Portfolio

Hand Detection

Hand tracking in this algorithm is implemented using OpenCV’s Haar cascade-based detection, where a grayscale image is processed to identify and track hand positions in real-time. The function first converts a 16-bit grayscale image into an 8-bit format, followed by a color conversion to RGB to enhance detection accuracy. A Haar cascade classifier (aGest.xml) is used to detect hand regions by scanning the grayscale version of the image. The detected bounding boxes are then refined by scaling and padding, ensuring that the hand’s position is accurately captured. The largest detected hand is selected, and its bounding box and center position are recorded.

Once a hand is detected, it is tracked across frames by updating the _trackedHands list, ensuring smooth tracking. The algorithm also estimates the distance of the hand from the camera by correlating depth values from the grayscale image, providing real-world depth estimation for hand positions. For visualization, a bounding box is drawn around the detected hand, and the estimated distance is overlaid as text. The processed image is then converted into a WriteableBitmap and displayed in the UI using Dispatcher, ensuring real-time updates. This efficient and scalable approach makes it ideal for gesture recognition, augmented reality (AR), virtual reality (VR), and human-computer interaction applications, where precise hand movement tracking is essential.

A Convolutional Neural Network Approach for Skin Cancer Detection

Abstract— Skin cancer is among the top ten most common cancers worldwide. Early detection significantly increases the likelihood of successful treatment. In this paper, we propose a CNN-based framework to detect and classify various types of skin cancer from clinical images. Our architecture consists of three convolutional blocks employing the swish activation function, followed by pooling, flattening, and fully connected layers. Experimental evaluation on a benchmark dataset demonstrates the model’s potential, achieving competitive classification accuracy across 7 skin cancer classes.

The proposed model is designed to address the inherent challenges in skin cancer detection by leveraging a deep learning architecture that efficiently extracts discriminative features from dermoscopic images. The architecture comprises three convolutional blocks followed by a dense classifier, as illustrated in Figure 1.

Each convolutional block is engineered to gradually extract low to high-level features. The first block applies 32 filters with a 2×2 kernel, utilizing the swish activation function to promote smooth gradients during training. Swish, being a non-monotonic activation function, has been shown to improve model performance and training stability compared to more traditional functions such as ReLU. Following the convolution, a 2×2 max pooling layer is employed to reduce the spatial dimensions, thereby decreasing the computational burden while retaining the most salient features.

The second block repeats this configuration with 32 filters, reinforcing the extraction of finer details from the input images. By stacking multiple convolutional layers, the network is able to learn increasingly abstract representations that capture subtle differences between various types of skin lesions.

In the third convolutional block, the number of filters is increased to 64 while maintaining the 2×2 kernel size. This block further enhances the network’s capacity to detect complex features by capturing more intricate patterns that are characteristic of malignant and benign lesions alike. The subsequent max pooling operation ensures that only the most significant features are retained for further processing.

Following the convolutional layers, the feature maps are flattened into a one-dimensional vector, serving as the input to the fully connected layers. The dense classifier consists of a 64-unit layer with swish activation, which integrates the extracted features into a high-level representation. A dropout layer with a rate of 0.5 is incorporated at this stage to mitigate overfitting by randomly deactivating neurons during training. This regularization step is particularly important given the high dimensionality of the feature space and the potential variability in clinical image datasets.

The final dense layer consists of 7 units corresponding to the 7 classes of skin lesions under study. A SoftMax activation function is applied to generate normalized class probabilities, facilitating multi-class classification by indicating the likeliho

This article presented a CNN-based approach for the automated detection of skin cancer, leveraging swish activation and a three-block convolutional architecture. The proposed model achieved a classification accuracy of 98%, demonstrating its effectiveness indistinguishing among seven different types of skin lesions. A detailed examination of the confusion matrix indicates that the model maintains high precision and recall across all classes, underscoring its robustness and potential clinical utility. Despite the promising results, future work will focus on expanding the dataset and exploring advanced regularization strategies to further enhance the model’s generalization. Additionally, integrating attention mechanisms or transfer learning approaches may offer additional gains in performance and facilitate wider adoption in clinical settings.

Depth Estimation

Depth Anything and MiDaS are both powerful monocular depth estimation models, but they cater to different needs. Depth Anything is designed for absolute depth estimation, making it suitable for real-world applications like robotics, 3D reconstruction, AR/VR, and autonomous navigation. It leverages transformer-based architecture and large-scale datasets to deliver high accuracy and generalization across various scenes. However, it requires a high-end GPU for optimal performance.

On the other hand, MiDaS focuses on relative depth estimation, excelling in image-based depth perception, SLAM, and artistic applications. It offers lighter models that can run on mid-range GPUs or even CPUs, making it a more accessible option for applications that don’t require absolute depth values. While Depth Anything provides better precision for real-world depth measurement, MiDaS is faster and more efficient for general depth perception tasks. The choice depends on absolute depth accuracy or computational efficiency.

Depth estimation in this algorithm is achieved using the Depth Anything model, which processes a 16-bit grayscale depth image and maps it to a real-world depth representation. The function first converts the raw depth data into an 8-bit grayscale format, ensuring that the values are normalized within the expected input range of the model. The grayscale image is then resized to 518×518, matching the required input dimensions for the ONNX-based Depth Anything v2 model. Before inference, the image undergoes preprocessing, where each pixel is normalized using ImageNet mean and standard deviation values for the three RGB channels. The model processes this tensor through an ONNX inference session, generating a depth map where each pixel represents an estimated distance value.

To enhance visualization, the algorithm applies min-max normalization to scale the depth values between 0 and 255, making them suitable for display. A Jet colormap is then applied, converting the grayscale depth map into a colorized representation, where warmer colors indicate closer distances and cooler colors represent farther depths. To improve depth interpretation, a depth scale bar is generated and overlaid on the image, providing reference values from minimum to maximum depth in meters. Additionally, real-time UI updates are handled using Dispatcher, ensuring smooth rendering in a graphical application. This approach allows for accurate depth estimation and visualization, making it ideal for applications such as robotics, AR/VR, 3D reconstruction, and autonomous navigation, where precise depth perception is essential.

Heat Measurement

Heat measurement in this algorithm is performed by processing a 16-bit grayscale thermal image and mapping its intensity values to real-world temperature readings. The function first converts the raw pixel values into grayscale intensity, normalizing them within a specified temperature range (e.g., 70°F to 1200°F). To enhance accuracy, it calculates the average environmental temperature and determines a dynamic threshold based on the 99th percentile of intensity values, filtering out noise. Using OpenCV, the grayscale image is converted to an 8-bit format, and a binary thresholding method is applied to identify significant heat sources. The algorithm then detects heat regions by finding contours and filtering out small areas below a minimum region size. For visualization, a colormap (Jet) is applied, where heat sources are highlighted in colors corresponding to temperature levels. Bounding boxes are drawn around detected heat regions, and the average and maximum heat values for each region are calculated and displayed. To provide a reference scale, a heat intensity bar is generated on the right side of the image, annotated with temperature labels. Finally, the processed heatmap is converted into a WriteableBitmap for display in the GUI. This approach ensures accurate heat source identification and visualization, making it useful for applications such as thermal imaging, industrial monitoring, and safety inspections.