Search

With the promotion of artificial intelligence in agriculture and the popularization of plug tray seedling-raising technology, seedling raising and transplanting have become the most popular planting modes. Miss-seeding is one of the most serious problems affecting seedling raising and transplanting. It not only affects the germination rate of seeds but also reduces the utilization rate of the plug tray. The experimental analysis of traditional machine vision-based miss-seeding showed that because of uneven lighting, the plug tray was wrongly identified as a seed under bright light, but the seeds in the dark were not easy to identify. When using the seeding area to identify seeds and noise, sweet corn seeds in a small area can be easily screened out. This paper proposes a method using the ResNet network with an attention mechanism to solve the above-mentioned problems. In this paper, the captured image was segmented into the images of a single plug tray, and a residual attention network was built; the detection scheme of miss-seeding was also converted into a dichotomous picture recognition task. This paper demonstrates that the residual attention network can effectively recognize and detect the seed images of sweet corn with very high accuracy. The results of the experiment showed that the average accuracy of this recognition model was 98%. The feature visualization method was used to analyze the features, further proving the effectiveness of the classification method of plug tray seedlings.

Military object detection from Unmanned Aerial Vehicle (UAV) reconnaissance images faces challenges, including lack of image data, images with poor quality, and small objects. In this work, we simulate UAV low-altitude reconnaissance and construct the UAV reconnaissance image tank database UAVT-3. Then, we improve YOLOv5 and propose UAVT-YOLOv5 for object detection of UAV images. First, data augmentation of blurred images is introduced to improve the accuracy of fog and motion-blurred images. Secondly, a large-scale feature map together with multi-scale feedback is added to improve the recognition ability of small objects. Thirdly, we optimize the loss function by increasing the loss penalty of small objects and classes with fewer samples. Finally, the anchor boxes are optimized by clustering the ground truth object box of UAVT-3. The feature visualization technique Class Action Mapping (CAM) is introduced to explore the mechanisms of the proposed model. The experimental results of the improved model evaluated on UAVT-3 show that the mAP reaches 99.2%, an increase of 2.1% compared with YOLOv5, the detection speed is 40 frames per second, and data augmentation of blurred images yields an mAP increase of 20.4% and 26.6% for fog and motion blur images detection. The class action maps show the discriminant region of the tanks is the turret for UAVT-YOLOv5.

Nowadays, most deep-learning-based bearing fault diagnosis methods are studied under the condition of steady speed, while the performance of these models cannot be fully played under time-varying conditions. Therefore, in order to facilitate the practical application of a deep learning model in bearing fault diagnosis, a vibration–speed fusion network is proposed, which utilizes a transformer with a self-attention module to extract vibration features and utilizes a sparse autoencoder (SAE) network to extract sparse features from speed pulse signal. The vibration–speed fusion network enables the efficient fusion of different signals in a high-dimensional vector space with a high degree of model interpretability, without additional signal processing steps. After tuning the hyperparameters of the network, the key segments of the bearing’s time-domain vibration signals can be optimally extracted, the network performance is much better than traditional deep learning methods, and the classification accuracy can reach 95.18% and 99.85% on the two public bearing datasets from the Xi’an Jiaotong University and the University of Ottawa.

Konvolucijske neuronske mreže su poznate po njihovim nevjerojatnim postignućima u području računalnog vida, ali kao modeli crne kutije pate od manjka transparentnosti. Postoje metode objašnjivosti koji pokušavaju dati uvide u metodologiju koju mreže koriste kada stvaraju zaključke. Ove metode su veoma subjektivne, i ne postoji objektivni, opće prihvačeni način za mjerenje njihove korisnosti. Metode su detaljno objašnjene, s opisima o napretku koji su ostvarili kroz vrijeme. Metode su zatim korištene i uspoređene na raznim skupinama podataka, s naglaskom na njihovu primjenjivost, pronicljivost, i lakoću korištenja. Convolutional neural networks are known for being capable of increadible feats in computer vision, but as a black box model they suffer from a lack of transparency. There are explanatory methods that attempt to give insights into the inference methodology of these networks. These methods are highly subjective in nature, and there is no objective, universally agreed way to measure their usefulness. These methods are described in detail, with descriptions on the progress that has been made on their improvement over time. The methods are then used and compared on various datasets, with attention given to their applicability, insightfulness, and ease of use.

