ABSTRACT
Detecting, locating, and classifying objects correctly is a hard task. It is essential to have reliable models with high-accuracy levels, especially when handling challenging images, such as Infrared (IR) scenarios. A convolutional neural network (CNN) is a highly accurate classifier that may detect and identify objects within an image as stated in [Voulodimos et al., 2018], including IR images. CNN presents more image recognition and object detection advantages than standard fully connected neural networks. Many models perform all the object detection, location, and recognition tasks with a single CNN, but as a result, they become larger, more complex, and consume more computational resources [Nielsen, 2015]. Some examples are [Malik et al., 2020], [Olmez and Dokur, 2003], and [Akula et al., 2017] which¨ present a CNN-based deep learning framework for the recognition of targets in IR images. Another good example of object identification and recognition for IR imaging using complex CNN is addressed in [d'Acremont et al., 2019], which is performed in the wild for defense applications where no large-scale datasets are available. In the case of IR images, the number of features is limited, and a grayscale single-channel is the only available information source. This makes them much more challenging to work with than the common color images (RGB), which have three information channels available, among other things. The chapter aims to provide relevant theoretical frameworks and some of the latest empirical research findings. It is written for professionals who want to improve their understanding of the strategic techniques to detect and locate objects in challenging images, and the measurements employed to assess them. Finally, an emerging model using a small CNN and wavelet filters based on a lifting scheme (multi-resolution analysis (MRA)) to improve the feature extraction process is presented. The performance of the proposed hybrid model is assessed using an IR benchmark dataset, comparing its performance with the standard pooling methods: maximum, average, and mixed pooling.
