognition classification task than the other three models in terms of accuracy, indicating that the attention mechanism in the model can somewhat mitigate the impact of context irrelevant information on the classification outcome. The online website where the trained analytical model is deployed in this work also has an easy-to-use interface that is practical for medical staff to use.

08. System for Detecting Sitting Posture Based on the Keras Framework

A real-time sitting posture detection system based on deep learning is built with a low-power embedded real-time sitting posture detection system in mind. A thin-film pressure sensor is used in the system to measure the human body's sitting posture pressure. The system then collects and analyses data on human body pressure in various sitting positions and creates an analytical model using the Keras framework.

Burn the model onto the STM32 using cubemax to provide real-time data collecting, processing, and sitting position identification. Finally, the MQTT protocol, which enables the real-time detection and classification of the sitting position and provides the pertinent sitting posture correction cues, accomplishes communication between the STM32 and the Android application.

09. SE-ResNeXt50 model to identify and control strawberry diseases and pests.

Since there aren't many high-quality open image datasets available at this time, studies on strawberry disease and pest diagnosis and control are rare. In light of this, initially, online and offline independent collections of 13 types of frequent strawberry pests and diseases were made in order to create databases. Second, the SE-ResNeXt50 model was developed, which was more practical than the ResNet50 residual network model.

To be more precise, the Inception model was combined with the ResNet50 model to widen the network, 32 branches were set, and the attention mechanism, the squeeze and excitation module (SE), was also imported. This solution to the complex image background and information interference problems improved the model's identification efficiency and accuracy. According to the findings, the SE-ResNeXt50 model's accuracy, which reached 89.3%, was 8% greater than the ResNet50 model's. After 15 iterations, the SEResNeXt50 model reached a plateau, demonstrating its successful identification.

Additionally, the SEResNeXt50 model, which was created using data collected in the actual world, has high resilience and generalisation capabilities, better addressing the needs of strawberry growers. Based on the SE-ResNeXt50 model, a WeChat mini-program for strawberry disease and pest identification was created, allowing fruit growers to quickly identify strawberry pests and illnesses and receive prevention recommendations, supporting the growth of the strawberry sector.

10. Technology for detecting IoT intrusions based on deep learning

A network intrusion detection model including convolutional neural networks and gated recurrent units is presented to solve the issue of low accuracy of existing network intrusion detection models for multi-classification of intrusion behaviours and redundancy of data characteristics. The random forest algorithm and Pearson correlation analysis are used to solve the issue of feature redundancy.

Next, the temporal features of the data are extracted using TCN and GRU, while attention module is introduced to assign different weights to the features, thereby reducing overhead and enhancing model performance. Finally, the classification problem is solved using the Softmax function. In this study, a model with a 99.99% accuracy is evaluated on the Bot-Lot dataset.

11. Enhanced particle swarm optimization technique and neural network for human fall detection.

Fall detection has become a widespread worry in the area of public safety as the world's population ages. Elderly fall injuries can be effectively reduced by quickly and precisely identifying falling behaviours in surveillance cameras and promptly dispatching rescue signals.

For real-time fall detection in indoor situations, this research suggests a hybrid approach based on an enhanced particle swarm optimization technique and a neural network. Alpha pose is used to extract human key points from video frames, and an upgraded particle swarm optimization neural network model is used to classify the human keypoints in real time. According to experimental findings, this technique may accurately identify falling actions in interior scenes.

12. Deep Convolutional Generative Adversarial Networks to Create Human Faces.

They are incredibly potent methods for producing sights, sounds, texts, or films that are identical to actual data. The goal of this research is to create realistic-looking human faces from scratch by starting with random noise and using Deep Convolutional Generative Adversarial Networks.

13. Differential Autoencoders

The potential for VAE, or variational autoencoder, in deep learning is enormous. They are able to create fresh data that is comparable to the training data. They come equipped with both an encoder and a decoder. You can start by generating numbers using the MNIST dataset. Variational Autoencoders

14. Photos colouring of old and black and white images using Deep CNN

The goal of this project is to create a model that can convert vintage black-and-white photos into colourful ones. While it used to take digital painters many hours to colour a picture, deep learning today makes it feasible to colour an image in only a few seconds.

15. Translator of languages

While it takes humans almost a year to learn a language, computers can do it in just a day. We can create a language translator app for this project that can translate between English and to other languages including regional languages.

16. Cucumber Disease detection in Image Recognition Using a New Attention Module

Deep convolutional neural networks are a significant advancement in the world of computers, but they need a lot of labelled data for model training. In the agricultural industry, it is challenging to create a sizable imaging record of crop disease. In this study, we offer a channel and position block-based solution to this issue (called a CPB module). The CPB module can infer attention maps from an intermediate feature map in addition to the channel and position.

The input feature maps for adaptive feature refinement can then be created by multiplying the attention maps. Using our own crop disease image dataset, we conducted rigorous tests to verify the CPB module's efficacy. Class activation map (CAM) display characteristics were used in the trials. The outcomes demonstrated that the CPB module is capable of accurately classifying images of crop diseases.

17. NB-IoT technology for water quality monitoring system

The STM32F103 microcontroller and Cloud IoT platform, based on narrowband Internet of Things technology, are used in this research to design a water quality monitoring system that addresses the issues of complexity and low efficiency of conventional water quality monitoring methods.

This device can continuously gather information on the target's automatically monitored waters' temperature, pH, TDS, and ORP levels. Real-time data can be automatically uploaded to the cloud platform, and users can access real-time monitoring data on the CLOUD IoT platform. This innovation and application value overcomes the drawbacks of conventional water quality monitoring systems, such as lengthy data collection cycles and subpar real-time performance.

18. Facial appearance method for recognising personalities

In this study, we offer an automatic personality assessment approach using portrait photos that is motivated by psychology research that shows that a person's facial appearance can be seen as the important clue in judging their personality. Crowd-sourced data was used to create the CelebAMask-HQ-Personality face-personality dataset. This dataset includes of annotated personality traits, facial geometric attributes, and facial photos.

By removing hidden ones, the CelebAMask-HQ facial photos are improved. Under the direction of psychological research, we build geometric characteristics using face landmark information. Volunteers use the Ten-Item Personality Inventory to collect the necessary number of personality annotations (TIPI). Then, utilising a few sample classifiers, such as multi-layer perceptrons, random forests, support vector machines, and k-nearest neighbours, etc., the relationships between face features and personality are investigated. According to the results of experiments, it is possible to determine a person's personality from their facial features.