Invasive pest fall armyworm (FAW) Spodoptera frugiperda is one of the serious threats to the food safety. Early warning and control plays a key role in FAW management. Nowadays, deep learning technology has been applied to recognize the image of FAW. However, there is a serious lack of training dataset in the current researches, which may mislead the model to learn features unrelated to the key visual characteristics (ring pattern, reniform pattern, etc.) of FAW adults and its close related species. Therefore, this research established a database of 10,177 images belonging to 7 species of noctuid adults, including FAW and 6 FAW close related species. Based on the small-scale dataset, transfer learning was used to build the recognition model of FAW adults by employing three deep learning models (VGG-16, ResNet-50 and DenseNet-121) pretrained on ImageNet. All of the models got more than 98% recognition accuracy on the same testing dataset. Moreover, by using feature visualization techniques, this research visualized the features learned by deep learning models and compared them to the related key visual characteristics recognized by human experts. The results showed that there was a high consistency between the two counterparts, i.e., the average feature recognition rate of ResNet-50 and DenseNet-121 was around 85%, which further demonstrated that it was possible to use the deep learning technology for the real-time monitoring of FAW adults. In addition, this study also found that the learning abilities of key visual characteristics among different models were different even though they have similar recognition accuracy. Herein, we suggest that when evaluating the model capacity, we should not only focus on the recognition rate, the ability of learning individual visual characteristics should be allocated importance for evaluating the model performance. For those important taxonomical traits, if the visualization results indicated that the model didn't learnt them, we should then modify our datasets or adjusting the training strategies to increase the learning ability. In conclusion, this study verified that visualizing the features learnt by the model is a good way to evaluate the learning ability of deep learning models, and to provide a possible way for other researchers in the field who want to understand the features learnt by deep learning models.

This paper employs state-of-the-art Deep Convolutional Neural Networks (CNNs), namely AlexNet, VGGNet, Inception, ResNet and ResNeXt in a first experimental study of ear recognition on the unconstrained EarVN1.0 dataset. As the dataset size is still insufficient to train deep CNNs from scratch, we utilize transfer learning and propose different domain adaptation strategies. The experiments show that our networks, which are fine-tuned using custom-sized inputs determined specifically for each CNN architecture, obtain state-of-the-art recognition performance where a single ResNeXt101 model achieves a rank-1 recognition accuracy of 93.45%. Moreover, we achieve the best rank-1 recognition accuracy of 95.85% using an ensemble of fine-tuned ResNeXt101 models. In order to explain the performance differences between models and make our results more interpretable, we employ the t-SNE algorithm to explore and visualize the learned features. Feature visualizations show well-separated clusters representing ear images of the different subjects. This indicates that discriminative and ear-specific features are learned when applying our proposed learning strategies.

Steel defect detection is used to detect defects on the surface of the steel and to improve the quality of the steel surface. However, traditional image detection algorithms cannot meet the detection requirements because of small defect features and low contrast between background and features about steel surface defect datasets. A novel recognition algorithm for steel surface defects based on improved deep learning network models using feature visualization and quality evaluation is proposed in this paper. Firstly, the VGG19 is used to pre-train the steel surface defect classification task and the corresponding DVGG19 is established to extract the feature images in different layers from defects weight model. Secondly, the SSIM and decision tree are used to evaluate the feature image quality and adjust the parameters and structure of VGG19. On this basis, a new VSD network is obtained and used for the classification of steel surface defects. Comparing with ResNet and VGG19 methods, experiment results show that the proposed method markedly can improve the average accuracy of classification, and the model is able to converge quickly, which can be good for steel surface defect recognition using VSD network model of feature visualization and quality evaluation.

