The tiny size of defects and the noise information in the Ball-Grid-Array (BGA) Chip images have been challenging the image-processing based visual inspection systems in the Integrated Circuit (IC) manufacturing industry. Moreover, the gradient vanishing and high time-consuming problems of deep neural network models are a big obstacle for its application to solve industrial projects. This paper focuses on proposing a modified version of the YOLOv3 model, a leading object detection and classification method in terms of speed; and its application to deal with the problem of detecting and classifying defects on BGA Chip Images. There are five modifications constructed on (4) YOLOv3 architecture that are aimed to enhance the ability of feature extraction towards small objects, strengthen the flow of information inside the network and eliminate the problem of redundant information. With the application of this model into BGA Chip Defects estimation, 49 patches (320x320 pixels), extracted from a single high-resolution BGA Chip image (1, 450x1,450 pixels), are continuously fed into the modified YOLOv3 model to detect and classify the inner defects. As a result, the problem of BGA Chip Defects estimation is solved with the highest performance achieves an average precision of 86% at IoU (Intersection over Union) of 0.75 and an average recall of 99%.