Abstract
Ordinal classification tasks are present in a large number of different domains. However, common losses for deep neural networks, such as cross-entropy, do not properly weight the relative ordering between classes. For that reason, many losses have been proposed in the literature, which model the output probabilities as following a unimodal distribution. This manuscript reviews many of these losses on three different datasets and suggests a potential improvement that focuses the unimodal constraint on the neighborhood around the true class, allowing for a more flexible distribution, aptly called quasi-unimodal loss. For this purpose, two constraints are proposed: A first constraint concerns the relative order of the top-three probabilities, and a second constraint ensures that the remaining output probabilities are not higher than the top three. Therefore, gradient descent focuses on improving the decision boundary around the true class in detriment to the more distant classes. The proposed loss is found to be competitive in several cases.

Abstract
Cervical cancer is the fourth leading cause of cancer-related deaths in women, especially in low to middle-income countries. Despite the outburst of recent scientific advances, there is no totally effective treatment, especially when diagnosed in an advanced stage. Screening tests, such as cytology or colposcopy, have been responsible for a substantial decrease in cervical cancer deaths. Cervical cancer automatic screening via Pap smear is a highly valuable cell imaging-based detection tool, where cells must be classified as being within one of a multitude of ordinal classes, ranging from abnormal to normal. Current approaches to ordinal inference for neural networks are found to not sufficiently take advantage of the ordinal problem or to be too uncompromising. A non-parametric ordinal loss for neuronal networks is proposed that promotes the output probabilities to follow a unimodal distribution. This is done by imposing a set of different constraints over all pairs of consecutive labels which allows for a more flexible decision boundary relative to approaches from the literature. Our proposed loss is contrasted against other methods from the literature by using a plethora of deep architectures. A first conclusion is the benefit of using non-parametric ordinal losses against parametric losses in cervical cancer risk prediction. Additionally, the proposed loss is found to be the top-performer in several cases. The best performing model scores an accuracy of 75.6% for seven classes and 81.3% for four classes.

Abstract
Convolutional neural networks are shown to be vulnerable to changes in the background. The proposed method is an end-to-end method that augments the training set by introducing new backgrounds during the training process. These backgrounds are created by a generative network that is trained as an adversary to the model. A case study is explored based on overhead power line insulators detection using a drone - a training set is prepared from photographs taken inside a laboratory and then evaluated using photographs that are harder to collect from outside the laboratory. The proposed method improves performance by over 20% for this case study.

Abstract

Abstract
Overhead Power Distribution Lines (OPDLs) correspond to a large percentage of the medium-voltage electrical systems. In these networks, visual inspection activities are usually performed without resorting to automated systems, requiring a significant investment of time and human resources. We present a methodology to identify the defect and type of insulators using Convolutional Neural Networks (CNNs). More than 2500 photographs were collected both from inside a studio and from a realistic OPDL. A classification model is proposed to automatically recognize the insulators conformity. This model is able to learn from indoors photographs by augmenting these images with realistic details such as top ties and real-world backgrounds. Furthermore, Multi-Task Learning (MTL) was used to improve performance of defect detection by also predicting the insulator class. The proposed methodology is able to achieve an accuracy of 92% for material classification and 85% for defect detection, with F1-score of 0.75, surpassing available solutions.