It is very important and necessary to diagnose bearing faults timely, quickly, and accurately in practical applications, because the operation status of the bearings is directly related to the performance and reliability of the rotating machinery. Therefore, a generic intelligent bearing fault diagnosis system based on AlexNet with transfer learning was proposed to automatically identify and classify different bearing faults. Transfer learning was used to avoid overfitting problem of deep network. Five bearing faults at two different motor loads and speeds were selected from the Case Western Reserve University (CWRU) bearing dataset to validate the performance of the proposed method. Results showed that compared to previous methods, the proposed method achieved excellent performance, with overall classification accuracy over 99.7%, and fast training and testing times. Feature visualization displayed the common and high-level features of spectrograms of vibration signals learned by the trained classification model. And strongest activations demonstrated the classification model had learned the correct features of each bearing fault. Importantly, there were two testing datasets employed in this study, where the training dataset and testing dataset (2) were completely independent and the number of sub-spectrograms used for testing was 3-5 times greater than those used for training. Thus, all the results suggest that the proposed method is stable, reliable, suitable for diagnosing different bearing faults and has a great potential in practical applications.

We investigated whether the use of technological tools can effectively help in manipulating the increasing volume of audio data available through the use of long field recordings. We also explored whether we can address, by using these recordings and tools, audio data analysis, feature extraction and determine predominant patterns in the data. Similarly, we explored whether we can visualize feature clusters in the data and automatically detect sonic events. Our focus was primarily on enhancing the importance of natural-urban hybrid habitats within cities, which benefit communities in various ways, specifically through the natural soundscapes of these habitats that evoke memories and reinforce a sense of belonging for inhabitants. The loss of sonic heritage can be a precursor to the extinction of biodiversity within these habitats. By quantifying changes in the soundscape of these habitats over long periods of time, we can collect relevant information linked to this eventual loss. In this respect, we developed two approaches. The first was the comparison among habitats that progressively changed from natural to urban. The second was the optimization of the field recordings’ labeling process. This was performed with labels corresponding to the annotations of classes of sonic events and their respective start and end times, including events temporarily superimposed on one another. We compared three habitats over time by using their sonic characteristics collected in field conditions. Comparisons of sonic similarity or dissimilarity among patches were made based on the Jaccard coefficient and uniform manifold approximation and projection (UMAP). Our SEDnet model achieves a F1-score of 0.79 with error rate 0.377 and with the area under PSD-ROC curve of 71.0. In terms of computational efficiency, the model is able to detect sound events from an audio file in a time of 14.49 s. With these results, we confirm the usefulness of the methods used in this work for the process of labeling field recordings.

The fingerprint is an important biological feature of the human body, which contains abundant biometric information. At present, the academic exploration of fingerprint gender characteristics is generally at the level of understanding, and the standardization research is quite limited. A robust approach is presented in this article, Dense Dilated Convolution ResNet Autoencoder, to extract valid gender information from fingerprints. By replacing the normal convolution operations with the atrous convolution in the backbone, prior knowledge is provided to keep the edge details, and the global reception field can be extended. The results were explored from three aspects: (1) Efficiency of DDC-ResNet. We conducted experiments using a combination of 6 typical automatic feature extractors with 9 classifiers for a total of 54 combinations are evaluated in our dataset; the experimental results show that the combination of methods we used achieved an average accuracy of 96.5%, with a classification accuracy of 97.52% for males and 95.48% for females, which outperformed the other experimental combinations. (2) The effect of the finger. The results showed that the right ring finger was the most effective for finger classification by gender. (3) The effect of specific features. We used the Class Activating Mapping method to plot fingerprint concentration thermograms, which allowed us to infer that fingerprint epidermal texture features are related to gender. The results demonstrated that autoencoder networks are a powerful method for extracting gender-specific features to help hide the privacy information of the user’s gender contained in the fingerprint. Our experiments also identified three levels of features in fingerprints that are important for gender differentiation, including loops and whorls shape, bifurcations shape, and line shapes.

Deep neural networks have become a standard framework for image analytics. Besides the traditional applications, such as object classification and detection, the latest studies have started to expand the scope of the applications to include artworks. However, popular art forms, such as comics, have been ignored in this trend. This study investigates visual features for comic classification using deep neural networks. An effective input format for comic classification is first defined, and a convolutional neural network is used to classify comic images into eight different artist categories. Using a publicly available dataset, the trained model obtains a mean F1 score of 84% for the classification. A feature visualization technique is also applied to the trained classifier, to verify the internal visual characteristics that succeed in classification. The experimental result shows that the visualized features are significantly different from those of general object classification. This work represents one of the first attempts to examine the visual characteristics of comics using feature visualization, in terms of comic author classification with deep neural networks.

