CENet: improve counting performance of X-ray surface mounted chip counter via scale favor and cell extraction

The X-ray Surface Mounted Chip Counter (X-SMDCC) relies on a counting algorithm to count the number of surface-mounted chips, enabling convenient and fast counting. It is an efficient auxiliary equipment for SMT material management. However, most existing counting algorithms use crowd counting algorithms for fine-tuning, without designing a special structure to optimize processing based on the differences and characteristics of data in crowd counting and X-SMDCC, leading to inaccurate counting results under the condition of chip scale change or adhesion. In this work, we propose a cell extraction network to address the issues of scale difference and adhesion, which improves the counting accuracy of X-SMDCC. Firstly, we present a scale-favoring module to handle scale differences between different images, as we notice that the scale difference only appears between different images. Furthermore, we propose a cell extraction module to process adhesive regions since we discovered that the human eye can process adhesive regions through comparison while labeling data. Additionally, we recommend using a shape-constrained inverse distance transform map as a learning target. We conducted numerous experiments on the SMD-Chip-179 dataset and found that our method is significantly superior to current advanced counting methods.