Soutenance de thèse : Feng HE

Configurable Convolutional Neural Networks: Applications to Breast Cancer Explainable Classification and Display Panel Defect Detection

Doctorant : Feng HE

Laboratoire INSA : CREATIS

Ecole doctorale : ED160 : Electronique, Electrotechnique, Automatique

Current deep learning methods such as convolutional neural networks (CNNs) are often dedicated to a specific task and object; they are generally fixed in network architecture, which limits their generalizability and prevents them from addressing multiple scenarios with different objectives. To achieve both the explainable classification of breast cancer and the online defect detection of display panels, we propose a configurable convolutional neural network (ConfigNet) capable of being transformed into different configurations according to the tasks and objects in question. The ConfigNet presents two main functional configurations. The first is composed of a feature extraction module (FEM), a decision map generator (DMG) and a classifier; it is devoted to image explainable classification, for which we propose two DMG structures and a weighted average pooling (WAP) classifier for histopathological breast cancer images. The second is an encoder- decoder configuration devoted to object segmentation and localization. In this second configuration, we propose an efficiency-favored decoder and an element-wise feature fusion module (EFFM) guiding the skip connection between the encoder and decoder for online defect detection of display panels. In addition, we develop a spatial and channel attention-guided feature fusion module (SCAFFM) and a bottleneck-structured decoder for breast tumor segmentation. The FEM or encoder in these two configurations is constructed through transfer learning from existing CNNs having deep convolutional layers.