Finding depth information with stereo matching using a deep learning algorithm for embedded systems has recently gained significant attention owing to emerging high-performance mobile graphics processing units (GPUs). Several researchers have proposed feasible small-scale CNNs that can run on a local GPU, but they still suffer from low accuracy and/or high computational requirements. In the method proposed in this study, pooling layers with padding and an asymmetric convolution filter are used to reduce computational costs and simultaneously maintain the accuracy of disparity. The patch size and number of layers are adjusted by analyzing the feature and activation maps. The proposed method forms a small-scale network algorithm suitable for a vision system at the edge and still exhibits high-disparity accuracy and low computational loads as compared to existing stereo-matching networks.

Deep neural networks (DNNs) have attained human-level performance on dozens of challenging tasks via an end-to-end deep learning strategy. Deep learning allows data representations that have multiple levels of abstraction; however, it does not explicitly provide any insights into the internal operations of DNNs. Deep learning's success is appealing to neuroscientists not only as a method for applying DNNs to model biological neural systems but also as a means of adopting concepts and methods from cognitive neuroscience to understand the internal representations of DNNs. Although general deep learning frameworks, such as PyTorch and TensorFlow, could be used to allow such cross-disciplinary investigations, the use of these frameworks typically requires high-level programming expertise and comprehensive mathematical knowledge. A toolbox specifically designed as a mechanism for cognitive neuroscientists to map both DNNs and brains is urgently needed. Here, we present DNNBrain, a Python-based toolbox designed for exploring the internal representations of DNNs as well as brains. Through the integration of DNN software packages and well-established brain imaging tools, DNNBrain provides application programming and command line interfaces for a variety of research scenarios. These include extracting DNN activation, probing and visualizing DNN representations, and mapping DNN representations onto the brain. We expect that our toolbox will accelerate scientific research by both applying DNNs to model biological neural systems and utilizing paradigms of cognitive neuroscience to unveil the black box of DNNs.

This paper explores how well deep learning models trained on chest CT images can diagnose COVID-19 infected people in a fast and automated process. To this end, we adopted advanced deep network architectures and proposed a transfer learning strategy using custom-sized input tailored for each deep architecture to achieve the best performance. We conducted extensive sets of experiments on two CT image datasets, namely, the SARS-CoV-2 CT-scan and the COVID19-CT. The results show superior performances for our models compared with previous studies. Our best models achieved average accuracy, precision, sensitivity, specificity, and F1-score values of 99.4%, 99.6%, 99.8%, 99.6%, and 99.4% on the SARS-CoV-2 dataset, and 92.9%, 91.3%, 93.7%, 92.2%, and 92.5% on the COVID19-CT dataset, respectively. For better interpretability of the results, we applied visualization techniques to provide visual explanations for the models&rsquo, predictions. Feature visualizations of the learned features show well-separated clusters representing CT images of COVID-19 and non-COVID-19 cases. Moreover, the visualizations indicate that our models are not only capable of identifying COVID-19 cases but also provide accurate localization of the COVID-19-associated regions, as indicated by well-trained radiologists.

The broad availability of smartphones and Inertial Measurement Units in particular brings them into the focus of recent research. Inertial Measurement Unit data is used for a variety of tasks. One important task is the classification of the mode of transportation. In the first step, we present a deep-learning-based algorithm that combines long-short-term-memory (LSTM) layer and convolutional layer to classify eight different modes of transportation on the Sussex&ndash, Huawei Locomotion-Transportation (SHL) dataset. The inputs of our model are the accelerometer, gyroscope, linear acceleration, magnetometer, gravity and pressure values as well as the orientation information. In the second step, we analyze the contribution of each sensor modality to the classification score and to the different modes of transportation. For this analysis, we subtract the baseline confusion matrix from a confusion matrix of a network trained with a left-out sensor modality (difference confusion matrix) and we visualize the low-level features from the LSTM layers. This approach provides useful insights into the properties of the deep-learning algorithm and indicates the presence of redundant sensor modalities.

