Noise suppression algorithms have been used in various tasks such as computer vision, industrial inspection, and video surveillance, among others. The robust image processing systems need to be fed with images closer to a real scene; however, sometimes, due to external factors, the data that represent the image captured are altered, which is translated into a loss of information. In this way, there are required procedures to recover data information closest to the real scene. This research project proposes a Denoising Vanilla Autoencoding (DVA) architecture by means of unsupervised neural networks for Gaussian denoising in color and grayscale images. The methodology improves other state-of-the-art architectures by means of objective numerical results. Additionally, a validation set and a high-resolution noisy image set are used, which reveal that our proposal outperforms other types of neural networks responsible for suppressing noise in images.

We develop a swift, robust, and practical tool for detecting brain lesions with minimal user intervention to assist clinicians and researchers in the diagnosis process, radiosurgery planning, and assessment of the patient’s response to the therapy. We propose a unified gravitational fuzzy clustering-based segmentation algorithm, which integrates the Newtonian concept of gravity into fuzzy clustering. We first perform fuzzy rule-based image enhancement on our database which is comprised of T1/T2 weighted magnetic resonance (MR) and fluid-attenuated inversion recovery (FLAIR) images to facilitate a smoother segmentation. The scalar output obtained is fed into a gravitational fuzzy clustering algorithm, which separates healthy structures from the unhealthy. Finally, the lesion contour is automatically outlined through the initialization-free level set evolution method. An advantage of this lesion detection algorithm is its precision and its simultaneous use of features computed from the intensity properties of the MR scan in a cascading pattern, which makes the computation fast, robust, and self-contained. Furthermore, we validate our algorithm with large-scale experiments using clinical and synthetic brain lesion datasets. As a result, an 84%–93% overlap performance is obtained, with an emphasis on robustness with respect to different and heterogeneous types of lesion and a swift computation time.

Associative memories in min and max algebra are of great interest for pattern recognition. One property of these is that they are one-shot, that is, in an attempt they converge to the solution without having to iterate. These memories have proven to be very efficient, but they manifest some weakness with mixed noise. If an appropriate kernel is not used, that is, a subset of the pattern to be recalled that is not affected by noise, memories fail noticeably. A possible problem for building kernels with sufficient conditions, using binary and gray-scale images, is not knowing how the noise is registered in these images. A solution to this problem is presented by analyzing the behavior of the acquisition noise. What is new about this analysis is that, noise can be mapped to a distance obtained by a distance transform. Furthermore, this analysis provides the basis for a new model of min heteroassociative memory that is robust to the acquisition/mixed noise. The proposed model is novel because min associative memories are typically inoperative to mixed noise. The new model of heteroassocitative memory obtains very interesting results with this type of noise.

In the self-driving vehicles domain, steering control is a process that transforms information obtained from sensors into commands that steer the vehicle on the road and avoid obstacles. Although a greater number of sensors improves perception and increases control precision, it also increases the computational cost and the number of processes. To reduce the cost and allow data fusion and vehicle control as a single process, this research proposes a data fusion approach by formulating a neurofuzzy aggregation deep learning layer; this approach integrates aggregation using fuzzy measures $\mu$ as fuzzy synaptic weights, hidden state using the Choquet fuzzy integral, and a fuzzy backpropagation algorithm, creating a data processing from different sources. In addition, implementing a previous approach, a self-driving neural model is proposed based on the aggregation of a steering control model and another for obstacle detection. This was tested in an ROS simulation environment and in a scale prototype. Experimentation showed that the proposed approach generates an average autonomy of $95\%$ and improves driving smoothness by $9\%$ compared to other state-of-the-art methods.