Particle Detection of Complex Images Based on Convolutional Neural Network

Particle Detection plays important roles in industrial and medical image processing. Non-learning based methods can only deal with the particle images with simple backgrounds and sparse particles. And most neural network based methods can also hardly to deal with overlapping objects and small particles, besides, the huge amount of parameters of these methods also has a great demand for annotated data, which is difficult to get. Therefore, we use bright pixels to represent the spatial position of particles, and artificially divid the sizes of particles into multiple intervals. A feature representation called Light-spots Maps is proposed to replace the bounding box and pixel classifications in our detection networks. Some Light-weight networks based on ResNets is used to transform the particle images into Light-spots maps to detect the particle information. Through experiments, we prove the effectiveness of our method.