The identification work based on inertial data is not limited by space, and has high flexibility and concealment. Previous research has shown that inertial data contains information related to behavior categories. This article discusses whether inertial data contains information related to human identity. The classification experiment, based on the neural network feature fitting function, achieves 98.17% accuracy on the test set, confirming that the inertial data can be used for human identification. The accuracy of the classification method without feature extraction on the test set is only 63.84%, which further indicates the need for extracting features related to human identity from the changes in inertial data. In addition, the research on classification accuracy based on statistical features discusses the effect of different feature extraction functions on the results. The article also discusses the dimensionality reduction processing and visualization results of the collected data and the extracted features, which helps to intuitively assess the existence of features and the quality of different feature extraction effects.

Ear recognition is an active research area in the biometrics community with the ultimate goal to recognize individuals effectively from ear images. Traditional ear recognition methods based on handcrafted features and conventional machine learning classifiers were the prominent techniques during the last two decades. Arguably, feature extraction is the crucial phase for the success of these methods due to the difficulty in designing robust features to cope with the variations in the given images. Currently, ear recognition research is shifting towards features extracted by Convolutional Neural Networks (CNNs), which have the ability to learn more specific features robust to the wide image variations and achieving state-of-the-art recognition performance. This paper presents and compares ear recognition models built with handcrafted and CNN features. First, we experiment with seven top performing handcrafted descriptors to extract the discriminating ear image features and then train Support Vector Machines (SVMs) on the extracted features to learn a suitable model. Second, we introduce four CNN based models using a variant of the AlexNet architecture. The experimental results on three ear datasets show the superior performance of the CNN based models by 22%. To further substantiate the comparison, we perform visualization of the handcrafted and CNN features using the t-distributed Stochastic Neighboring Embedding (t-SNE) visualization technique and the characteristics of features are discussed. Moreover, we conduct experiments to investigate the symmetry of the left and right ears and the obtained results on two datasets indicate the existence of a high degree of symmetry between the ears, while a fair degree of asymmetry also exists.

The recognition performance of visual recognition systems is highly dependent on extracting and representing the discriminative characteristics of image data. Convolutional neural networks (CNNs) have shown unprecedented success in a variety of visual recognition tasks due to their capability to provide in-depth representations exploiting visual image features of appearance, color, and texture. This paper presents a novel system for ear recognition based on ensembles of deep CNN-based models and more specifically the Visual Geometry Group (VGG)-like network architectures for extracting discriminative deep features from ear images. We began by training different networks of increasing depth on ear images with random weight initialization. Then, we examined pretrained models as feature extractors as well as fine-tuning them on ear images. After that, we built ensembles of the best models to further improve the recognition performance. We evaluated the proposed ensembles through identification experiments using ear images acquired under controlled and uncontrolled conditions from mathematical analysis of images (AMI), AMI cropped (AMIC) (introduced here), and West Pomeranian University of Technology (WPUT) ear datasets. The experimental results indicate that our ensembles of models yield the best performance with significant improvements over the recently published results. Moreover, we provide visual explanations of the learned features by highlighting the relevant image regions utilized by the models for making decisions or predictions.

Recognition of physical and mental responses to stress is important for stress assessment and management as its negative effects in health can be prevented or reduced. Wearable technology, mainly using electroencephalogram (EEG), provides information such as tracking fitness activity, disease detection, and monitoring neurologicalstates of individuals. However, the recording of EEG signals from a wearable device is inconvenient, expensive, and uncomfortable during normal daily activities. This study introduces the application of tensor decomposition of non-EEG data for visualizing and classifying neurological statuses with application to human stress recognition. The multimodal dataset of non-EEG physiological signals publicly available from the PhysioNet database was used for testing the proposed method. To visualize the biosignals in a low dimensional feature space, the multi-way factorization technique known as the PARAFAC was applied for feature extraction. Results show visualizations that well separate the four groups of neurological statuses obtained from twenty healthy subjects. The extracted features were then used for pattern classification. Two statistical classifiers, which are the multinomial logit regression(MLR) and linear discriminant analysis (LDA), were implemented. The results show that the MLR and LDA can identify the four neurological statuses with accuracies of 95% and 98.8%, respectively. This study suggests the potential application of tensor decomposition for the analysis of physiological measurements collected from multiple sensors. Moreover, the proposed study contributes to the advancement of wearable technology for human stress monitoring. With tensor decomposition of complex multi-sensor or multi-channel data, simple classification techniques can be employed to achieve similar results obtained using sophisticated machine-learning techniques.

