Measuring difficulty of images using objects of interest - selective attention

• Main Author: Hee, Joshua
• Other Authors: Lye Sun Woh
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Mechanical engineering
• Online Access: https://hdl.handle.net/10356/168330
• Institution: Nanyang Technological University

The objective of this report is to investigate the phenomenon of selective attention by quantifying the saliency of objects in diverse scenarios, and to establish a task capability metric. This metric will serve as a means to objectively evaluate images and determine their level of difficulty, thereby aiding in the selection and training of individuals who are best suited to perform the given task. The afore mentioned is a metric of quantifying difficulty of an image. This results in obtaining a regression model that’s able to determine the difficulty of each image. This report aims to evaluate and discover the prevailing issues of corresponding research lacunae in extant literature. Moreover, a novel solution to quantitatively measure saliency in selective attention will be presented. This report dives comprehensively into the experimental framework, delineating the procedures employed and the obtained results, and subsequently appraising how these outcomes could pave the way for significant headway in future research pursuits. A 9-minute experiment was conducted on 30 participants where their accuracy of results and perceived difficulty of the image was recorded. The experiment required participants to exhibit the ability to pick out the targeted objects when presented with distractions and clutter. The experiment varied the number of targets needed to be picked out, the amount of distracting objects being placed around the targets, as well as the amount of clutter scattered around the image. These variations were extracted to study the participants’ ability to accuracy exercise their selective attention capacity in relation to the 3 variables which would subsequently quantify the difficulty of each image. The difficulty of each and ultimately any image would be determined with a regression model. From the 30 experimental studies, the accuracy of answers were tabulated together with the perceived difficulty of each question. It was seen that the less accurate the answers provided, the higher the level of difficulty perceived. This proves to be sensible. The difficulty of each image was ranked according to the accumulated average of each response provided. In experiments involving a low object count across variables, difficulty was revealed to be low, a value of 1.87 (out of 5). Images with a higher target count in general saw an increase in difficulty, a difficulty ranking of medium to hard, values of 2.74 and 4.2 (out of 5). Images with a higher distractor count in general saw an increase in difficulty, a difficulty ranking of medium to hard, values of 2.48, 4.13 and 4.37 (out of 5). Images with a higher clutter count in general saw an increase in difficulty of hard, values of 4.2 and 4.37 (out of 5). The findings offer a measurable technique for categorizing image complexity based on the factors of targets, distractors, and clutter, with potential applicability to diverse domains and objects. The identification of the Saliency formula, regression model, opens up prospects for investigating attentional levels concerning visual task participation, as well as for gauging the proficiency level needed to undertake a specific job. Factors relating to cueing, in addition to existing variables, can be added. Future research can proceed on from perception to comprehension.