Object Counting via Convolutional Neural Network

Counting is a crucial information acquisition capability with significant impact across various fields such as congestion management and cell counting. While various counting methods exist, visually-based counting is the most intuitive and easiest for humans to understand. However, manually counting a large number of objects in images is not user-friendly, labour-intensive, and prone to errors. Therefore, the computer vision approach has been employed for object counting. The current state-of-the-art in computer vision based counting methods is dominated by convolutional neural network (CNN) based deep learning techniques. These CNN-based counting methods can provide more accurate, efficient, and adaptable solutions for object counting.

However, due to the vast differences in countable object features and the varying scenarios in which counting models can be used, a single CNN model is difficult to solve all counting tasks. This necessitates the use of numerous datasets in current research, resulting in increasingly bulky models that require extensive computational resources for training. Existing datasets designed for counting tasks primarily focus on human targets, lacking specialized datasets for other objects. Addressing the diverse counting needs under such data conditions is a practical challenge for CNN-based counting research. Driven by data-centric approaches, this research aims to achieve counting for different complex scenarios through CNN-based counting architectures to address these research gaps.

In this research, three counting frameworks have been developed for different situations: density map-based counting, point-based counting, and non-point-based counting. First, the Multi-Fusion Convolutional Neural Network (MFCNN) is used in hospital monitoring. This density map-based method implements crowd counting in an indoor scenario. The method combines multidimensional data to enable comprehensive assessment and evidence-based decision-making on the allocation of healthcare resources within a hospital. Second, recognising the systematic errors inherent in the data transformation of density map-based counting, a Point-Detection-based Counting Network (PDCNet) that avoids these errors by utilising direct point labelling is developed. The PDCNet combines three sophisticated point-matching algorithms with a dynamically optimised trained adaptive loss function, reducing computational requirements when compared to traditional methods. Last, to alleviate the requirement of high-accuracy labelling for traditional counting frameworks, this research also proposes a non-point-based counting framework, an Efficient Lightweight Multi-scale-feature-fusion Multi-task GAN (ELMGAN) model. To address the challenges of multi-task training, the framework makes innovative use of the generative adversarial network (GAN) to improve the training performance of generative multi-task models. The effectiveness of the model in generating high-quality segmented images and improving computational efficiency is demonstrated in various public datasets of medical cells.