Cardiovascular disease is a serious life-threatening disease. It can occur suddenly and progresses rapidly. Finding the right disease features in the early stage is important to decrease the number of deaths and to make sure that the patient can fully recover. Though there are several methods of examination, describing heart activities in signal form is the most cost-effective way. In this case, ECG is the best choice because it can record heart activity in signal form and it is safer, faster and more convenient than other methods of examination. However, there are still problems involved in the ECG. For example, not all the ECG features are clear and easily understood. In addition, the frequency features are not present in the traditional ECG. To solve these problems, the project uses the optimized CWT algorithm to transform data from the time domain into the time-frequency domain. The result is evaluated by three data mining algorithms with different mechanisms. The evaluation proves that the features in the ECG are successfully extracted and important diagnostic information in the ECG is preserved. A user interface is designed increasing efficiency, which facilitates the implementation.

The existence of adversarial attacks on convolutional neural networks (CNN) questions the fitness of such models for serious applications. The attacks manipulate an input image such that misclassification is evoked while still looking normal to a human observer -- they are thus not easily detectable. In a different context, backpropagated activations of CNN hidden layers -- "feature responses" to a given input -- have been helpful to visualize for a human "debugger" what the CNN "looks at" while computing its output. In this work, we propose a novel detection method for adversarial examples to prevent attacks. We do so by tracking adversarial perturbations in feature responses, allowing for automatic detection using average local spatial entropy. The method does not alter the original network architecture and is fully human-interpretable. Experiments confirm the validity of our approach for state-of-the-art attacks on large-scale models trained on ImageNet., 13 pages, 6 figures

In order to recognize indoor scenarios, we extract image features for detecting objects, however, computers can make some unexpected mistakes. After visualizing the histogram of oriented gradient (HOG) features, we find that the world through the eyes of a computer is indeed different from human eyes, which assists researchers to see the reasons that cause a computer to make errors. Additionally, according to the visualization, we notice that the HOG features can obtain rich texture information. However, a large amount of background interference is also introduced. In order to enhance the robustness of the HOG feature, we propose an improved method for suppressing the background interference. On the basis of the original HOG feature, we introduce a principal component analysis (PCA) to extract the principal components of the image colour information. Then, a new hybrid feature descriptor, which is named HOG–PCA (HOGP), is made by deeply fusing these two features. Finally, the HOGP is compared to the state-of-the-art HOG feature descriptor in four scenes under different illumination. In the simulation and experimental tests, the qualitative and quantitative assessments indicate that the visualizing images of the HOGP feature are close to the observation results obtained by human eyes, which is better than the original HOG feature for object detection. Furthermore, the runtime of our proposed algorithm is hardly increased in comparison to the classic HOG feature.

