Context awareness based Sketch-DeepNet architecture for hand-drawn sketches classification and recognition in AIoT

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human sketch recognition is efficient but challenging for computers due to low detail.
• Existing methods like SIFT and BoW are complex and time-consuming.
• Deep neural networks (DNNs) struggle with sketch data due to its nature, requiring specialized approaches.

Purpose of the Study:

• To develop an effective deep learning model for accurate sketch recognition.
• To address the limitations of current computer vision models in understanding hand-drawn sketches.
• To improve the performance of sketch classification systems for AIoT applications.

Main Methods:

• Proposed a novel convolutional neural network (CNN) architecture named Sketch-DeepNet.
• Utilized the TU-Berlin dataset for training and evaluating the sketch classification model.
• Compared Sketch-DeepNet's performance against established sketch recognition methods and human accuracy.

Main Results:

• Sketch-DeepNet achieved a classification accuracy of 95.05% on the TU-Berlin dataset.
• The proposed model significantly outperformed existing methods, including DeformNet, Sketch-DNN, Sketch-a-Net, SketchNet, Thinning-DNN, CNN-PCA-SVM, and Hybrid-CNN.
• Sketch-DeepNet surpassed human recognition accuracy (73%) on the same dataset.

Conclusions:

• The developed Sketch-DeepNet architecture demonstrates superior performance in sketch recognition tasks.
• This advancement is crucial for developing robust artificial intelligence of things (AIoT) systems.
• The findings highlight the potential of specialized CNNs for interpreting low-detail graphical representations.