Abstract A system that reconstructs 3D models from a single camera monitoring fish transported on a conveyor system is investigated. Models are subsequently used for training a species classifier and for improving estimates of discarded biomass. It is demonstrated that a monocular camera, combined with a conveyor's linear motion produces a constrained form of multiview structure from motion, that allows the 3D scene to be reconstructed using a conventional stereo pipeline analogous to that of a binocular camera. Although motion stereo was proposed several decades ago, the present work is the first to compare the accuracy and precision of monocular and binocular stereo cameras monitoring conveyors and operationally deploy a system. The system exploits Convolutional Neural Networks (CNNs) for foreground segmentation and stereo matching. Results from a laboratory model show that when the camera is mounted 750 mm above the conveyor, a median accuracy of <5 mm can be achieved with an equivalent baseline of 62 mm. The precision is largely limited by error in determining the equivalent baseline (i.e. distance travelled by the conveyor belt). When ArUco markers are placed on the belt, the inter quartile range (IQR) of error in z (depth) near the optical centre was found to be ±4 mm.

This paper considers the problem of vessel segmentation in optical fundus images of the retina. We adopt an approach that uses a machine learning paradigm to identify texture features called textons and present a new filter bank (MR11) that includes bar detectors for vascular feature extraction and other kernels to detect edges and photometric variations in the image. Textons are generated by k-means clustering and texton maps representing vessels are derived by back-projecting pixel clusters onto hand labelled ground truth. A further step is implemented to ensure that the best combinations of textons are represented in the map and subsequently used to identify vessels in the test set. The experimental results on two benchmark datasets show that our proposed method performs well compared to other published work and the results of human experts. A further test of our system on an independent set of optical fundus images verified its consistent performance.

This paper describes the use of color image analysis to automatically discriminate between oesophagus, stomach, small intestine, and colon tissue in wireless capsule endoscopy (WCE). WCE uses "pill-cam" technology to recover color video imagery from the entire gastrointestinal tract. Accurately reviewing and reporting this data is a vital part of the examination, but it is tedious and time consuming. Automatic image analysis tools play an important role in supporting the clinician and speeding up this process. Our approach first divides the WCE image into subimages and rejects all subimages in which tissue is not clearly visible. We then create a feature vector combining color, texture, and motion information of the entire image and valid subimages. Color features are derived from hue saturation histograms, compressed using a hybrid transform, incorporating the discrete cosine transform and principal component analysis. A second feature combining color and texture information is derived using local binary patterns. The video is segmented into meaningful parts using support vector or multivariate Gaussian classifiers built within the framework of a hidden Markov model. We present experimental results that demonstrate the effectiveness of this method.

Wireless Capsule Endoscopy is a colour imaging technology that enables close examination of the interior of the entire small intestine. The Wireless Capsule Endoscope (WCE) operates for ∼ 8 hours and captures ∼ 40, 000 useful images. The images are viewed by a clinician as a video sequence, generally taking over an hour to analyse. In this paper we present a method of automatically discriminating stomach and intestine tissue which can significantly speed-up one key part of the video analysis time, namely the process of locating the Pylorus- the valve between the stomach and the intestine. We divide the WCE image into 28 sub-regions and process only those regions where tissue is clearly visible. We create a feature vector using colour and texture information. The colour features are derived from Hue Saturation chromaticity histograms of the useful regions, compressed using a hybrid transform, incorporating the Discrete Cosine Transform (DCT) and Principal Component Analysis (PCA). The texture features are derived by Singular Value Decomposition of the same tissue regions. After training the Support Vector classifier, we apply a discriminator algorithm, which scans the video with an increasing step and builds up a classification result sequence. By minimizing the number of misclassifications within this sequence, we predict the most probable position of the Pylorus. We present experimental results that demonstrate the effectiveness of this method.

This paper proposes a new representation for color texture using a set of multiscale illuminant invariant features. The approach was specifically developed to investigate the feasibility of using machine vision to automatically monitor populations of animal species in the Amazon Forest. The approach uses a combination of Finlayson's (1996) color angle idea and Gabor multichannel filters. Using a database of color textures from species of Amazonian monkey, and also a previously published reference database of color regions, we show that the approach performs better than methods based on color angles or Gabor filters alone. The Monkey database was compiled from texture segments extracted from a video of the Amazon Forest using a spatial-temporal segmentation algorithm. The approach is evaluated by applying two classification tests in order to measure the quality of the recognition features: root mean square (RMS) analysis and receiver operating characteristic (ROC) analysis.

This paper proposes a new representation for color texture using a set of multiscale illuminant invariant features. The approach was specifically developed to investigate the feasibility of using machine vision to automatically monitor populations of animal species in ecologically sensitive regions, such as the Amazon Forest. The approach uses a combination of Finlayson's (1994) color angle idea and Gabor multichannel filters and was inspired by the multichannel model of the human visual system (HVS). Using a database of color textures from three species of Amazonian monkey, and also a previously published reference database of color regions, we show that the approach performs better than methods based on color angles or Gabor filters alone. The Monkey database was compiled from texture segments extracted from a video of the Amazon Forest using a spatio-temporal segmentation algorithm. The approach is evaluated by applying two different classification tests in order to measure the quality of the recognition features root mean square (RMS) analysis and receiver operating characteristic (ROC) analysis.

Modern neural and adaptive systems often have complicated error performance surfaces with many local extrema. Visualising and understanding these surfaces is critical to effective tuning of these systems but almost all visualisation methods are confined to two dimensions. Here we show how to use a morphological scale-space transform to convert these multi-dimensional complex error surfaces into two-dimensional trees where the leaf nodes are local minima and other nodes represent decision points such as saddle points and points of inflection.

Images are re-mapped as scale-space trees. The minimal data structure is then augmented by "complement nodes" to increase the practical value of the representation. It is then shown how the resulting ctree can be used to remove noise from images, provide a hierarchical way to estimate a dense disparity map from a stereo pair and to provide a basic segmentation of images for image retrieval purposes.

The method known as SSD, sum of squared difference, is a popular and moderately robust method for dense stereo matching, but it has several known limitations. Various refinements have been proposed, including actively changing the window shape during processing and multi-resolution analysis. We use the multi-scale tree representation of the image to generate window shapes that are well matched to image objects. We use the structure of the tree to refine our disparity search, creating a disparity tree. This tree is a close representation of the objects in the scene, and is a step towards creating a symbolic representation of objects within a scene. We find that the new method performs particularly well in regions of low texture: operating conditions that cause conventional methods to fail.