L'incessant augment del nombre de seqüències genòmiques, juntament amb l'increment del nombre de tècniques experimentals de les que es disposa, permetrà obtenir el catàleg complet de les funcions cel.lulars de diferents organismes, incloent-hi la nostra espècie. Aquest catàleg definirà els fonaments sobre els que es podrà entendre millor com els organismes funcionen a nivell molecular. Al mateix temps es tindran més pistes sobre els canvis que estan associats amb les malalties. Per tant, la seqüència en brut, tal i com s'obté dels projectes de seqüenciació de genomes, no té cap valor sense les anàlisis i la subsegüent anotació de les característiques que defineixen aquestes funcions. Aquesta tesi presenta la nostra contribució en tres aspectes relacionats de l'anotació dels gens en genomes eucariotes. Primer, la comparació a nivell de seqüència entre els genomes humà i de ratolí es va dur a terme mitjançant un protocol semi-automàtic. El programa de predicció de gens SGP2 es va desenvolupar a partir d'elements d'aquest protocol. El concepte al darrera de l'SGP2 és que les regions de similaritat obtingudes amb el programa TBLASTX, es fan servir per augmentar la puntuació dels exons predits pel programa geneid, amb el que s obtenen conjunts d'anotacions més acurats d'estructures gèniques. SGP2 té una especificitat que és prou gran com per que es puguin validar experimentalment via RT-PCR. La validació de llocs d'splicing emprant la tècnica de la RT-PCR és un bon exemple de com la combinació d'aproximacions computacionals i experimentals produeix millors resultats que per separat. S'ha dut a terme l'anàlisi descriptiva a nivell de seqüència dels llocs d'splicing obtinguts sobre un conjunt fiable de gens ortòlegs per humà, ratolí, rata i pollastre. S'han explorat les diferències a nivell de nucleòtid entre llocs U2 i U12, pel conjunt d'introns ortòlegs que se'n deriva d'aquests gens. S'ha trobat que els senyals d'splicing ortòlegs entre humà i rossegadors, així com entre rossegadors, estan més conservats que els llocs no relacionats. Aquesta conservació addicional pot ser explicada però a nivell de conservació basal dels introns. D'altra banda, s'ha detectat més conservació de l'esperada entre llocs d'splicing ortòlegs entre mamífers i pollastre. Els resultats obtinguts també indiquen que les classes intròniques U2 i U12 han evolucionat independentment des de l'ancestre comú dels mamífers i les aus. Tampoc s'ha trobat cap cas convincent d'interconversió entre aquestes dues classes en el conjunt d'introns ortòlegs generat, ni cap cas de substitució entre els subtipus AT-AC i GT-AG d'introns U12. Al contrari, el pas de GT-AG a GC-AG, i viceversa, en introns U2 no sembla ser inusual. Finalment, s'han implementat una sèrie d'eines de visualització per integrar anotacions obtingudes pels programes de predicció de gens i per les anàlisis comparatives sobre genomes. Una d'aquestes eines, el gff2ps, s'ha emprat en la cartografia dels genomes humà, de la mosca del vinagre i del mosquit de la malària, entre d'altres. El programa gff2aplot i els filtres associats, han facilitat la tasca d'integrar anotacions de seqüència amb els resultats d'eines per la cerca d'homologia, com ara el BLAST. S'ha adaptat també el concepte de pictograma a l'anàlisi comparativa de llocs d splicing ortòlegs, amb el desenvolupament del programa compi., El aumento incesante del número de secuencias genómicas, junto con el incremento del número de técnicas experimentales de las que se dispone, permitirá la obtención del catálogo completo de las funciones celulares de los diferentes organismos, incluida nuestra especie. Este catálogo definirá las bases sobre las que se pueda entender mejor el funcionamiento de los organismos a nivel molecular. Al mismo tiempo, se obtendrán más pistas sobre los cambios asociados a enfermedades. Por tanto, la secuencia en bruto, tal y como se obtiene en los proyectos de secuenciación masiva, no tiene ningún valor sin los análisis y la posterior anotación de las características que definen estas funciones. Esta tesis presenta nuestra contribución a tres aspectos relacionados de la anotación de los genes en genomas eucariotas. Primero, la comparación a nivel de secuencia entre el genoma humano y el de ratón se llevó a cabo mediante un protocolo semi-automático. El programa de predicción de genes SGP2 se desarrolló a partir de elementos de dicho protocolo. El concepto sobre el que se fundamenta el SGP2 es que las regiones de similaridad obtenidas con el programa TBLASTX, se utilizan para aumentar la puntuación de los exones predichos por el programa geneid, con lo que se obtienen conjuntos más precisos de anotaciones de estructuras génicas. SGP2 tiene una especificidad suficiente como para validar esas anotaciones experimentalmente vía RT-PCR. La validación de los sitios de splicing mediante el uso de la técnica de la RT-PCR es un buen ejemplo de cómo la combinación de aproximaciones computacionales y experimentales produce mejores resultados que por separado. Se ha