We investigates the performance of deep Convolutional Neural Network (CNN) for recognizing highway traffic congestion state in surveillance camera images. Different from the usual images in ImageNet, images generated by highway surveillance cameras usually have much more extensive range of perspective and thus larger area of background. Therefore the objective road and vehicles are not as prominent as target object in ImageNet images. And also these images from cameras across a large number of highway sites could show a very rich variance of scenes, road configurations. We are very interested to study whether convolutional networks are still reliably able to classify such images, without any special previous processing such as segmentation of objective roads. Two classic convolutional networks, AlexNet and GoogLeNet are employed to classify congestion state. We build a highway imagery dataset using real-life traffic videos to evaluate the CNNs recognition performance. These images cover a wide range of road configurations, times of the day, weather and lighting conditions, and have been labeled with one of the two states, congestion or non-congestion. The experimental results indicate that under the current strategy of feeding images directly into networks, both AlexNet and GoogLeNet can achieve an excellent recognition accuracy of 98% on held-out test samples. And many of the misclassified images turn out to be borderline cases. More results include that scale and perspective in photography could affect the recognition result.
Regular textures are frequently found in man-made environments and some biological and physical images. There are a wide range of applications for recognizing and locating regular textures. In this work, we used deep convolutional neural networks (CNNs) as a general method for modelling and classifying regular and irregular textures. We created a new regular texture database and investigated two sets of deep CNNs-based methods for regular and irregular texture classification. First, the classic CNN models (e.g. inception, residual network,etc.) were used in a standard way. These two-class CNN classifiers were trained by fine-tuning networks using our new regular texture database. Next, we transformed the trained filter features of the last convolutional layer into a vector representation using Fisher Vector pooling (FV). Such representations can be efficiently used for a wide range of machine learning tasks such as classification or clustering, thus more transferable from one domain to another. Our experiments show that the standard CNNs attained sufficient accuracy for regular texture recognition tasks. The Fisher representations combined with support vector machine (SVM) also showed high performance for regular and irregular texture classification. We also find CNNs performs sub-optimally for long-range patterns, despite the fact that their fully-connected layers pool local features into a global image representation.