Third-generation thermal cameras have high dynamic range (HDR), up to 14bit. They are designed to collect images with contrasts exceeding the range of traditional display devices. Sophisticated techniques are therefore required for signal visualization while maintaining and possibly improving object visibility and image contrast. Although this impediment has been tackled in the visual wavelength spectrum, it has scarcely been addressed in the IR. Current means of visualizing HDR images distinguish between contrast-enhancement and dynamic-range-compression techniques.1 Both methods suffer from noise amplification in uniform areas and generate artifacts due to overenhancement of high-frequency details, known as halo artifacts. We have developed a computationally fast approach that fuses both techniques, leading to artifact reduction and enhanced visibility.
One of the most common contrast-enhancement techniques uses histogram equalization. However, this procedure requires an excessive number of output levels to map large flat areas even if they exhibit little temperature variation, while other important image regions will be mapped more sparsely even if they contain interesting structures. This represents a serious problem for IR applications. A ‘contrast-limited adaptive-histogram equalization’ (CLAHE) algorithm was developed2 to deal with this difficulty. When the intensity histogram exceeds a given fractional value, the excess intensity is redistributed uniformly over the entire histogram, thus limiting the slope of the equalization function. We recently presented a ‘balanced CLAHE’ approach1where we optimized the algorithm for use with IR images. Its output is a mapping function that enhances the medium- to low-contrast perceptibility.
Figure 1 illustrates the proposed balanced CLAHE contrast-enhancement approach. The method is applied to the input image (which is divided into Bblocks for increased efficiency) to obtain the basic mapping function. This is subsequently modified using local contrast properties, thus generating a set of mapping functions suitable for local contrast enhancement using a small number of digital levels. Finally, the results obtained for each block are combined to remove any discretization artifacts.
Figure 1. Flow chart of the proposed local-contrast-enhancement system. The input image is divided into a user-defined number (represented by the green arrow) of B blocks, for each of which the local contrast is enhanced (Contr. Enh.) by applying our new balanced contrast-limited adaptive-histogram-equalization (Bal. CLAHE) technique.
The contrast-enhancement performance of this balanced CLAHE approach was evaluated by analysis of real HDR IR images. Each of these presented challenging scenarios, including the presence of a horizon, ‘flat regions,’ small hot objects, and other limitations typical of IR applications. The results were compared with the outcomes from two well-established dynamic-range-compression methods currently offering the best visual quality, Fattal's gradient-domain technique3 and a homomorphic Frankle–McCann Retinex algorithm,4 and from a traditional histogram-equalization method.
Figure 2. Performance comparison of a range of current contrast-enhancement techniques. (a) Input image (14bit), (b) histogram-equalization procedure, (c) Fattal's technique, (d) Frankle–McCann Retinex method, and (e) the proposed balanced CLAHE approach with B = 9.
Our new technique is robust in the presence of a horizon (see Figure 2) since it attenuates the large intensity difference between sky and ground regions. Figure 2 also shows that the method successfully deals with the presence of small low-contrast structures. A quantitative performance analysis was also carried out on the basis of a ‘mean-opinion score’ involving a group of 15 observers. The analysis showed that our new technique outperforms Fattal's, the Frankle–McCann Retinex, and histogram-equalization methods in terms of the perceived accuracy, visibility of details, ability to handle images containing small hot objects, and overall quality of the output image.1 The algorithm also successfully displays the acquired signal, enhances the contrast of potential threats such as aerial cables and trees, and improves the perception of objects in darker areas. Our ongoing work focuses on applying this method to IR sequences to test its performance in real-time processing for applications such as area surveillance and object detection and recognition.
Francesco Branchitta, Marco Diani, Giovanni Corsini
Department of Information Engineering
University of Pisa
Campi Bisenzio, Firenze, Italy
2. S. M. Pizer, E. P. Amburn, J. D. Austin, R. Cromartie, A. Geselowitz, T. Greer, B. T. H. Romeny, J. B. Zimmerman, Adaptive histogram equalization and its variations, Comput. Vision Graph. 39, pp. 355-368, 1987.