llevado a cabo el análisis descriptivo a nivel de secuencia de los sitios de splicing obtenidos sobre un conjunto fiable de genes ortólogos para humano, ratón, rata y pollo. Se han explorado las diferencias a nivel de nucleótido entre sitios U2 y U12 para el conjunto de intrones ortólogos derivado de esos genes. Se ha visto que las señales de splicing ortólogas entre humanos y roedores, así como entre roedores, están más conservadas que las no ortólogas. Esta conservación puede ser explicada en parte a nivel de conservación basal de los intrones. Por otro lado, se ha detectado mayor conservación de la esperada entre sitios de splicing ortólogos entre mamíferos y pollo. Los resultados obtenidos indican también que las clases intrónicas U2 y U12 han evolucionado independientemente desde el ancestro común de mamíferos y aves. Tampoco se ha hallado ningún caso convincente de interconversión entre estas dos clases en el conjunto de intrones ortólogos generado, ni ningún caso de substitución entre los subtipos AT-AC y GT-AG en intrones U12. Por el contrario, el paso de GT-AG a GC-AG, y viceversa, en intrones U2 no parece ser inusual. Finalmente, se han implementado una serie de herramientas de visualización para integrar anotaciones obtenidas por los programas de predicción de genes y por los análisis comparativos sobre genomas. Una de estas herramientas, gff2ps, se ha utilizado para cartografiar los genomas humano, de la mosca del vinagre y del mosquito de la malaria. El programa gff2aplot y los filtros asociados, han facilitado la tarea de integrar anotaciones a nivel de secuencia con los resultados obtenidos por herramientas de búsqueda de homología, como BLAST. Se ha adaptado también el concepto de pictograma al análisis comparativo de los sitios de splicing ortólogos, con el desarrollo del programa compi., The constantly increasing amount of available genome sequences, along with an increasing number of experimental techniques, will help to produce the complete catalog of cellular functions for different organisms, including humans. Such a catalog will define the base from which we will better understand how organisms work at the molecular level. At the same time it will shed light on which changes are associated with disease. Therefore, the raw sequence from genome sequencing projects is worthless without the complete analysis and further annotation of the genomic features that define those functions. This dissertation presents our contribution to three related aspects of gene annotation on eukaryotic genomes. First, a comparison at sequence level of human and mouse genomes was performed by developing a semi-automatic analysis pipeline. The SGP2 gene-finding tool was developed from procedures used in this pipeline. The concept behind SGP2 is that similarity regions obtained by TBLASTX are used to increase the score of exons predicted by geneid, in order to produce a more accurate set of gene structures. SGP2 provides a specificity that is high enough for its predictions to be experimentally verified by RT-PCR. The RT-PCR validation of predicted splice junctions also serves as example of how combined computational and experimental approaches will yield the best results. Then, we performed a descriptive analysis at sequence level of the splice site signals from a reliable set of orthologous genes for human, mouse, rat and chicken. We have explored the differences at nucleotide sequence level between U2 and U12 for the set of orthologous introns derived from those genes. We found that orthologous splice signals between human and rodents and within rodents are more conserved than unrelated splice sites. However, additional conservation can be explained mostly by background intron conservation. Additional conservation over background is detectable in orthologous mammalian and chicken splice sites. Our results also indicate that the U2 and U12 intron classes have evolved independently since the split of mammals and birds. We found neither convincing case of interconversion between these two classes in our sets of orthologous introns, nor any single case of switching between AT-AC and GT-AG subtypes within U12 introns. In contrast, switching between GT-AG and GC-AG U2 subtypes does not appear to be unusual. Finally, we implemented visualization tools to integrate annotation features for gene- finding and comparative analyses. One of those tools, gff2ps, was used to draw the whole genome maps for human, fruitfly and mosquito. gff2aplot and the accompanying parsers facilitate the task of integrating sequence annotations with the output of homologybased tools, like BLAST.We have also adapted the concept of pictograms to the comparative analysis of orthologous splice sites, by developing compi.

Deep Learning has attained state-of-the-art performance in the recent years, but it is still hard to determine the reasoning behind each prediction. This project will cover the latest advances on interpretability and propose a new method for pixel attribution on image classifiers.